CN111651692A

CN111651692A - Information recommendation method and device based on artificial intelligence and electronic equipment

Info

Publication number: CN111651692A
Application number: CN202010488298.0A
Authority: CN
Inventors: 徐宝川; 张伸正
Original assignee: Tencent Technology Beijing Co Ltd
Current assignee: Tencent Technology Beijing Co Ltd
Priority date: 2020-06-02
Filing date: 2020-06-02
Publication date: 2020-09-11

Abstract

The invention provides an information recommendation method, device, electronic equipment and computer readable storage medium based on artificial intelligence; the method comprises the following steps: acquiring the characteristics of each piece of information in an information set to be recommended; based on the relevance among a plurality of pieces of information in the information set, carrying out relevance processing on the characteristics of each piece of information to obtain the relevance characteristics of each piece of information; determining a first click rate corresponding to a plurality of combinations when each piece of information is displayed in the page in the plurality of combinations of display positions and display styles based on the associated characteristics of each piece of information in the information set; determining an arrangement mode which enables the overall click rate of the page to be the highest based on the first click rates of the information corresponding to the multiple combinations respectively, wherein the arrangement mode comprises the display position and the display style of the information in the page; and executing the recommendation operation based on the information set with the arrangement mode. According to the invention, the information recommendation accuracy can be improved, so that the evaluation index of a recommendation system is improved.

Description

Information recommendation method and device based on artificial intelligence and electronic equipment

Technical Field

The present invention relates to information recommendation technologies based on artificial intelligence, and in particular, to an information recommendation method and apparatus based on artificial intelligence, an electronic device, and a computer-readable storage medium.

Background

Cloud Computing (Cloud Computing) is a Computing model that distributes Computing tasks over a resource pool of large numbers of computers, enabling various application systems to obtain Computing power, storage space, and information services as needed. The network that provides the resources is referred to as the "cloud". Resources in the "cloud" appear to the user as being infinitely expandable and available at any time, available on demand, expandable at any time, and paid for on-demand.

Artificial Intelligence (AI) is a theory, method and technique and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and expand human Intelligence, perceive the environment, acquire knowledge and use the knowledge to obtain the best results.

Information recommendation is an important application of artificial intelligence, a rearrangement module in the related technology is a final stage of personalized recommendation of a recommendation system, firstly, information sorted by a fine-arrangement model is collected according to the maximum quantity limit of each category, then, the display style of the collected information and the display position in a display list are determined, and the information is finally applied to the collected information and displayed to a user, but the problems that a reordering strategy is single and the recommendation lacks personalization exist.

Disclosure of Invention

The embodiment of the invention provides an information recommendation method and device based on artificial intelligence, an electronic device and a computer readable storage medium, which can improve the information recommendation accuracy rate, thereby improving each evaluation index of a recommendation system.

The technical scheme of the embodiment of the invention is realized as follows:

the embodiment of the invention provides an information recommendation method based on artificial intelligence, which comprises the following steps:

based on the relevance among a plurality of pieces of information in an information set, carrying out relevance processing on the characteristics of each piece of information to obtain the relevance characteristics of each piece of information;

determining a first click rate corresponding to a plurality of combinations of display positions and display styles of each piece of information when the information is displayed in the page in the plurality of combinations based on the associated characteristics of each piece of information in the information set;

determining an arrangement mode which enables the overall click rate of the page to be the highest based on the first click rates of the information corresponding to the combinations respectively;

the arrangement mode comprises a display position and a display style of each piece of information in the page;

and executing recommendation operation based on the information set to which the arrangement mode is applied.

The embodiment of the invention provides an information recommendation device based on artificial intelligence, which comprises:

the association processing module is used for carrying out association degree processing on the characteristics of each piece of information based on the association degrees among a plurality of pieces of information in the information set to obtain the association characteristics of each piece of information;

the click rate determining module is used for determining first click rates respectively corresponding to a plurality of combinations of display positions and display styles when each piece of information is displayed in the page according to the associated characteristics of each piece of information in the information set;

the arrangement mode determining module is used for determining an arrangement mode which enables the overall click rate of the page to be the highest based on the first click rates of the information corresponding to the combinations respectively, wherein the arrangement mode comprises the display position and the display style of the information in the page;

and the recommending module is used for executing recommending operation based on the information set to which the arrangement mode is applied.

In the above scheme, before obtaining the feature of each piece of information in the information set to be recommended, the click rate determining module is further configured to:

acquiring the characteristics of each piece of information in the information base to determine a corresponding second click rate;

performing descending sorting processing on the information base based on the second click rate of each piece of information, and selecting a plurality of pieces of information sorted at the front from descending sorting results to form the information set;

wherein the number of information in the set of information is the same as the number of display positions in the page.

In the above solution, the apparatus further comprises: a feature acquisition module further configured to:

performing the following for each information in the set of information:

inquiring a characteristic vector corresponding to the characteristic of the information from a pre-established characteristic vector matrix;

and carrying out fusion processing on the feature vector of each piece of information to obtain the feature corresponding to each piece of information.

In the foregoing solution, the association processing module is further configured to:

performing the following for each information in the set of information:

performing attention coding processing on the characteristics of each piece of information in the information set to obtain the association degree between the information and each piece of information in the information set;

and determining the association characteristics of each information based on the association degree between the information and each information in the information set.

carrying out linear transformation processing on the characteristics of the information to obtain a query vector, a key vector and a value vector which respectively correspond to the characteristics;

performing point multiplication on the query vector of the information and the key vector of each information in the information set, and performing normalization processing on a point multiplication processing result based on a maximum likelihood function to obtain the association degree between the information and each information in the information set;

determining the degree of association as an attention weight of a vector of values corresponding to the each information;

and carrying out weighting processing on the value vector based on the attention weight to obtain the associated characteristics of the information based on attention coding processing.

In the foregoing solution, the click rate determining module is further configured to:

and performing full connection processing on the associated features of each piece of information to obtain a plurality of first click rates which are in one-to-one correspondence with a plurality of combinations of display positions and display styles when each piece of information is displayed in the page in the plurality of combinations.

In the foregoing solution, the arrangement determining module is further configured to:

for each information in the set of information, performing the following:

binding the information with each display position respectively to obtain an information position binding result of the information corresponding to each display position;

each information position binding result has a first click rate respectively corresponding to a plurality of display styles;

for each information position binding result, executing the following processing:

in the first click rate corresponding to a plurality of display styles, taking the display style corresponding to the highest first click rate as the display style corresponding to the information position binding result;

performing path search processing on the plurality of information position binding results to obtain a path which enables the overall click rate of the page to be the highest;

and determining a plurality of information position binding results and respectively corresponding display styles included in the path search processing result as an arrangement mode which enables the overall click rate of the page to be the highest.

for information I in the information set_nThe information I is processed_nAnd the display position P_qBinding to obtain an information position binding result R_nq；

N is more than or equal to 1 and less than or equal to K, q is more than or equal to 1 and less than or equal to K, K is the number of information in the information set and is an integer more than or equal to 2, and J is the number of the display styles and is an integer more than or equal to 2;

binding the result at the information locationR_nqDetermining a presentation style with the highest first click rate among first click rates corresponding to a plurality of presentation styles as the information I_nAt the display position P_qThe display style used in the display.

selecting m information sequences which enable the overall click rate of the first k display positions to be the highest from the information set to serve as candidate information sequences of the first k display positions;

wherein m is the path search size of the path search processing, and the value range satisfies that m is more than or equal to 1 and less than or equal to K;

k is an integer with the value increasing from 1, and the value range of K satisfies that K is more than or equal to 1 and is less than K; the information in the information sequence corresponds to the first k display positions one by one;

and selecting the information sequence with the highest overall click rate from the candidate information sequences of the previous K display positions to serve as a path enabling the overall click rate of the page to be the highest.

determining a first click rate of a corresponding display position in the front k-1 display positions aiming at each information in the m candidate information sequences of the front k-1 display positions;

for each of the m candidate information sequences, performing the following:

adding the first click rates corresponding to each information in the candidate information sequence to obtain the whole click rate of the candidate information sequence corresponding to the first k-1 display positions;

acquiring a first click rate of each piece of residual information at a kth display position, wherein the residual information is information in the information set except information in the candidate information sequence;

adding the first click rate of each piece of residual information at the kth display position with the integral click rate respectively to obtain integral click rates of a plurality of information sequences corresponding to the first k display positions;

wherein the plurality of information sequences are combinations of the candidate information sequences and a plurality of the remaining information sequences, respectively, the remaining information being ordered after the candidate information sequence in each of the information sequences;

and selecting m information sequences with the highest overall click rate ranking from a plurality of information sequences obtained based on the m candidate information sequences.

