CN111159242A - Client reordering method and system based on edge calculation - Google Patents

Client reordering method and system based on edge calculation Download PDF

Info

Publication number
CN111159242A
CN111159242A CN201911390108.5A CN201911390108A CN111159242A CN 111159242 A CN111159242 A CN 111159242A CN 201911390108 A CN201911390108 A CN 201911390108A CN 111159242 A CN111159242 A CN 111159242A
Authority
CN
China
Prior art keywords
information
client
attention
lstm
click
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201911390108.5A
Other languages
Chinese (zh)
Other versions
CN111159242B (en
Inventor
范俊
顾湘余
李文杰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hangzhou Quwei Science & Technology Co ltd
Original Assignee
Hangzhou Quwei Science & Technology Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hangzhou Quwei Science & Technology Co ltd filed Critical Hangzhou Quwei Science & Technology Co ltd
Priority to CN201911390108.5A priority Critical patent/CN111159242B/en
Publication of CN111159242A publication Critical patent/CN111159242A/en
Application granted granted Critical
Publication of CN111159242B publication Critical patent/CN111159242B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24—Querying
    • G06F16/245—Query processing
    • G06F16/2457—Query processing with adaptation to user needs
    • G06F16/24578—Query processing with adaptation to user needs using ranking
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention discloses a client reordering method and a system based on edge calculation, wherein the reordering method comprises the following steps: s1, acquiring a first information sequence which is recommended by a server and comprises k pieces of information; s2, randomly arranging the K information to obtain K! Sorting the second information; s3, extracting the recommended information in each second information sequence and the characteristics of the context information thereof; and S4, sequentially inputting the characteristics of each second information sequence into the Attention-LSTM-MLP model to obtain corresponding click through rate estimated values, and selecting the second information sequence corresponding to the highest click through rate estimated value for display. The invention fully considers the display characteristics of each client, reorders the sequence at the client, and improves the user experience and the recommendation effect.