In the above solution, the characteristic of the information includes at least one of:

basic attribute characteristics used for representing basic information of a user to be recommended; the interest label characteristics are used for representing interest preferences of the user to be recommended; the environment characteristics are used for representing a recommendation environment for recommending the information to the user to be recommended; a category feature for characterizing a category of the information; source characteristics for characterizing a source of the information; content characteristics for characterizing the content of the information.

In the above scheme, a first click rate at which each piece of information in the information set is displayed in a different display style at each display position is determined by calling a click rate prediction model, where the click rate prediction model includes an attention coding structure; the device further comprises: a training module to: before obtaining the characteristics of each information in the information set to be recommended,

generating a training sample set for training the click rate prediction model;

each training data sample in the training sample set comprises a sample information sequence, and the information with the preset number ranked at the top in the sample information sequence is effective sample information;

performing attention coding processing on the characteristics of each sample information in the sample information sequence to obtain the correlation degree between the sample information and each effective sample information in the sample information sequence;

determining an association characteristic of each sample information based on the association degree between the sample information and each effective sample information in the sample information sequence;

carrying out forward propagation on the associated features of each sample information in each training data sample in the click rate prediction model to obtain a predicted first click rate of each sample information of each training data sample displayed in a sample display mode at a sample display position;

and performing back propagation in the click rate prediction model based on the difference between the predicted first click rate and the real first click rate, and updating the parameters of the click rate prediction model in the process of back propagation.

An embodiment of the present invention provides an electronic device, including:

a memory for storing executable instructions;

and the processor is used for realizing the artificial intelligence based information recommendation method provided by the embodiment of the invention when the executable instructions stored in the memory are executed.

The embodiment of the invention provides a computer-readable storage medium, which stores executable instructions and is used for realizing the artificial intelligence-based information recommendation method provided by the embodiment of the invention when being executed by a processor.

The embodiment of the invention has the following beneficial effects:

the method comprises the steps of obtaining association characteristics capable of representing association influence among information by performing association processing on characteristics of the information, and determining predicted click rates of the information at various positions and presented in various modes through association characteristic quantification so as to obtain an optimal arrangement mode of information display modes and positions and enable indexes of a recommendation system to grow in a forward direction.

Drawings

FIG. 1 is a schematic diagram of an architecture of an artificial intelligence-based information recommendation system provided by an embodiment of the present invention;

fig. 2 is a schematic structural diagram of a server 200 applying an artificial intelligence-based information recommendation method according to an embodiment of the present invention;

FIG. 3 is a diagram of an overall model architecture of an artificial intelligence-based information recommendation method according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of a model attention structure of an artificial intelligence-based information recommendation method provided by an embodiment of the present invention;

FIG. 5 is a schematic diagram of a model output of an artificial intelligence-based information recommendation method provided by an embodiment of the present invention;

6A-6D are schematic flow diagrams of artificial intelligence based information recommendation methods provided by embodiments of the invention;

FIG. 7 is an overall architecture diagram of an artificial intelligence-based information recommendation method provided by an embodiment of the present invention;

fig. 8 is a schematic diagram of path search optimization of an artificial intelligence based information recommendation method according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail with reference to the accompanying drawings, the described embodiments should not be construed as limiting the present invention, and all other embodiments obtained by a person of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

In the following description, reference is made to "some embodiments" which describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict.

In the following description, references to the terms "first \ second \ third" are only to distinguish similar objects and do not denote a particular order, but rather the terms "first \ second \ third" are used to interchange specific orders or sequences, where appropriate, to enable embodiments of the invention described herein to be practiced in other than the order shown or described herein.

Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to be limiting of the invention.

Before further detailed description of the embodiments of the present invention, terms and expressions mentioned in the embodiments of the present invention are explained, and the terms and expressions mentioned in the embodiments of the present invention are applied to the following explanations.

1) The recommendation system comprises: recommendation systems are a tool for automatically contacting users and information, which can help users find information interesting to them in an information overload environment, and can push information to users interested in them.

2) A rearrangement module: the rearrangement module is the final stage of the recommendation system, and the recommendation system selects and rearranges the candidate information to further determine the display style and the display position of the information.

3) Information personalized recommendation: recommending information interested by the user to the user according to the interest characteristics and the browsing behavior of the user, wherein the information comprises: merchandise, news, etc.

4) Click Through Rate (CTR, Click-Through-Rate), actual Click times of recommendation information in a recommendation system are divided by the display number of the information, and the predicted Click Through Rate is the score of the user clicking the information to be recommended, which is usually used for fine ranking and sorting and the like.

The following technical solutions exist for the rearrangement problem in the related art: the display style and the display position of the information are determined by using a simple strategy, the display positions of the information are determined in a random arrangement mode, or the display positions of the information are determined by arranging according to the sequence of the precisely-arranged predicted click rate, the display style is determined based on the priority of the display style, the influence of user characteristics and the characteristics of the information on the rearrangement result is not considered, the correlation influence between the information is not considered, the quantitative prediction of the click rate of the information at each display position based on various display styles is not considered, so that the personalized rearrangement cannot be realized, and the recommendation accuracy is reduced.

Aiming at the problems that recommendation is not accurate, personalized rearrangement cannot be realized and the like in the methods provided by the related technology, the embodiment of the invention provides an information recommendation method based on artificial intelligence, a device, electronic equipment and a computer readable storage medium, which can solve the problems that recommendation accuracy is low and recommendation lacks personalization, and is an information recommendation method based on a neural network model.

An exemplary application of the electronic device provided by the embodiment of the present invention is described below, and the electronic device provided by the embodiment of the present invention may be implemented as a server. In the following, an exemplary application will be explained when the electronic device is implemented as a server.

The server may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server providing basic cloud computing services such as cloud service, a cloud database, cloud computing, a cloud function, cloud storage, network service, cloud communication, middleware service, domain name service, security service, CDN, and a big data and artificial intelligence platform.

The so-called artificial intelligence cloud Service is also generally called AIaaS (AI as a Service, chinese). The method is a service mode of an artificial intelligence platform, and particularly, the AIaaS platform splits several types of common AI services and provides independent or packaged services at a cloud. This service model is similar to the one opened in an AI theme mall: all developers can access one or more artificial intelligence services provided by the platform through an API (application programming interface), and part of the qualified developers can also use an AI framework and an AI infrastructure provided by the platform to deploy and operate and maintain the self-dedicated cloud artificial intelligence services.

Referring to fig. 1, fig. 1 is a schematic diagram of an architecture of an artificial intelligence based information recommendation system according to an embodiment of the present invention, where the information recommendation system may be used to support recommendation scenes of various information, such as an application scene for recommending news, an application scene for recommending goods, an application scene for recommending videos, and the like, and according to different application scenes, the information may be news, actual goods, video information, graphics, and the like, in the information recommendation system, a terminal 400 is connected to a server 200 through a network 300, the network 300 may be a wide area network or a local area network, or a combination of the two, in response to receiving a recommendation information request of the terminal 400, the function of the information recommendation system is implemented based on various modules in the server 200, in response to receiving an information recommendation request of the terminal 400 by the server 200, a click rate determination module 2553 in the server 200 performs first click rate prediction and ranking on information recalled from an information database 500, and selecting K pieces of information in the top ranking to form an information set, a feature obtaining module 2551 in the server 200 obtaining features of the information in the information set obtained after the first ranking, determining click rates of the information displayed in each display style at each display position through a click rate determining module 2553, performing optimization processing based on the click rates through an arrangement mode determining module 2554 to obtain an optimal arrangement mode, and a recommending module 2555 displaying the information according to the optimal arrangement mode to recommend the information to the terminal 400.

Referring to fig. 2, fig. 2 is a schematic structural diagram of a server 200 applying an artificial intelligence-based information recommendation method according to an embodiment of the present invention, where the server 200 includes: at least one processor 210, memory 250, and at least one network interface 220. The various components in server 200 are coupled together by a bus system 240. It is understood that the bus system 240 is used to enable communications among the components. The bus system 240 includes a power bus, a control bus, and a status signal bus in addition to a data bus. For clarity of illustration, however, the various buses are labeled as bus system 240 in fig. 2.

The Processor 210 may be an integrated circuit chip having Signal processing capabilities, such as a general purpose Processor, a Digital Signal Processor (DSP), or other programmable logic device, discrete gate or transistor logic device, discrete hardware components, or the like, wherein the general purpose Processor may be a microprocessor or any conventional Processor, or the like.

The memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid state memory, hard disk drives, optical disk drives, and the like. Memory 250 optionally includes one or more storage devices physically located remotely from processor 210.

In some embodiments, memory 250 is capable of storing data, examples of which include programs, modules, and data structures, or a subset or superset thereof, to support various operations, as exemplified below.

An operating system 251 including system programs for processing various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks;

a network communication module 252 for communicating to other computing devices via one or more (wired or wireless) network interfaces 220, exemplary network interfaces 220 including: bluetooth, wireless compatibility authentication (WiFi), and Universal Serial Bus (USB), among others.

In some embodiments, the artificial intelligence based information recommendation apparatus provided by the embodiments of the present invention may be implemented in software, and fig. 2 illustrates an artificial intelligence based information recommendation apparatus 255 stored in a memory 250, which includes a plurality of modules of an information recommendation system, where the modules may be software in the form of programs and plug-ins, and include the following software modules: the feature obtaining module 2551, the association processing module 2552, the click rate determining module 2553, the arrangement determining module 2554, the recommending module 2555 and the training module 2556 are logical modules, and therefore, the functions of the modules may be arbitrarily combined or further divided according to the implemented functions, and the functions of the modules will be described below.

The information recommendation method based on artificial intelligence provided by the embodiment of the present invention will be described below with reference to an exemplary application and implementation of the information recommendation system provided by the embodiment of the present invention, where the information recommendation system includes a training phase and an application phase.

Next, a model used in the artificial intelligence based information recommendation method and training performed by the model according to the embodiment of the present invention will be described.

Referring to fig. 3, fig. 3 is an overall model structure diagram of the artificial intelligence based information recommendation method according to the embodiment of the present invention, where a model applied by the artificial intelligence based information recommendation method according to the embodiment of the present invention is a click rate prediction model, the click rate prediction model is actually a neural network model, and is composed of an input layer, a vectorization presentation layer, a feature merging layer, an attention layer, a full connection layer, and an output layer, where the input layer simultaneously inputs a plurality of features of each information (information 1, information 2, …, information K) in an information set, the information in the information set is the first N pieces of information with higher click rate obtained after the recalled information is sorted for the first time, and the information is determined to be information that needs to be presented to a user, that is, to be recommended information corresponding to a brush operation, and therefore, when a rearrangement process is performed, the features of the information are simultaneously used as inputs, the vectorization presentation layer is used for acquiring vectorization presentation of a plurality of input features corresponding to each piece of information, the vectorization presentation is realized by encoding the features through Hash processing, then querying corresponding feature vectors from a preset feature vector matrix according to the encoding, the feature merging layer is used for merging a plurality of feature vectors related to the same information, the number of information in an information set is 10, the feature merging layer can output 10 fusion features respectively corresponding to each piece of information, the self-attention layer performs self-attention processing (association processing) between information on the 10 fusion features respectively corresponding to each piece of information output by the feature merging layer to obtain associated features respectively corresponding to each piece of information, the number of fully-connected layers can be 1 or more, the number of layers of the fully-connected layers is determined according to the number of training sample data during model training, the number of layers of the full-connection layer is positively correlated with the number of the training sample data, the full-connection layer performs full-connection processing on the correlation characteristics respectively corresponding to the information to obtain a first click rate of displaying the information in each display position in each display style, if the number of the information is 10, the number of the display positions is 10 as same as the number of the information, and the number of the display styles is 4, the number of the output first click rate is 10 x 4, and the output first click rate matrix can be represented as a size [10,10,4 ].

In some embodiments, a first click rate of each information in the information set to be recommended displayed in each display style at each display position is obtained by calling a click rate prediction model, and the click rate prediction model comprises an attention mechanism structure; before obtaining the characteristics of each piece of information in an information set to be recommended, executing the following training scheme to generate a training sample set for training a click rate prediction model; each training data sample in the training sample set comprises a sample information sequence, and the information with the preset number ranked in the top in the sample information sequence is effective sample information; performing attention coding processing on the characteristics of each sample information in the sample information sequence to obtain the association degree between the sample information and each effective sample information in the sample information sequence; determining the correlation characteristics of each sample information based on the correlation degree between the sample information and each effective sample information in the sample information sequence; carrying out forward propagation on the associated characteristics of each sample information in each training data sample in a click rate prediction model to obtain a predicted first click rate of each sample information of each training data sample displayed in a real display mode at a real display position; and performing back propagation in the click rate prediction model based on the difference between the predicted first click rate and the real first click rate, and updating the parameters of the click rate prediction model in the process of back propagation.

By way of example, forward transfer of a training data sample is completed through the click rate prediction model shown in fig. 3, the training data sample includes a sample information sequence, and a true click behavior for each sample information in the sample information sequence, a tag having a click behavior for the sample information is 1, a tag having no click behavior for the sample information is 0, a preset number of pieces of information in the sample information sequence ranked in the top is valid sample information, and the applicant finds that, in implementing the embodiment of the present invention, a user does not generally browse the last pieces of information in a piece of brush information when browsing a piece of brush information (sample information sequence) presented, that is, the number of valid sample information (the first pieces of information in a piece of brush information are generally browsed) can be determined according to historical behavior data of the user, so as to determine the determined number of valid sample information in the sample information sequence, therefore, when the associated features of each piece of information are generated in the forward transmission, the feature of each piece of sample information in the sample information sequence is processed by attention coding, so as to obtain the degree of association between the piece of sample information and each piece of valid sample information in the sample information sequence, as shown in fig. 4, fig. 4 is a schematic diagram of a model attention structure of an artificial intelligence-based information recommendation method according to an embodiment of the present invention, in which attention weights (relevance) of sample information of non-valid sample information (information K-1, information K) in a sample information sequence (information 1, information 2, …, information K) are uniformly set to 0, namely, the sample information of the non-effective sample information in the sample information sequence has negligible relevance to other information, and only the relevance between each sample information and each effective sample information in the sample information sequence is determined, so that the first click rate prediction is performed based on the obtained relevance characteristics.

Through the processing of distinguishing the valid sample information from the invalid sample information, the use behavior of the user can be effectively simulated, so that the trained model can more accurately predict the first click rate corresponding to the display of each piece of information in each display position in each display style, and further the recommendation accuracy is improved.