Description

Client reordering method and system based on edge calculation
Technical Field
The invention relates to the field of recommendation sequencing, in particular to a client side reordering method and system based on edge calculation.
Background
The ranking To Rank (LTR) is mainly used in the field of information retrieval, and ranks search results by comprehensively considering a plurality of ranking features. LTR is a widely and extensively studied problem, whether in the search, advertising, or recommendation fields. Since the LTR in most cases targets the Click Through Rate, the LTR problem is also referred to as Click Through Rate (CTR) prediction, which refers to the Click Through Rate of the web advertisement (photo advertisement/text advertisement/keyword advertisement/ranked advertisement/information advertisement, etc.), i.e. the actual number of clicks of the advertisement (strictly speaking, the number of pages to target) is divided by the advertisement presentation amount (Show content).
The general practice of the recommendation scenario is: the application requests a recommendation list from the server; the server side recalls and sorts the information according to the user characteristics, the information characteristics and the context characteristics, returns the information to the application side, sorts the information from high to low according to the click rate, and displays the information to the user according to the sequence of the list. Whether the recommendation information is clicked or not is closely related to the order and position when presented to the user. Thus, LTRs are generally classified into the following three types: pointwise, the algorithm is simple, provided that whether each piece of recommended information is clicked is only relevant to itself, and other information around the information is not considered; the pairwise optimizes the partial order relation of the click probabilities of different pieces of information, and considers the relation between each piece of information and other pieces of information in the list; listwise, the global sequence of the whole list is optimized, and the defect is that the algorithm complexity is high. In the existing recommendation environment, considering the requirement of online on algorithm performance, most LTRs adopt pointwise algorithm, and the commonly used algorithm includes: logistic version, gbdt, factitious machine, and dnn, which are currently more popular. However, for the poitwise algorithm, documents of the same class cannot be ranked; entirely from the classification point of view of a single document, without taking into account the relative order between the documents.
Further, as previously mentioned, whether the recommendation information is clicked, in addition to meeting the user interest preferences, is also location dependent. The same recommendation list shows different styles on different devices. As shown in fig. 1, the same recommendation list is presented on different devices, giving the user a completely different visual experience and experience. The same recommended video list is displayed on the equipment on the left side, and videos of different categories are arranged at intervals, so that the people feel more diversified. And the videos of the same type are gathered on one side to influence the user experience and further influence the click when the videos are displayed on the equipment on the right side. However, the presentation form of the recommendation result cannot be determined by the server, especially some H5 pages, which allow the user to zoom the display window on the server at any time, so that it is very important to perform personalized reordering on the recommendation list on the client. However, almost all the existing CTR predictions are completed at the server, so that the advantage of strong computing power of the server is fully utilized, and the same recommendation ranking is adopted for different clients without considering the influence of different display forms of the different clients.
Therefore, how to implement the information ordering adapted to the display form of each client is a problem to be solved in the field.
Disclosure of Invention
The invention aims to provide a client reordering method and a client reordering system based on edge calculation aiming at the defects of the prior art. The invention fully considers the display characteristics of each client, reorders the sequence at the client, and improves the user experience and the recommendation effect.
In order to achieve the purpose, the invention adopts the following technical scheme:
a client reordering method based on edge calculation comprises the following steps:
s1, acquiring a first information sequence which is recommended by a server and comprises k pieces of information;
s2, randomly arranging the K information to obtain K! Sorting the second information;
s3, extracting the recommended information in each second information sequence and the characteristics of the context information thereof;
and S4, sequentially inputting the characteristics of each second information sequence into the Attention-LSTM-MLP model to obtain corresponding click through rate estimated values, and selecting the second information sequence corresponding to the highest click through rate estimated value for display.
Further, the characteristics of the recommendation information include: the click through rate estimation result output by the information semantic vector, the information category and the server side sequencing model; the characteristics of the context information include information surrounding the currently recommended information in the current user interface presentation form.
Further, the Attention-LSTM-MLP model is generated at the server and sent to the client.
Further, the generation of the Attention-LSTM-MLP model specifically includes:
s41, constructing an Attention-LSTM-MLP network;
s42, extracting the characteristics of the recommendation information and the context information thereof, calculating the click through rate according to the user click number of the recommendation information, and training the Attention-LSTM-MLP network based on the characteristics and the click through rate.
Further, the Attention-LSTM-MLP network is specifically:
an Attention network is connected behind an implicit layer in the LSTM network; the hidden layer of the LSTM network outputs the extracted features and inputs the features into the Attention network; the Attention network is converted into a weight coefficient of each node through a softmax function, the value of each node in the Attention network is multiplied by the weight coefficient to serve as the output of the node, the final Attention-LSTM coding output is obtained, and the coding output is input into an MLP network to obtain the click through rate pre-estimated value.
Further, the loss function of the Attention-LSTM-MLP network is:
Figure BDA0002342023030000031
wherein y isiIs the calculated true click through rate, piIs a predicted value calculated according to the Attention-LSTM-MLP network, n is the number of samples, yiThe specific calculation is as follows:
Figure BDA0002342023030000032
where num is the number of information clicked by the user in the recommended information sequence that has been displayed and exposed at one time.
Further, the characteristics of the context information are set to 8 positions, and missing ones are filled with a default value of 0.
The invention also provides a client reordering system based on edge calculation, which comprises a client and specifically comprises the following components:
the acquisition module is used for acquiring a first information sequence which is recommended by the server and comprises k pieces of information;
a random ordering module for randomly arranging the K pieces of information to obtain K! Sorting the second information;
the extraction module is used for extracting the recommended information in each second information sequence and the characteristics of the context information of the recommended information;
and the estimation module is used for sequentially inputting the characteristics of each second information sequence into the Attention-LSTM-MLP model to obtain a corresponding click through rate estimated value, and selecting the second information sequence corresponding to the highest click through rate estimated value for display.
Further, the reordering system further includes a server, and specifically includes:
the building module is used for building an Attention-LSTM-MLP network;
and the training module is used for extracting the characteristics of the recommendation information and the context information thereof, calculating the click through rate according to the user click number of the recommendation information, and training the Attention-LSTM-MLP network based on the characteristics and the click through rate.
Further, the characteristics of the recommendation information include: the click through rate estimation result output by the information semantic vector, the information category and the server side sequencing model; the characteristics of the context information include information surrounding the currently recommended information in the current user interface presentation form.
Compared with the prior art, the invention has the following effects:
(1) according to the method, the recommendation results are reordered on the client side by combining the display form of a specific client side, the display characteristics of the client side are fully considered, so that the ordering of the same server side results on different client sides can be different, and the user experience and the recommendation effect are improved;
(2) the invention adopts the listwise algorithm to carry out global target optimization on the whole list, and because candidate information to be ordered reordered on the end is very small, the additional cost introduced on the basis of the optimized ordering by adopting the listwise algorithm is small;
(3) in view of the fact that large-scale users and information characteristics cannot be transmitted to the client, the method utilizes fewer characteristics, mainly influences the characteristics of user vision and context characteristics, and further optimizes the sequencing performance;
(4) the invention learns the preferences of human brain on the dispersion and aggregation of different contents by adding an attention mechanism, reduces the data processing amount in the reordering process and improves the data processing efficiency;
(5) the invention fully utilizes the characteristics of the server and the client, trains and generates the Attention-LSTM-MLP model at the server and sends the model to the client, thereby reducing the burden of the client.
Drawings
FIG. 1 is an example of the display of recommendation information on different devices;
FIG. 2 is a flowchart illustrating a method for reordering clients based on edge computation according to an embodiment;
FIG. 3 is a comparison graph of the display of recommendation information on a web page and a mobile terminal;
FIG. 4 is a diagram of the Attention-LSTM-MLP network architecture;
fig. 5 is a block diagram of a client reordering system based on edge computation according to the second embodiment.
Detailed Description
The embodiments of the present invention are described below with reference to specific embodiments, and other advantages and effects of the present invention will be easily understood by those skilled in the art from the disclosure of the present specification. The invention is capable of other and different embodiments and of being practiced or of being carried out in various ways, and its several details are capable of modification in various respects, all without departing from the spirit and scope of the present invention. It is to be noted that the features in the following embodiments and examples may be combined with each other without conflict.
It should be noted that the drawings provided in the following embodiments are only for illustrating the basic idea of the present invention, and the components related to the present invention are only shown in the drawings rather than drawn according to the number, shape and size of the components in actual implementation, and the type, quantity and proportion of the components in actual implementation may be changed freely, and the layout of the components may be more complicated.
The invention is further described with reference to the following drawings and specific examples, which are not intended to be limiting.
Example one
As shown in fig. 2, the present embodiment provides a client reordering method based on edge calculation, including:
s1, acquiring a first information sequence which is recommended by a server and comprises k pieces of information;
in the information recommendation field, the server side recalls and sorts information according to the user characteristics, the recommendation information characteristics and the context characteristics, and returns the information to the client side, wherein the information is sorted from high to low according to the click rate. In order to adapt to the display characteristics of different clients and realize the optimal information sequencing aiming at each client, the received recommended information sequencing is reordered at the client.
And aiming at the information sequencing of different clients, the display characteristics of the clients need to be considered, and corresponding sequencing is generated for each client. The server is connected to a large number of clients, which is undoubtedly a huge burden if the corresponding ordering is customized for each client. With the popularization of mobile devices and the development of the future internet of things, terminal devices are becoming more diversified, and the performance on a client is stronger, so that the migration of part of computing tasks from a server to a mobile terminal is possible. Therefore, the client is used as the edge node, and the recommendation information reordering is realized by utilizing edge calculation. The edge calculation is a distributed calculation structure, which moves the calculation of application program, data and service from the central node of the network to the edge node of the network logic for processing.
Specifically, after receiving the information sequence recommended by the server, the client does not immediately display the information sequence, but intercepts the information for further reordering.
S2, randomly arranging the K information to obtain K! Sorting the second information;
in order to comprehensively consider the mutual influence of the ordering among all information and overcome the defects of the existing pointwise and pairwise, the invention adopts a listwise algorithm to order the information so as to carry out global target optimization on the whole list. Because the candidate information to be sorted reordered on the client is very small (generally 6-12), the listwise algorithm can be adopted, and the obvious sorting performance reduction can not occur.