Referring to fig. 5, fig. 5 is a schematic diagram of a model output of an artificial intelligence based information recommendation method provided in an embodiment of the present invention, where an output layer outputs a click rate prediction model for first click rates of all display positions and all display patterns, the number of information predicted each time is k, the number of display positions is k, and the display patterns of information are s, then the first click rate number output for each piece of information is k × s, the first click rate number of all information in one recommendation is k × s, at the stage of training of the click rate model, X represents the click rate output of k × k in fig. 5, and the final output is actually 4X, 4 is the number of display patterns, and a combination between positions and information may form multiple arrangement modes, but only one arrangement mode is true, so that each training sample is model optimized only for a part of output nodes, that is, optimization is performed only based on a true arrangement mode (for example, diagonal nodes shown on the right side of fig. 4, the true arrangement mode is that information shown at a first position in a true information sequence is shown at a first position, information shown at a second position is shown at a second position, and the like), the arrangement mode is used as a target for predicting and reducing loss of a model, the output of nodes of an output layer is adjusted to be a matrix with the size of 10 × 4(k × 10, s × 4), corresponding nodes are obtained according to a determined display pattern of a training sample, the size of the matrix is reduced to be 10 × 10, then model optimization is performed on nodes effective on the diagonal, in a click rate prediction model of a rearrangement module, input k pieces of information have k labels, and output k × s first click rates, so that loss is calculated only for output nodes (diagonal nodes) corresponding to display positions and display patterns of each piece of information in a true sample In addition, the mean square error loss of the overall real click and the overall predicted click of the k pieces of information in each sample is added into a loss function for optimization, and a formula (1) of the loss function of the click rate prediction model of the rearrangement module is as follows:

wherein, y_iThe label of the ith information in the sample is 1 when clicked, 0 when not clicked, p_iThe first click rate of the ith information in the sample is displayed in the corresponding display position in the corresponding display style is determined for the combination of the information, the position and the style due to the loss function aiming at the real arrangement mode_iIs the true presentation style, pos, of the ith information_iIs the real display position of the ith piece of information.

Next, an application of the model in the artificial intelligence based information recommendation method provided by the embodiment of the present invention is described. Referring to fig. 6A, fig. 6A is a flowchart illustrating an artificial intelligence based information recommendation method according to an embodiment of the present invention, which will be described with reference to steps 101-105 shown in fig. 6A.

In step 101, the server obtains the characteristics of each piece of information in the set of information to be recommended.

In some embodiments, before obtaining the features of each piece of information in the information set to be recommended, the following technical scheme may be further performed: acquiring the characteristics of each piece of information in the information base to determine a corresponding second click rate; performing descending sorting processing on the information base based on the second click rate of each piece of information, and selecting a plurality of pieces of information sorted at the front from the descending sorting result to form an information set to be recommended; and the number of the information in the information set to be recommended is the same as the number of the display positions in the page.

As an example, the information set is implemented by a fine ranking module of the recommendation system, the fine ranking module may use a simpler linear model (logistic regression model) to perform second click rate prediction by learning the click behavior of the user, and may finally rank the recalled information (information in the information base) according to a descending order of the size of the second click rate value, select N pieces of information ranked at the top as the information in the information set to be recommended, the number of the information in the information set to be recommended is the same as the number of the display positions in the page, in addition, the linear model may be combined with a deep learning model, the linear model finds correlations between information or features from the history data, the deep learning model partially performs a generalization function, i.e., performs a transfer of correlations, finds new feature combinations that occur little or none in the history data, looking for new preferences of the user, a balance can be better made between the historical interest of the user and exploring new points of interest by combining the two models.

In some embodiments, the obtaining of the characteristics of each piece of information in the information set to be recommended in step 101 may be implemented by the following technical solutions: executing the following processing for each information in the information set to be recommended: inquiring a characteristic vector corresponding to the characteristic data of the information from a pre-established characteristic vector matrix; and performing fusion processing on the feature vector of each piece of information to obtain the feature corresponding to each piece of information.

As an example, the fusion process of the feature vectors may be a summation process of the feature vectors, thereby obtaining a feature corresponding to each piece of information.

In some embodiments, the feature data of the information may be encoded in the feature engineering to obtain the feature vector of the information, which may specifically be implemented by the following technical solutions: converting the characteristic value of the characteristic data into a characteristic index, and performing hash processing on the characteristic index to obtain a characteristic index code; and carrying out Hash processing on the feature names of the feature data to obtain feature name codes, and combining the feature name codes and the feature index codes to obtain the feature vectors of the information.

As an example, in the process of performing hash coding, it is necessary to calculate an index corresponding to feature data and encode the index, in the machine learning process, in order to facilitate implementation of a correlation algorithm, it is often necessary to convert tag data (generally, a character string) into an integer index, or to restore the integer index into a corresponding tag after the calculation is completed, a converter may encode a list of class attribute features (or tags) so as to digitize the class attribute features, the range of the index starts from 0, the process may index corresponding features so that certain algorithms that cannot accept class type features can be used, and the efficiency of a machine learning algorithm such as a decision tree is improved, if input is numerical data, the input may be converted into character type data and then encoded, and a feature name actually represents a feature class, such as an attribute of a feature, the feature data may be mapped to a 64-bit hash space, with the first 16 bits of the 64-bit hash space representing the feature name and the last 48 bits representing the feature index.

In some embodiments, the characteristic of the information comprises at least one of: basic attribute characteristics used for representing basic information of a user to be recommended; the interest label characteristics are used for representing interest preferences of the user to be recommended; the environment characteristics are used for representing a recommendation environment for recommending the information to the user to be recommended; a category feature for characterizing a category of the information; source characteristics for characterizing a source of the information; content characteristics for characterizing the content of the information.

As an example, characteristics of a user side, such as gender, age, region, and family structure of the user, characteristics of an information side, such as information category and information source, may be acquired, characteristics of an environment side, such as a network environment where the user is located, a network device used by the user, and the like, may be acquired, and characteristics of each dimension are acquired to model the first click rate of each presentation style of information at each presentation position, so as to implement more personalized recommendation.

In step 102, the server performs relevance processing on the feature of each information based on the relevance between a plurality of information in the information set to obtain the relevance feature of each information.

Based on fig. 6A and fig. 6B, fig. 6B is a schematic flow chart of an information recommendation method based on artificial intelligence according to an embodiment of the present invention, in step 102, based on the association degree between a plurality of pieces of information in an information set to be recommended, the feature of each piece of information is associated, and the obtained associated feature of each piece of information can be realized through step 1021-.

The following steps 1021-.

In step 1021, the feature of each piece of information in the set of information to be recommended is subjected to attention coding processing, so as to obtain the association degree between the piece of information and each piece of information in the set of information to be recommended.

In step 1022, the association feature of each information is determined based on the association degree between the information and each information in the information set to be recommended.

In some embodiments, the attention coding processing is performed on the features of each piece of information in the information set to be recommended to obtain the association degree between the piece of information and each piece of information in the information set to be recommended, and the following technical solutions may be implemented: carrying out linear transformation processing on the characteristics of the information to obtain a query vector, a key vector and a value vector which respectively correspond to the characteristics; performing point multiplication on the query vector of the information and the key vector of each information in the information set to be recommended, and performing normalization processing on the result of the point multiplication processing based on a maximum likelihood function to obtain the association degree between the information and each information in the information set to be recommended; the above determining the association characteristic of each information based on the association degree between the information and each information in the information set to be recommended can be realized by the following technical scheme: determining the degree of association as an attention weight of a value vector corresponding to each piece of information; and carrying out weighting processing on the value vector based on the attention weight to obtain the relevant characteristics of the information based on the attention coding processing.