Specifically, for sorting information returned by a server arriving at a client, a sequence of length K, such as a click-through rate, randomly arranges all information in the sequence, and the number of rearrangements available is K! .
S3, extracting the recommended information in each second information sequence and the characteristics of the context information thereof;
the present invention considers the position relationship between information in addition to the characteristics of the recommendation information itself. And the second information ordering of each candidate is different in style presented on different clients. Therefore, the invention obtains K! After the second information is sequenced, the context information of each piece of recommended information is obtained according to the display style of the client.
In the ranking model of the server, in order to improve the learning ability of the model, the final feature dimension may be billion dimensions. The characteristic dimension of the client is not too high, limited by the computing power and the amount of data transmitted by the network. In view of the fact that large-scale users and recommended information features cannot be transmitted to the client, the present invention utilizes fewer features, mainly features affecting the vision of the users, and context (context) features. The recommended information features used in the re-ranking of the present invention include: information semantic vector, information category, CTR of information. The characteristics of the context information include information surrounding the current information item. The method specifically comprises the following steps:
(1) information semantic vector: the floating point unit vector with 128 dimensions of dimensionality is calculated by the server and is transmitted to the client;
(2) information type: the discrete type characteristic is transmitted to the client by the server, and is represented by one-hot;
(3) CTR of information: the CTR estimation result output by the server side sequencing model;
(4) context information: information surrounding the current information item in the current user interface presentation form.
As shown in fig. 3, the left image is typical context information under a web page, the middle dark color is a target information item, and the surrounding 8 light colored squares represent its context information item; the right diagram is typical context information on a handset device. In order to unify feature dimensions, the context features of the present invention are all set to 8 positions, and all missing ones are filled with a default value of 0. For example, for the dark target information in the right image, the left information entry is missing due to the display limitation of the client, and the present invention adopts 0 to represent the corresponding left information.
And S4, sequentially inputting the characteristics of each second information sequence into the Attention-LSTM-MLP model to obtain corresponding click through rate estimated values, and selecting the second information sequence corresponding to the highest click through rate estimated value for display.
The recommendation information returned by the server is reordered at the client, and the click through rate pre-estimated value is estimated through an Attention-LSTM-MLP model generated by training. Specifically, the invention adopts a neural network to encode the returned information sequence, wherein RNN is a recurrent neural network, and the neural network is a general term of a series of neural networks capable of processing sequence data. A Long Short Term Memory (LSTM) network is a specific form of RNN, and shows strong adaptability in time sequence data analysis, hidden layer neurons of the LSTM network are connected, the LSTM network can memorize previous information and apply the previous information to current calculation output, and the Long-term dependence problem in time sequence information is solved. Thus, the present invention utilizes an LSTM neural network.
The features of the second information ordering are encoded using LSTM units that learn the input sequence and encode it as a fixed length vector representation. Thus, for input sequences of shorter length, the LSTM neural network is able to learn a corresponding reasonable vector representation. However, it is difficult for the LSTM neural network to learn a reasonable vector representation when the input sequence is very long. Based on this, the present invention introduces an Attention (Attention) mechanism that enables the model to view different parts of the data in a targeted manner. An attention mechanism is added to learn the preferences of the human brain for different content dispersion and aggregation.
The Attention mechanism is implemented by retaining intermediate output results of the LSTM neural network on input sequences, then training a model to selectively learn these inputs and associate the output sequences with them as the model is output.
After the characteristics of the second information ordering are coded through the Attention-LSTM network, the input codes are subjected to click through rate prediction by adopting a Multi-Layer Perceptron (MLP), and the codes are used as input to obtain a score through an MLP network. Finally, we will choose the highest scoring sequence as the final result after reordering.
It is noted that the execution subjects of the above steps S1-S4 are clients. The training of the Attention-LSTM-MLP model needs to process a large amount of data, the resource limit of a client is considered, and the Attention-LSTM-MLP model is suitable for a large amount of clients, so the training of the whole model of the invention is carried out at a server. After training is finished, the model is transmitted to the client, and click through rate prediction is carried out on the client. The specific steps for generating the Attention-LSTM-MLP model are as follows:
s41, constructing an Attention-LSTM-MLP network;
as shown in FIG. 4, the present invention pre-joins an Attention network after the hidden layer in the conventional LSTM network. The LSTM network encodes each second information ordered feature and converts the second information ordered feature into a high-level feature, and the hidden-layer output vector of the LSTM network is the extracted feature hi. The Attention network is converted into a weight coefficient w of each node through a softmax function in the forward calculation processiMultiplying the value of each node in the attention network by the weight coefficient to obtain the output of the nodeAnd (6) discharging. The influence degree of each dimension in the hidden layer on the result is judged by adding an attention network, and the larger the weight coefficient is, the larger the influence on the result is, so that the network is focused on the change of one or more dimensions. And obtaining the final encoded output y of the Attention-LSTM through the output of each node.
The encoding output y of Attention-LSTM is input to a MLP network, which is an artificial neural network of forward structure that maps a set of input vectors to a set of output vectors. An MLP can be viewed as a directed graph, consisting of multiple layers of nodes, each layer being fully connected to the next. Each node, except the input nodes, is a neuron (or processing unit) with a nonlinear activation function. The present invention adopts the conventional MLP network structure, which is not described herein again.
S42, extracting the characteristics of the recommendation information and the context information thereof, calculating the click through rate according to the user click number of the recommendation information, and training the Attention-LSTM-MLP network based on the characteristics and the click through rate.
After the Attention-LSTM-MLP network is constructed, the invention trains the network by adopting a large amount of sample data. As with the sequences to be sorted, for each sequence of recommendation information, the characteristics of the recommendation information and its context information are extracted. And the sample data is a displayed and exposed recommended information sequence, and the actual user click number of the user is obtained, so that the click passing rate is calculated.
LTR belongs to one of supervised learning, assuming that the training sample is represented as { (x)i，yi),i∈[1,n]}. Here, unlike pointwise algorithms, xiNot characteristic of an item of information, but xiAnd its contextual characteristics. And click through rate yiCalculated using the following formula:
Figure BDA0002342023030000091
where num is the number of information clicked by the user in the recommended information sequence that has been displayed and exposed at one time. The range of values of this function is [0,1 ].
For the loss function, the invention uses the root mean square error, the formula is as follows:
Figure BDA0002342023030000092
wherein y isiIs the calculated true click through rate, piIs a predicted value calculated according to the Attention-LSTM-MLP network, and n is the number of samples.
The invention inputs the sample data into the Attention-LSTM-MLP network, optimizes the Attention-LSTM-MLP model by calculating the loss function of the whole reordering model, and trains to generate the Attention-LSTM-MLP model.
Example two
As shown in fig. 5, the present embodiment provides a client reordering system based on edge calculation, where the client specifically includes:
the acquisition module is used for acquiring a first information sequence which is recommended by the server and comprises k pieces of information;
in the information recommendation field, the server side recalls and sorts information according to the user characteristics, the recommendation information characteristics and the context characteristics, and returns the information to the client side, wherein the information is sorted from high to low according to the click rate. In order to adapt to the display characteristics of different clients and realize the optimal information sequencing aiming at each client, the received recommended information sequencing is reordered at the client.
And aiming at the information sequencing of different clients, the display characteristics of the clients need to be considered, and corresponding sequencing is generated for each client. The server is connected to a large number of clients, which is undoubtedly a huge burden if the corresponding ordering is customized for each client. With the popularization of mobile devices and the development of the future internet of things, terminal devices are becoming more diversified, and the performance on a client is stronger, so that the migration of part of computing tasks from a server to a mobile terminal is possible. Therefore, the client is used as the edge node, and the recommendation information reordering is realized by utilizing edge calculation. The edge calculation is a distributed calculation structure, which moves the calculation of application program, data and service from the central node of the network to the edge node of the network logic for processing.
Specifically, after receiving the information sequence recommended by the server, the client does not immediately display the information sequence, but intercepts the information for further reordering.
A random ordering module for randomly arranging the K pieces of information to obtain K! Sorting the second information;
in order to comprehensively consider the mutual influence of the ordering among all information and overcome the defects of the existing pointwise and pairwise, the invention adopts a listwise algorithm to order the information so as to carry out global target optimization on the whole list. Because the candidate information to be sorted reordered on the client is very small (generally 6-12), the listwise algorithm can be adopted, and the obvious sorting performance reduction can not occur.
Specifically, for sorting information returned by a server arriving at a client, a sequence of length K, such as a click-through rate, randomly arranges all information in the sequence, and the number of rearrangements available is K! .
The extraction module is used for extracting the recommended information in each second information sequence and the characteristics of the context information of the recommended information;
the present invention considers the position relationship between information in addition to the characteristics of the recommendation information itself. And the second information ordering of each candidate is different in style presented on different clients. Therefore, the invention obtains K! After the second information is sequenced, the context information of each piece of recommended information is obtained according to the display style of the client.
In the ranking model of the server, in order to improve the learning ability of the model, the final feature dimension may be billion dimensions. The characteristic dimension of the client is not too high, limited by the computing power and the amount of data transmitted by the network. In view of the fact that large-scale users and recommended information features cannot be transmitted to the client, the present invention utilizes fewer features, mainly features affecting the vision of the users, and context (context) features. The recommended information features used in the re-ranking of the present invention include: information semantic vector, information category, CTR of information. The characteristics of the context information include information surrounding the current information item. The method specifically comprises the following steps:
(1) information semantic vector: the floating point unit vector with 128 dimensions of dimensionality is calculated by the server and is transmitted to the client;
(2) information type: the discrete type characteristic is transmitted to the client by the server, and is represented by one-hot;
(3) CTR of information: the CTR estimation result output by the server side sequencing model;
(4) context information: information surrounding the current information item in the current user interface presentation form.
And the estimation module is used for sequentially inputting the characteristics of each second information sequence into the Attention-LSTM-MLP model to obtain a corresponding click through rate estimated value, and selecting the second information sequence corresponding to the highest click through rate estimated value for display.
The recommendation information returned by the server is reordered at the client, and the click through rate pre-estimated value is estimated through an Attention-LSTM-MLP model generated by training. Specifically, the invention adopts a neural network to encode the returned information sequence, wherein RNN is a recurrent neural network, and the neural network is a general term of a series of neural networks capable of processing sequence data. A Long Short Term Memory (LSTM) network is a specific form of RNN, and shows strong adaptability in time sequence data analysis, hidden layer neurons of the LSTM network are connected, the LSTM network can memorize previous information and apply the previous information to current calculation output, and the Long-term dependence problem in time sequence information is solved. Thus, the present invention utilizes an LSTM neural network.