As an example, the attention coding processing is performed on the feature of each information in the information set to be recommended, the attention score of each information for a certain information is obtained, the attention score is used as the relevance of the information and each information, the information itself is included, the calculation of the attention score of each information for a certain information is mainly realized by the dot multiplication processing of a query vector and a key vector and the maximum likelihood processing, after the dot multiplication processing and before the maximum likelihood processing are performed, the dot multiplication processing result can be divided by the square root of the dimension of the key vector, the gradient is more stable, the maximum likelihood processing is realized by a softmax function, the attention scores of all the information are normalized, the obtained scores are all positive values and have a sum of 1, the attention score determines the contribution of each information to the information, and the query vector of the feature of each information is obtained, The key vector and the value vector are obtained by linearly transforming the feature vector of the information, that is, the feature vector of the information is multiplied by the query parameter, the key parameter and the value parameter, which are obtained by model training, in order to simplify the training process, the query parameter, the key parameter and the value parameter can be uniformly set to 1, that is, the linear transformation processing is to multiply the feature vector of the information by 1, so that the obtained query vector, the key vector and the value vector are the same and are all feature vectors, and the attention score of each information relative to a certain information and the value vector corresponding to each information are weighted and summed, that is, the associated feature of the information based on the attention coding processing is obtained.

Through the attention coding processing, the information dependency relationship inside the information set can be learned, so that the internal information relationship of the information set can be captured, and in the long-distance dependency problem, because the attention coding processing is to calculate the attention scores of all information and each information, the maximum path length is only 1 no matter how long the information is, so that the long-distance dependency relationship can be captured.

In step 103, the server determines, based on the associated features of each piece of information in the information set, first click rates respectively corresponding to a plurality of combinations when each piece of information is displayed in the page in the plurality of combinations of the display position and the display style.

In some embodiments, in step 103, based on the associated features of each piece of information in the information set, when it is determined that each piece of information is displayed in multiple combinations of display positions and display styles in the page, the first click rates respectively corresponding to the multiple combinations may be implemented by the following technical solutions: and performing full connection processing on the associated characteristics of each piece of information to obtain a plurality of first click rates which are in one-to-one correspondence with a plurality of combinations when each piece of information is displayed in the page in the plurality of combinations of the display position and the display style.

As an example, the fully-connected processing maps the learned distributed feature representations to the sample mark space, in actual use, the fully-connected core operation is that the matrix vector product y is Wx, a first click rate matrix is output through the fully-connected processing, the first click rate matrix is a three-dimensional matrix and corresponds to information, a display position and a display style respectively, each first click rate in the first click rate matrix represents a first click rate when corresponding information is displayed in a corresponding display style at the corresponding display position, for example, a first click rate when information 1 is displayed in a display style 3 at a display position 2, and when the display positions are 10, the display styles are 4, and the number of information in the information set is 10, there are 10 × 4 first click rates.

In step 104, the server determines an arrangement mode which enables the overall click rate of the page to be the highest based on the first click rates of the plurality of combinations corresponding to each piece of information, wherein the arrangement mode comprises a display position and a display style of each piece of information in the page.

Based on fig. 6A and fig. 6C, fig. 6C is a schematic flow chart of the information recommendation method based on artificial intelligence provided in the embodiment of the present invention, and the determining of the arrangement mode that enables the overall click rate of the page to be the highest based on the first click rates of the plurality of combinations respectively corresponding to each piece of information in step 104 may be implemented by steps 1041-.

In step 1041, for each piece of information in the set of information, the following processing is performed: binding the information with each display position respectively to obtain an information position binding result of the information corresponding to each display position; and each information position binding result has a first click rate respectively corresponding to a plurality of display styles.

In some embodiments, the information is respectively bound to each display position to obtain an information position binding result of the information corresponding to each display position, and the information binding result can be obtained by the following technical solutions: for information I in information set_nTo transmit information I_nAnd a display position P_qBinding to obtain an information position binding result R_nq(ii) a Wherein n is more than or equal to 1 and less than or equal to K, q is more than or equal to 1 and less than or equal to K, K is the number of information in the information set and is an integer more than or equal to 2, and J is the number of display styles and is an integer more than or equal to 2.

As an example, according to the first click rate matrix with the shape of [ k, k, s ] output in step 103, determining a final display position and a display style, where k is the number of information in the information set, k is also the number of display positions of the page, and s is the number of display styles, the display positions of the information are not repeatable, and the display styles of the information are repeatable, and binding the information and the display positions can obtain k × k information position binding results, for example, an information position binding result formed by information 1 and position 2.

In step 1042, for each information location binding result, the following processing is performed: and in the first click rate corresponding to the plurality of display styles, taking the display style corresponding to the highest first click rate as the display style corresponding to the information position binding result.

In some embodiments, the above-mentioned taking the presentation style corresponding to the highest first click rate among the first click rates corresponding to the multiple presentation styles as the presentation style corresponding to the information position binding result may be implemented by the following technical solutions: binding result R at information location_nqDetermining the display style with the highest first click rate as information I in the first click rates corresponding to the plurality of display styles_nIn the display position P_qThe display style used in the display.

In some embodiments, for an information position binding result formed by information 1 and position 2, there are s display styles, which correspond to s first click rates, and if s is 4, the display style corresponding to the highest first click rate is display style a, the display style a is determined as the display style of information 1 when the position 2 is displayed, that is, the display style with the highest first click rate at each display position of each piece of information is selected as the display style of the information at the position, so that the first click rate matrix with the size [ k, k, s ] is simplified into the first click rate matrix with the size [ k, k ].

As an example, the number of the presentation positions (position 1 and position 2) and the information (information 1 and information 2) is 2, the number of the presentation styles (style 1-style 4) is 4, the presentation style of the information 1 at the position 1 (selected from style 1-style 4), the presentation style of the information 1 at the position 2, the presentation style of the information 2 at the position 1, and the presentation style of the information 2 at the position 2 are respectively selected, a specific process of selecting the presentation style of the information 1 at the position 1 (selected from style 1-style 4) is as follows, a first click rate when the information 1 is presented in the presentation style 1 at the position 1 is P1, a first click rate when the information 1 is presented in the presentation style 2 at the position 1 is P2, a first click rate when the information 1 is presented in the presentation style 3 at the position 1 is P3, the first click rate of the information 1 displayed in the display style 4 at the position 1 is P4, and the display style corresponding to the highest value in P1-P4 is determined as the display style of the selected information 1 at the position 1.

In step 1043, a path search process is performed on the multiple information position binding results to obtain a path that maximizes the overall click rate of the page.

As an example, three methods may be used to derive the dimension [ k, k ] from]The display position of each piece of information is determined in the first click rate matrix, for example, a full arrangement method, a cluster search method and a greedy method, the full arrangement method can find the optimal click profit solution of the arrangement mode of the display positions of each piece of information, the click profit is the click rate of a page, for example, 10 pieces of information and 10 display positions exist in the page, if the arrangement mode is as follows, the information 1 is displayed at the position 1 (assuming that the determined display style 1 corresponds to the P in the first click rate matrix₁₁₁) …, message 10 is shown at position 10 (assuming that the determined presentation style 10, corresponding to P in the first click rate matrix₁₀₁₀₁₀) The click profit (page entire click rate) is P₁₁₁+，…，+P₁₀₁₀₁₀However, the time complexity of the full permutation method is a factorial of k, the time complexity is high, and a better benefit effect can be realized in an experiment within the allowable time complexity by adopting the cluster searching method.

In step 1044, the plurality of information location binding results and the corresponding presentation styles included in the path search processing result are determined as the arrangement mode that maximizes the overall click rate of the page.

In some embodiments, the above-mentioned performing path search processing on the multiple information position binding results to obtain a path that enables the overall click rate of the page to be the highest may be implemented by the following technical solutions: selecting m information sequences which enable the overall click rate of the first k display positions to be the highest from the information set to serve as candidate information sequences of the first k display positions; wherein m is the path search size of the path search processing, and the value range satisfies that m is more than or equal to 1 and less than or equal to K; k is an integer with the value increasing from 1, and the value range of K satisfies that K is more than or equal to 1 and is less than K; the information in the information sequence corresponds to the first k display positions one by one; and selecting the information sequence with the highest overall click rate from the candidate information sequences of the previous K display positions to serve as a path enabling the overall click rate of the page to be the highest.