The features of the second information ordering are encoded using LSTM units that learn the input sequence and encode it as a fixed length vector representation. Thus, for input sequences of shorter length, the LSTM neural network is able to learn a corresponding reasonable vector representation. However, it is difficult for the LSTM neural network to learn a reasonable vector representation when the input sequence is very long. Based on this, the present invention introduces an Attention (Attention) mechanism that enables the model to view different parts of the data in a targeted manner. An attention mechanism is added to learn the preferences of the human brain for different content dispersion and aggregation.
The Attention mechanism is implemented by retaining intermediate output results of the LSTM neural network on input sequences, then training a model to selectively learn these inputs and associate the output sequences with them as the model is output.
After the characteristics of the second information ordering are coded through the Attention-LSTM network, the input codes are subjected to click through rate prediction by adopting a Multi-Layer Perceptron (MLP), and the codes are used as input to obtain a score through an MLP network. Finally, we will choose the highest scoring sequence as the final result after reordering.
It should be noted that the training of the Attention-LSTM-MLP model requires processing a large amount of data, considering the resource limitation of the client, and the Attention-LSTM-MLP model is suitable for a large number of clients, so the training of the whole model of the present invention is performed at the server. After training is finished, the model is transmitted to the client, and click through rate prediction is carried out on the client. Therefore, the client reordering system based on edge calculation of the present invention further includes a server, which specifically includes:
the building module is used for building an Attention-LSTM-MLP network;
the invention adds an Attention network before the hidden layer in the traditional LSTM network. The LSTM network encodes each second information ordered feature and converts the second information ordered feature into a high-level feature, and the hidden-layer output vector of the LSTM network is the extracted feature hi. The Attention network is converted into a weight coefficient w of each node through a softmax function in the forward calculation processiMultiplying the value of each node in the attention network by the weight coefficient is the output of the node. The influence degree of each dimension in the hidden layer on the result is judged by adding an attention network, and the larger the weight coefficient is, the larger the influence on the result is, so that the network is focused on the change of one or more dimensions. Through the output of each node, obtainTo the final encoded output y of the Attention-LSTM.
The encoding output y of Attention-LSTM is input to a MLP network, which is an artificial neural network of forward structure that maps a set of input vectors to a set of output vectors. An MLP can be viewed as a directed graph, consisting of multiple layers of nodes, each layer being fully connected to the next. Each node, except the input nodes, is a neuron (or processing unit) with a nonlinear activation function. The present invention adopts the conventional MLP network structure, which is not described herein again.
And the training module is used for extracting the characteristics of the recommendation information and the context information thereof, calculating the click through rate according to the user click number of the recommendation information, and training the Attention-LSTM-MLP network based on the characteristics and the click through rate.
After the Attention-LSTM-MLP network is constructed, the invention trains the network by adopting a large amount of sample data. As with the sequences to be sorted, for each sequence of recommendation information, the characteristics of the recommendation information and its context information are extracted. And the sample data is a displayed and exposed recommended information sequence, and the actual user click number of the user is obtained, so that the click passing rate is calculated.
LTR belongs to one of supervised learning, assuming that the training sample is represented as { (x)i,yi),i∈[1,n]}. Here, unlike pointwise algorithms, xiNot characteristic of an item of information, but xiAnd its contextual characteristics. And click through rate yiCalculated using the following formula:
Figure BDA0002342023030000121
where num is the number of information clicked by the user in the recommended information sequence that has been displayed and exposed at one time. The range of values of this function is [0,1 ].
For the loss function, the invention uses the root mean square error, the formula is as follows:
Figure BDA0002342023030000131
wherein y isiIs the calculated true click through rate, piIs a predicted value calculated according to the Attention-LSTM-MLP network, and n is the number of samples.
The invention inputs the sample data into the Attention-LSTM-MLP network, optimizes the Attention-LSTM-MLP model by calculating the loss function of the whole reordering model, and trains to generate the Attention-LSTM-MLP model.
Therefore, the client reordering method and system based on edge calculation, provided by the invention, reorder the recommendation result on the client by combining the display form of the specific client, fully consider the display characteristics of the client, so that the same server result can be ordered differently at different clients, and improve the user experience and recommendation effect; global target optimization is carried out on the whole list by adopting a listwise algorithm, and because candidate information to be reordered on the end is very small, additional cost introduced on the basis of optimized ordering by adopting the listwise algorithm is small; in view of the fact that large-scale users and information characteristics cannot be transmitted to the client, the method utilizes fewer characteristics, mainly influences the characteristics of user vision and context characteristics, and further optimizes the sequencing performance; the invention learns the preferences of human brain on the dispersion and aggregation of different contents by adding an attention mechanism, reduces the data processing amount in the reordering process and improves the data processing efficiency; the invention fully utilizes the characteristics of the server and the client, trains and generates the Attention-LSTM-MLP model at the server and sends the model to the client, thereby reducing the burden of the client.
It is to be noted that the foregoing is only illustrative of the preferred embodiments of the present invention and the technical principles employed. It will be understood by those skilled in the art that the present invention is not limited to the particular embodiments described herein, but is capable of various obvious changes, rearrangements and substitutions as will now become apparent to those skilled in the art without departing from the scope of the invention. Therefore, although the present invention has been described in greater detail by the above embodiments, the present invention is not limited to the above embodiments, and may include other equivalent embodiments without departing from the spirit of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims (10)

1. A client reordering method based on edge calculation is characterized by comprising the following steps:
s1, acquiring a first information sequence which is recommended by a server and comprises k pieces of information;
s2, randomly arranging the K information to obtain K! Sorting the second information;
s3, extracting the recommended information in each second information sequence and the characteristics of the context information thereof;
and S4, sequentially inputting the characteristics of each second information sequence into the Attention-LSTM-MLP model to obtain corresponding click through rate estimated values, and selecting the second information sequence corresponding to the highest click through rate estimated value for display.
2. The client reordering method of claim 1 wherein the characteristics of the recommendation information comprise: the click through rate estimation result output by the information semantic vector, the information category and the server side sequencing model; the characteristics of the context information include information surrounding the currently recommended information in the current user interface presentation form.
3. The client reordering method of claim 1 wherein the Attention-LSTM-MLP model is generated at the server and sent to the client.
4. The client reordering method of claim 2, wherein generating the Attention-LSTM-MLP model is specifically:
s41, constructing an Attention-LSTM-MLP network;
s42, extracting the characteristics of the recommendation information and the context information thereof, calculating the click through rate according to the user click number of the recommendation information, and training the Attention-LSTM-MLP network based on the characteristics and the click through rate.
5. The client reordering method of claim 4, wherein the Attention-LSTM-MLP network is specifically:
an Attention network is connected behind an implicit layer in the LSTM network; the hidden layer of the LSTM network outputs the extracted features and inputs the features into the Attention network; the Attention network is converted into a weight coefficient of each node through a softmax function, the value of each node in the Attention network is multiplied by the weight coefficient to serve as the output of the node, the final Attention-LSTM coding output is obtained, and the coding output is input into an MLP network to obtain the click through rate pre-estimated value.
6. The client reordering method of claim 5 wherein the loss function of the Attention-LSTM-MLP network is:
Figure FDA0002342023020000021
wherein y isiIs the calculated true click through rate, piIs a predicted value calculated according to the Attention-LSTM-MLP network, n is the number of samples, yiThe specific calculation is as follows:
Figure FDA0002342023020000022
where num is the number of information clicked by the user in the recommended information sequence that has been displayed and exposed at one time.
7. The client reordering method of claim 3 wherein the context information is characterized by 8 positions, and missing is filled with a default value of 0.
8. The utility model provides a client reordering system based on edge calculation, includes the client, its characterized in that specifically includes:
the acquisition module is used for acquiring a first information sequence which is recommended by the server and comprises k pieces of information;
a random ordering module for randomly arranging the K pieces of information to obtain K! Sorting the second information;
the extraction module is used for extracting the recommended information in each second information sequence and the characteristics of the context information of the recommended information;
and the estimation module is used for sequentially inputting the characteristics of each second information sequence into the Attention-LSTM-MLP model to obtain a corresponding click through rate estimated value, and selecting the second information sequence corresponding to the highest click through rate estimated value for display.
9. The client reordering system of claim 8 wherein the reordering system further comprises a server, comprising:
the building module is used for building an Attention-LSTM-MLP network;
and the training module is used for extracting the characteristics of the recommendation information and the context information thereof, calculating the click through rate according to the user click number of the recommendation information, and training the Attention-LSTM-MLP network based on the characteristics and the click through rate.
10. The client reordering system of claim 8 wherein the characteristics of the recommendation information comprise: the click through rate estimation result output by the information semantic vector, the information category and the server side sequencing model; the characteristics of the context information include information surrounding the currently recommended information in the current user interface presentation form.
CN201911390108.5A 2019-12-27 2019-12-27 Client reordering method and system based on edge calculation Active CN111159242B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911390108.5A CN111159242B (en) 2019-12-27 2019-12-27 Client reordering method and system based on edge calculation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911390108.5A CN111159242B (en) 2019-12-27 2019-12-27 Client reordering method and system based on edge calculation