In some embodiments, the selecting of the m information sequences with the highest overall click rate at the first k presentation positions from the information set to be recommended may be implemented by the following technical solutions: determining a first click rate of a corresponding display position in the front k-1 display positions aiming at each information in the m candidate information sequences of the front k-1 display positions; for each of the m candidate information sequences, performing the following: adding the first click rates of all information in the corresponding candidate information sequences to obtain the overall click rate of the candidate information sequences corresponding to the first k-1 display positions; acquiring a first click rate of each piece of residual information at a kth display position, wherein the residual information is information in the information set except information in the candidate information sequence; adding the first click rate of each piece of residual information at the kth display position with the integral click rate respectively to obtain integral click rates of a plurality of information sequences corresponding to the first k display positions; wherein the plurality of information sequences are combinations of the candidate information sequences and a plurality of the remaining information sequences, respectively, the remaining information being ordered after the candidate information sequence in each of the information sequences; and selecting m information sequences with the highest overall click rate ranking from a plurality of information sequences obtained based on the m candidate information sequences.

As an example, assuming there are 4 presentation positions and 4 information (A, B, C, D) needs to be presented, i.e. K4, the information presented at the first presentation position in the page is first determined, a is presented at the first presentation position with a size [ K, K ] of]The click rate corresponding to the first click rate matrix is P_1AB is presented at the first presentation position at a size of [ k, k ]]The click rate corresponding to the first click rate matrix is P_1BC is presented in the first presentation position at a size of [ k, k ]]The click rate corresponding to the first click rate matrix is P_1CD is presented at the first presentation position at a size [ k, k ]]Corresponding to the first click rate matrixClick rate of P_1DWherein P is_1BMaximum, P_1CNext, assuming that the path search size of the path search processing is 2, the situation corresponding to the 2 first click rates with the highest first click rate is selected from the four situations as the display result of the first display position, that is, the candidate information sequence on the first display position is two sequences of information B and information C, and then the information of the second display position is determined, where the second display position may be A, C and D when the first display position displays B, and the second display position may be A, B and D when the first display position displays C, so as to obtain the following sequences: BA (sum of click rates P1B + P2A), BC (sum of click rates P1B + P2C), BD (sum of click rates P1B + P2D), CA (sum of click rates P1C + P2A), CB (sum of click rates P1C + P2B), CD (sum of click rates P1C + P2D), 2 (path search size 2) results with the highest sum result (assuming that P1C + P2B is the highest and P1B + P2D times) are selected from the 6 sum results of click rates as candidate information sequences (CB and BD) for the first presentation position and the second presentation position, and then 2 (path search size 2) information sequences of the first 3 presentation position (CB and BD) for the first presentation position, the second presentation position and the third presentation position are determined in the same manner until the last 4 information sequences of the last presentation position are obtained, and the result and the highest respective sum results of click rates are selected as the last information sequences of the selection results and the highest click rates of the respective result information sequences, for example, the last sequence searched out is CBAD, i.e., the first, second, third and fourth presentation positions respectively present information C, B, A and D.

The optimal or approximately optimal arrangement mode can be obtained through the path searching method, the time complexity of the path searching method is relatively low compared with a full arrangement method, the accuracy of the path searching method is higher compared with a greedy method, and a better arrangement mode can be obtained.

In step 105, the server performs a recommendation operation based on the information set to which the arrangement mode is applied.

As an example, the arrangement mode with the highest overall click rate is obtained through step 104, where the arrangement mode is constrained, and the display position and the display style of each piece of information are displayed, and a recommendation operation is performed on the information set according to the constrained arrangement mode, so that the information in the information set is presented in the page according to the constrained arrangement mode.

Referring to fig. 6D, fig. 6D is a schematic flowchart of an artificial intelligence based information recommendation method according to an embodiment of the present invention, in step 201, a server receives an information recommendation request sent by a terminal; in step 202, the server obtains the characteristics of each piece of information in the information base to determine a corresponding second click rate; in step 203, performing descending sorting processing on the information base based on the second click rate of each piece of information, and selecting a plurality of pieces of information sorted at the front from the descending sorting result to form an information set to be recommended; in step 204, the server obtains the characteristics of each piece of information in the information set to be recommended; in step 205, performing attention coding processing on the features of each piece of information in the information set to be recommended to obtain a degree of association between the piece of information and each piece of information in the information set to be recommended; in step 206, determining the association characteristic of each piece of information based on the association degree between the information and each piece of information in the information set to be recommended; in step 207, the server determines, based on the associated features of each piece of information in the information set, first click rates respectively corresponding to a plurality of combinations when each piece of information is displayed in the page in the plurality of combinations of the display position and the display style; in step 208, the server selects a display style of each piece of information corresponding to the highest first click rate at each display position, and determines the selected display style as the display style of the corresponding piece of information displayed at the corresponding display position; in step 209, the server performs path search processing based on the first click rate on each piece of information and each display position to obtain a path search processing result with the highest overall click rate, which is used as information displayed at each display position; in step 210, the server determines the display position of each piece of information and the display style of the corresponding information displayed on the corresponding display position as the arrangement mode which enables the overall click rate of the page to be the highest; in step 211, the server executes a recommendation operation to the terminal based on the information set to which the arrangement mode is applied; in step 212, the information arranged in the arrangement is presented on the terminal used by the user.

An exemplary application of the artificial intelligence based information recommendation method provided by the embodiment of the invention in a practical application scenario will be described below.

For example, the artificial intelligence based information recommendation method provided by the embodiment of the invention can be applied to a rearrangement module of a recommendation system of a news client, after the recalled news is sorted for the first time by a fine ranking module of the recommendation system, N news are selected from high to low according to a second click rate obtained by predicting the recalled information, and then the artificial intelligence based information recommendation method provided by the embodiment of the invention is used for determining the display position and the display style of the news, and in addition, the artificial intelligence based information recommendation method can be adapted to any information stream and e-commerce sequence recommendation scenes.

Referring to fig. 7, fig. 7 is an overall architecture diagram of an artificial intelligence based information recommendation method provided in an embodiment of the present invention, which is divided into two parts in practical application, one part is offline model training, the other part is online rearrangement of a recommendation system, a user log is obtained from an online recommendation service, a training sample training click rate prediction model (neural network rearrangement model) is formed by display data, click data, behavior data, session data, user side features, information side features and environment side features in the user log, in a refinement module, a display position and a display style of news are unknown and cannot be added as features to a second click rate prediction, the refinement model can only predict a second click rate for news to be predicted individually, and factors influencing each other between the displayed news cannot be considered, and a click rate prediction model in the rearrangement module adopts a self-attention mechanism to acquire the correlation characteristics of the information displayed on the same page at the same time, an output layer of the click rate prediction model performs first click rate estimation on the combination of all display positions and display styles of each piece of news, and then searches the arrangement mode of the news display positions and the display styles with the largest number of the clicked news of the user according to a first click rate matrix.