Publications (2)

Publication Number Publication Date
CN111159242A true CN111159242A (en) 2020-05-15
CN111159242B CN111159242B (en) 2023-04-25

Family

ID=70559297

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911390108.5A Active CN111159242B (en) 2019-12-27 2019-12-27 Client reordering method and system based on edge calculation

Country Status (1)

Country Link
CN (1) CN111159242B (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111915414A (en) * 2020-08-31 2020-11-10 支付宝(杭州)信息技术有限公司 Method and device for displaying target object sequence to target user
CN114401135A (en) * 2022-01-14 2022-04-26 国网河北省电力有限公司电力科学研究院 Internal threat detection method based on LSTM-Attention user and entity behavior analysis technology
WO2022083596A1 (en) * 2020-10-20 2022-04-28 北京沃东天骏信息技术有限公司 Sorting method, apparatus and device, and computer storage medium

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105117491A (en) * 2015-09-22 2015-12-02 北京百度网讯科技有限公司 Page pushing method and device
CN108959603A (en) * 2018-07-13 2018-12-07 北京印刷学院 Personalized recommendation system and method based on deep neural network
CN109858806A (en) * 2019-01-30 2019-06-07 网易（杭州）网络有限公司 Method, apparatus, medium and the electronic equipment of data processing
CN110162703A (en) * 2019-05-13 2019-08-23 腾讯科技（深圳）有限公司 Content recommendation method, training method, device, equipment and storage medium
US20190311807A1 (en) * 2018-04-06 2019-10-10 Curai, Inc. Systems and methods for responding to healthcare inquiries

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105117491A (en) * 2015-09-22 2015-12-02 北京百度网讯科技有限公司 Page pushing method and device
US20190311807A1 (en) * 2018-04-06 2019-10-10 Curai, Inc. Systems and methods for responding to healthcare inquiries
CN108959603A (en) * 2018-07-13 2018-12-07 北京印刷学院 Personalized recommendation system and method based on deep neural network
CN109858806A (en) * 2019-01-30 2019-06-07 网易（杭州）网络有限公司 Method, apparatus, medium and the electronic equipment of data processing
CN110162703A (en) * 2019-05-13 2019-08-23 腾讯科技（深圳）有限公司 Content recommendation method, training method, device, equipment and storage medium

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
XIANRONG ZHENG ETAL: "Ranking-Based Cloud Service Recommendation" *
俞春花;刘学军;李斌;章玮;: "基于上下文相似度和社会网络的移动服务推荐方法" *
臧铖: "个性化搜索中隐私保护的关键问题研究" *
黎邦群;: "基于检索行为的非个性化图书推荐" *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111915414A (en) * 2020-08-31 2020-11-10 支付宝(杭州)信息技术有限公司 Method and device for displaying target object sequence to target user
WO2022083596A1 (en) * 2020-10-20 2022-04-28 北京沃东天骏信息技术有限公司 Sorting method, apparatus and device, and computer storage medium
CN114401135A (en) * 2022-01-14 2022-04-26 国网河北省电力有限公司电力科学研究院 Internal threat detection method based on LSTM-Attention user and entity behavior analysis technology

Also Published As

Publication number Publication date
CN111159242B (en) 2023-04-25

Similar Documents

Publication Publication Date Title
CN112150210B (en) Improved neural network recommendation method and system based on GGNN (global warming network)
CN111444428B (en) Information recommendation method and device based on artificial intelligence, electronic equipment and storage medium
CN111382361B (en) Information pushing method, device, storage medium and computer equipment
CN111581510A (en) Shared content processing method and device, computer equipment and storage medium
CN112119388A (en) Training image embedding model and text embedding model
CN108921624B (en) Advertisement fusion method and device, storage medium and terminal equipment
CN111159242B (en) Client reordering method and system based on edge calculation
WO2022016556A1 (en) Neural network distillation method and apparatus
CN111143684B (en) Artificial intelligence-based generalized model training method and device
CN111400603A (en) Information pushing method, device and equipment and computer readable storage medium
CN111563770A (en) Click rate estimation method based on feature differentiation learning
CN112989212B (en) Media content recommendation method, device and equipment and computer storage medium
WO2024041483A1 (en) Recommendation method and related device
WO2024002167A1 (en) Operation prediction method and related apparatus
CN115510313A (en) Information recommendation method and device, storage medium and computer equipment
CN116205700A (en) Recommendation method and device for target product, computer equipment and storage medium
CN112269943B (en) Information recommendation system and method
CN114817692A (en) Method, device and equipment for determining recommended object and computer storage medium
CN116910357A (en) Data processing method and related device
CN116956183A (en) Multimedia resource recommendation method, model training method, device and storage medium
CN115238188A (en) Object recommendation method and system and object recommendation model system
CN112418402A (en) Method for recommending object, neural network and training method thereof, and computing device
CN116089712B (en) Hot conference recommending method and system based on data mining and analysis
CN114996561B (en) Information recommendation method and device based on artificial intelligence
CN116049532A (en) Object recommendation method, device, apparatus, storage medium and computer program product

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
CB02 Change of applicant information
CB02 Change of applicant information

Address after: 22nd floor, block a, Huaxing Times Square, 478 Wensan Road, Xihu District, Hangzhou, Zhejiang 310000

Applicant after: Hangzhou Xiaoying Innovation Technology Co.,Ltd.

Address before: 16 / F, HANGGANG Metallurgical Science and technology building, 294 Tianmushan Road, Xihu District, Hangzhou City, Zhejiang Province, 310012

Applicant before: HANGZHOU QUWEI SCIENCE & TECHNOLOGY Co.,Ltd.

GR01 Patent grant
GR01 Patent grant