Referring to fig. 3, the input X of the click-through rate prediction model is the maximum n features of K articles (information 1, information 2, …, information K) collected by the rearrangement module, and may be represented as X ═ X₁,x₂,…,x_i…,x_k}，x_iThe set of feature labels for the ith news item input to the click-through rate prediction model may be represented as x_i＝{f₁,f₂,…,f_jJ is less than or equal to n, and the random initialization size is [10000000,64 ] according to the size of the characteristic engineering hash]According to the input feature identifier, searching the corresponding feature vector from the feature vector matrix, which can be expressed as x_i＝{vec_f1，vec_f2,…,vec_fjThe formula for feature merging is

Referring to fig. 4, in the artificial intelligence based information recommendation method according to the embodiment of the present invention, when performing the self-attention process, Q, K and V matrices may be the same, that is, all the three matrices are output S after feature merging, that is, S ═ S { (S) }₁,s₂,…,s_kDetermine the associated features by the formula

Referring to fig. 5, the output layer output click rate prediction model has k information amount predicted each time, k display position amount, and s display styles of information, so that the output first click rate amount for each news is k × s, the first click rate amount for all news in one recommendation is k × s, at the stage of click rate model training, each training sample only performs model optimization for part of output nodes, that is, optimization is performed only based on one real arrangement mode, the output of output layer nodes is adjusted to a matrix of 10 × 4(k × 10, s 4) size, corresponding nodes are obtained according to the determined display styles of the training samples, and the matrix size is reduced to 10 × 10, howeverAnd performing model optimization on nodes effective on the diagonal line, wherein all parameters in the click rate prediction model are trained because the previous layer of the output layer is a full-connection layer, and in the refined model, the cross entropy loss is generally used as a loss function formula (2) of the refined model:

wherein, y_iFor the ith news tag in the sample, the clicked tag is 1, the unchecked tag is 0, p_iFor the second click rate of the ith news in the sample, in the click rate prediction model of the rearrangement module, the input k articles have k labels, and the output k × s first click rates are output, so that a loss function is calculated only for output nodes corresponding to the display positions and the display styles of each article in the real sample, in addition, the mean square error loss of the whole real clicks and the whole predicted clicks of the k news in each sample is also added into the loss function for optimization, and a loss function formula (3) of the click rate prediction model of the rearrangement module is as follows:

referring to fig. 8, fig. 8 is a schematic diagram of path search optimization of an artificial intelligence based information recommendation method provided in an embodiment of the present invention, in an online rearrangement module, according to a full matrix with a size [ k, k, s ] output by a click rate prediction model, a final display position and a final display style of each piece of information are determined, the display position of news is unrepeatable, and the display style of news is repeatable, so that a first highest click rate display style of each article at each display position is first selected as a display style of the article at the display position, and thus a first click rate matrix (full matrix AllMatrix) with a size [ k, k, s ] is simplified into a matrix with a size [ k, k ] (information position matrix PosMatrix), and a simplified formula is as follows: the PosMatrix [ i ] [ j ] ═ max (AllMatrix [ i ] [ j ]), then three methods can be used to determine the presentation position of each news from the PosMatrix, which are respectively the full permutation method, the cluster search method and the greedy method, as shown in fig. 8, in the full permutation method, each of which is a full search calculation performed in the path search process for each position, for example, in the case of 10 information, 10 information positions, the click rate of 10 information at position 1 is first calculated, then in the path search for position 2, if the information at position 1 is 1, 9 information sequences exist at position 2, if the information at position 1 is 1, 9 information sequences exist, and so on, 10 information sequences exist for position 1 and position 2, and then a path search is performed for the 3 rd position based on this, which is equivalent to setting the path search size to the number of all information, that is, the full ranking process shown in fig. 8 is implemented, the full ranking method may find the optimal solution of click rate profit of the news display position and the display style, for example, the optimal solution may be the maximized optimal solution of the sum of click rates of a plurality of articles recommended at a time, but the time complexity is a factorial of k, and if 10 pieces of news are recommended, 3628800 times need to be calculated, which results in a large time consumption, so a cluster search method and a greedy method may be adopted to solve the approximately optimal solution, wherein the cluster search method may find the optimal solution better than or equal to the greedy method within the allowed time complexity, and has a better profit effect in experiments and online applications.

The artificial intelligence based information recommendation method provided by the embodiment of the invention is used for carrying out experiments and full release on the headline channel and the recommendation channel of the news client, and the income effects of the two channels are respectively as follows: in the headline channel, the long click of the per-user list is improved by 2.45 percent (significant), the time of the per-user channel is improved by 1.32 percent (insignificant), the time of the per-user dwell of the client participating user is improved by 0.93 percent (significant), the next day dwell of the client participating user is improved by 0.10 percent (insignificant), the overall user per-user time of the client is improved by 0.73 percent (significant), the total click rate is improved by 1.43 percent (significant), the total click page browsing of the per-user is improved by 1.28 percent (significant), in the recommended channel, the long click of the per-user list is improved by 1.2 percent (significant), the time of the per-user channel is improved by 3.66 percent (significant), the time of the per-user dwell of the client participating user is improved by 2.00 percent (significant), the overall user per-user time of the client is improved by 0.31 percent (significant), the per-user click is improved by 1, the real exposure of independent visitors on the bottom layer page is improved by 3.42 percent (remarkable), and the average video playing time of the bottom layer page is improved by 3.47 percent (remarkable).

Continuing with the exemplary structure of the artificial intelligence based information recommender 255 as implemented as a software module provided in the present invention, in some embodiments, as shown in FIG. 2, the software modules stored in the artificial intelligence based information recommender 255 of the memory 250 may include: the association processing module 2552 is configured to perform association degree processing on the feature of each piece of information based on the association degrees among the pieces of information in the information set to obtain an association feature of each piece of information; the click rate determining module 2553 is configured to determine, based on the associated features of each piece of information in the information set, first click rates respectively corresponding to a plurality of combinations when each piece of information is displayed in the page in the plurality of combinations of the display position and the display style; an arrangement mode determining module 2554, configured to determine, based on the first click rates of the multiple combinations respectively corresponding to each piece of information, an arrangement mode that maximizes the overall click rate of the page, where the arrangement mode includes a display position and a display style of each piece of information in the page; and the recommending module 2555 is used for executing recommending operation based on the information set with the arrangement mode.

In some embodiments, before obtaining the features of each information in the information set to be recommended, the click-through rate determining module 2553 is further configured to: acquiring the characteristics of each piece of information in the information base to determine a corresponding second click rate; performing descending sorting processing on the information base based on the second click rate of each piece of information, and selecting a plurality of pieces of information sorted at the front from the descending sorting result to form an information set to be recommended; and the number of the information in the information set to be recommended is the same as the number of the display positions in the page.

In some embodiments, the apparatus further comprises: a feature acquisition module 2551 to: executing the following processing for each information in the information set to be recommended: inquiring a feature vector corresponding to the feature of the information from a pre-established feature vector matrix; and performing fusion processing on the feature vector of each piece of information to obtain the feature corresponding to each piece of information.

In some embodiments, association processing module 2552 is further configured to: executing the following processing for each information in the information set to be recommended: performing attention coding processing on the characteristics of each piece of information in the information set to be recommended to obtain the association degree between the information and each piece of information in the information set to be recommended; and determining the association characteristics of each piece of information based on the association degree between the information and each piece of information in the information set to be recommended.

In some embodiments, association processing module 2552 is further configured to: carrying out linear transformation processing on the characteristics of the information to obtain a query vector, a key vector and a value vector which respectively correspond to the characteristics; performing point multiplication on the query vector of the information and the key vector of each information in the information set to be recommended, and performing normalization processing on the result of the point multiplication processing based on a maximum likelihood function to obtain the association degree between the information and each information in the information set to be recommended; determining the degree of association as an attention weight of a value vector corresponding to each piece of information; and carrying out weighting processing on the value vector based on the attention weight to obtain the relevant characteristics of the information based on the attention coding processing.

In some embodiments, the click rate determination module 2553 is further configured to: and performing full connection processing on the associated characteristics of each piece of information to obtain a plurality of first click rates which are in one-to-one correspondence with a plurality of combinations when each piece of information is displayed in the page in the plurality of combinations of the display position and the display style.

In some embodiments, the arrangement determination module 2554 is further configured to: for each information in the set of information, performing the following: binding the information with each display position respectively to obtain an information position binding result of the information corresponding to each display position; each information position binding result has a first click rate respectively corresponding to a plurality of display styles; for each information location binding result, the following processing is performed: in the first click rate corresponding to the plurality of display styles, the display style corresponding to the highest first click rate is used as the display style corresponding to the information position binding result; performing path search processing on the plurality of information position binding results to obtain a path which enables the overall click rate of the page to be the highest; and determining a plurality of information position binding results and respectively corresponding display styles included in the path search processing result as an arrangement mode which enables the overall click rate of the page to be the highest.

In some embodiments, the arrangement determination module 2554 is further configured to: for information I in information set_nTo transmit information I_nAnd a display position P_qBinding to obtain an information position binding result R_nq(ii) a N is more than or equal to 1 and less than or equal to K, q is more than or equal to 1 and less than or equal to K, K is the number of information in the information set and is an integer more than or equal to 2, and J is the number of display styles and is an integer more than or equal to 2; binding result R at information location_nqDetermining the display style with the highest first click rate as information I in the first click rates corresponding to the plurality of display styles_nIn the display position P_qThe display style used in the display.

In some embodiments, the arrangement determination module 2554 is further configured to: selecting m information sequences which enable the overall click rate of the first k display positions to be the highest from the information set to serve as candidate information sequences of the first k display positions; wherein m is a path search size of the path search processing, and the value range satisfies that m is more than or equal to 1 and less than or equal to K; k is an integer with the value increasing from 1, and the value range of K satisfies that K is more than or equal to 1 and is less than K; information in the information sequence corresponds to the first k display positions one by one; and selecting the information sequence with the highest overall click rate from the candidate information sequences of the previous K display positions to serve as a path enabling the overall click rate of the page to be the highest.

In some embodiments, the arrangement determination module 2554 is further configured to: determining a first click rate of a corresponding display position in the front k-1 display positions aiming at each information in the m candidate information sequences of the front k-1 display positions;

for each of the m candidate information sequences, performing the following:

adding the first click rates corresponding to each information in the candidate information sequence to obtain the whole click rate of the candidate information sequence corresponding to the first k-1 display positions; acquiring a first click rate of each piece of residual information at a kth display position, wherein the residual information is information in the information set except information in the candidate information sequence; adding the first click rate of each piece of residual information at the kth display position with the integral click rate respectively to obtain integral click rates of a plurality of information sequences corresponding to the first k display positions; wherein the plurality of information sequences are combinations of the candidate information sequences and a plurality of the remaining information sequences, respectively, the remaining information being ordered after the candidate information sequence in each of the information sequences; and selecting m information sequences with the highest overall click rate ranking from a plurality of information sequences obtained based on the m candidate information sequences.

In some embodiments, the characteristic of the information comprises at least one of: basic attribute characteristics used for representing basic information of a user to be recommended; the interest label characteristics are used for representing interest preferences of the user to be recommended; the environment characteristics are used for representing a recommendation environment for recommending information to a user to be recommended; a category feature for characterizing a category of information; source characteristics for characterizing a source of information; content characteristics for characterizing the content of the information.

In some embodiments, a first click rate of each information in the information set to be recommended displayed in each display style at each display position is obtained by calling a click rate prediction model, and the click rate prediction model comprises an attention mechanism structure; the apparatus 255 further comprises: a training module 2556 to: generating a training sample set for training a click rate prediction model before acquiring the characteristics of each piece of information in an information set to be recommended; each training data sample in the training sample set comprises a sample information sequence, and the information with the preset number ranked in the top in the sample information sequence is effective sample information; performing attention coding processing on the characteristics of each sample information in the sample information sequence to obtain the association degree between the sample information and each effective sample information in the sample information sequence; determining the correlation characteristics of each sample information based on the correlation degree between the sample information and each effective sample information in the sample information sequence; carrying out forward propagation on the associated characteristics of each sample information in each training data sample in a click rate prediction model to obtain a predicted first click rate of each sample information of each training data sample displayed in a real display mode at a real display position; and performing back propagation in the click rate prediction model based on the difference between the predicted first click rate and the real first click rate, and updating the parameters of the click rate prediction model in the process of back propagation.

Embodiments of the present invention provide a computer-readable storage medium storing executable instructions, which when executed by a processor, cause the processor to perform a method provided by embodiments of the present invention, for example, as shown in fig. 6A-6D, which illustrate an artificial intelligence based information recommendation method.

In some embodiments, the computer-readable storage medium may be memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash, magnetic surface memory, optical disk, or CD-ROM; or may be various devices including one or any combination of the above memories.

In some embodiments, executable instructions may be written in any form of programming language (including compiled or interpreted languages), in the form of programs, software modules, scripts or code, and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

By way of example, executable instructions may correspond, but do not necessarily have to correspond, to files in a file system, and may be stored in a portion of a file that holds other programs or data, such as in one or more scripts in a hypertext markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files that store one or more modules, sub-programs, or portions of code).

By way of example, executable instructions may be deployed to be executed on one computing device or on multiple computing devices at one site or distributed across multiple sites and interconnected by a communication network.

In summary, through the embodiments of the present invention, the features of each piece of information are associated, so as to obtain associated features capable of representing associated influences among the pieces of information, and the predicted click rate of each piece of information at each position and presented in each pattern is determined through the associated features in a quantitative manner, so as to obtain the optimal arrangement manner of the information presentation patterns and positions, so that the index of the recommendation system is increased in a forward direction.

The above description is only an example of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, and improvement made within the spirit and scope of the present invention are included in the protection scope of the present invention.

Claims

1. An artificial intelligence based information recommendation method, characterized in that the method comprises:

based on the relevance among a plurality of pieces of information in the information set, carrying out relevance processing on the characteristics of each piece of information to obtain the relevance characteristics of each piece of information;

2. The method of claim 1, wherein prior to obtaining the characteristics of each information in the set of information to be recommended, the method further comprises:

3. The method of claim 1, further comprising:

performing the following for each information in the set of information:

inquiring a characteristic vector corresponding to the characteristic data of the information from a pre-established characteristic vector matrix;

4. The method according to claim 1, wherein the associating the feature of each information based on the degree of association between the plurality of information in the information set to obtain the associated feature of each information comprises:

performing the following for each information in the set of information:

5. The method of claim 4, wherein said performing attention coding on the feature of each information in the information set to obtain the degree of association between the information and each information in the information set comprises:

the determining the association characteristic of each information based on the association degree between the information and each information in the information set comprises:

6. The method of claim 1,

the determining, based on the associated features of each piece of information in the information set, a first click rate corresponding to each of a plurality of combinations of a display position and a display style when each piece of information is displayed in the page in the combination, includes:

7. The method according to claim 1, wherein the determining the arrangement mode that maximizes the overall click rate of the pages based on the first click rates of the combinations respectively corresponding to each piece of information comprises:

for each information in the set of information, performing the following:

8. The method of claim 7,

the binding the information with each display position respectively to obtain the information position binding result of the information corresponding to each display position comprises:

the step of taking the presentation style corresponding to the highest first click rate as the presentation style corresponding to the information position binding result in the first click rates corresponding to the plurality of presentation styles includes:

binding the result R at the information location_nqIn a first click rate corresponding to a plurality of presentation stylesDetermining the display style with the highest first click rate as the information I_nAt the display position P_qThe display style used in the display.

9. The method of claim 7, wherein the performing a path search process on the plurality of information location binding results to obtain a path that maximizes the overall click-through rate of the page comprises:

10. The method of claim 9, wherein said selecting m information sequences from said information set that maximize the overall click-through rate for the top k presentation positions comprises:

for each of the m candidate information sequences, performing the following:

11. The method of claim 1,

the characteristic of the information includes at least one of:

12. The method according to any one of claims 1 to 11,

the first click rate of each piece of information in the information set displayed in different display styles at each display position is determined by calling a click rate prediction model, and the click rate prediction model comprises an attention coding structure;

before obtaining the characteristics of each information in the information set to be recommended, the method further comprises the following steps:

generating a training sample set for training the click rate prediction model;

13. An artificial intelligence-based information recommendation device, comprising:

the characteristic acquisition module is used for acquiring the characteristics of each piece of information in the information set to be recommended;

14. An electronic device, comprising:

a memory for storing executable instructions;

a processor for implementing the artificial intelligence based information recommendation method of any one of claims 1 to 12 when executing the executable instructions stored in the memory.

15. A computer-readable storage medium storing executable instructions for implementing the artificial intelligence based information recommendation method of any one of claims 1 to 12 when executed by a processor.