CN115048578A

CN115048578A - Recommendation method and device

Info

Publication number: CN115048578A
Application number: CN202210681660.5A
Authority: CN
Inventors: 石晓文; 廖国钢; 王泽�; 吴晓旭; 王永康; 王兴星
Original assignee: Beijing Sankuai Online Technology Co Ltd
Current assignee: Beijing Sankuai Online Technology Co Ltd
Priority date: 2022-06-15
Filing date: 2022-06-15
Publication date: 2022-09-13

Abstract

The specification discloses a recommendation method and a recommendation device, wherein each candidate recommendation sequence determined by combining an appointed number of candidate recommendation objects according to different positions in the sequence and an appointed sequence of appointed operation executed by a user in history are determined, candidate recommendation features determined based on the content and the position of each candidate recommendation object in the candidate recommendation sequence and each appointed feature are obtained through a feature extraction layer of a click rate estimation model, and then the click rate of each candidate recommendation sequence is determined through a prediction layer of the click rate estimation model, so that the recommendation sequence recommended to the user is determined based on each click rate. When the click rate of each candidate recommendation sequence is predicted, the method is determined based on the content of the candidate recommendation objects contained in the candidate recommendation sequence and the positions of the candidate recommendation objects, so that the determined click rate is more accurate, and the recommendation efficiency is higher.

Description

Recommendation method and device

Technical Field

The present specification relates to the field of computer technologies, and in particular, to a recommendation method and apparatus.

Background

Currently, with the development of computer technology, how to recommend a suitable object for a user (e.g., a search result in a search scene, a natural result shown by a service provider to the user, etc.) has become one of the problems that the service provider needs to solve. The recommendation method can recommend suitable objects for the user based on the historical behaviors of the user and the characteristics of each object, and is widely applied to a scene that a service provider recommends objects for the user.

In the prior art, a common recommendation method is determined based on objects that a user has clicked on historically. Specifically, the service provider may obtain a historical behavior sequence of the user to determine objects that the user has clicked historically. Then, candidate recommendation objects may be obtained, and for each candidate recommendation object, a similarity between the candidate recommendation object and an object historically clicked by the user may be determined. And finally, recommending corresponding candidate recommendation objects to the user based on the determined similarity.

However, in the prior art, when recommendation is performed, the influence of other candidate recommendation objects in the page on the click rate of the recommendation object is not considered, so that when recommendation is performed to the user based on the determined candidate recommendation object, the recommendation efficiency is low.

Disclosure of Invention

The present specification provides a recommendation method and apparatus to partially solve the above problems in the prior art.

The technical scheme adopted by the specification is as follows:

the present specification provides a recommendation method including:

determining a specified number of candidate recommendation objects and each specified sequence of specified operations executed by the user in history according to a recommendation request sent by the user;

combining the candidate recommendation objects according to different positions in the sequence to determine candidate recommendation sequences;

for each candidate recommendation sequence, taking the candidate recommendation sequence and each appointed sequence as input, inputting a feature extraction layer of a pre-trained click rate estimation model, and obtaining candidate recommendation features and each appointed feature which are output by the feature extraction layer and determined according to the content of each candidate recommendation object and the position of each candidate recommendation object in the candidate recommendation sequence;

inputting the candidate recommendation characteristics and the designated characteristics as input into a prediction layer of the click rate prediction model to obtain the click rate of the candidate recommendation sequence output by the prediction layer;

and determining a recommendation sequence according to the click rate corresponding to each candidate recommendation sequence, and recommending to the user.

Optionally, the candidate recommendation sequence is used as an input, a feature extraction layer of a pre-trained click rate estimation model is input, and candidate recommendation features output by the feature extraction layer and determined according to the content of each candidate recommendation object included in the candidate recommendation sequence and the position of each candidate recommendation object are obtained, which specifically includes:

for each candidate recommendation sequence, inputting the candidate recommendation sequence serving as input into a feature extraction layer of a pre-trained click rate estimation model;

for each candidate recommendation object in the candidate recommendation sequence, determining the object characteristics of the candidate recommendation object according to the content of the candidate recommendation object and the position of the candidate recommendation object in the candidate recommendation sequence;

determining the similarity corresponding to the candidate recommendation object and other candidate recommendation objects in the candidate recommendation sequence respectively, and enhancing the object characteristics of the candidate recommendation object according to the similarity and the object characteristics of other candidate recommendation objects corresponding to the similarity respectively;

and determining the candidate recommendation characteristics of the candidate recommendation sequence according to the enhancement result of each object characteristic.

Optionally, the candidate recommendation features and the specified features are used as input, and the candidate recommendation features and the specified features are input to a prediction layer of the click rate prediction model to obtain the click rate of the candidate recommendation sequence output by the prediction layer, and the method specifically includes:

inputting the candidate recommended features and the specified features as input into a fusion layer of the click rate estimation model to obtain an enhancement result output by the fusion layer and used for enhancing the candidate recommended features according to the specified features and a fusion result output by the fusion layer and used for fusing the specified features as fusion features;

and inputting the fusion characteristics as input into a prediction layer of the click rate prediction model to obtain the click rate of the candidate recommendation sequence output by the prediction layer.

Optionally, the specified operation includes multiple operation types;

determining each appointed sequence of appointed operations executed by the user in history according to a recommendation request sent by the user, specifically comprising:

determining a user identifier carried in a recommendation request according to the recommendation request sent by a user, and determining sequences browsed by the user in history as browsing sequences according to the user identifier;

determining the historical behaviors of the user, determining the operation of the user on the browsing sequence according to the historical behaviors aiming at each browsing sequence, and determining the operation type of the operation as the label of the browsing sequence;

and for each operation type, determining each browsing sequence marked as the operation type from each browsing sequence as each designated sequence corresponding to the operation type.

Optionally, for each target sequence, the priorities of the objects included in the target sequence are not completely the same;

the method includes the steps of inputting the candidate recommendation sequence as input, inputting a feature extraction layer of a pre-trained click rate estimation model, and obtaining candidate recommendation features output by the feature extraction layer and determined according to the content of each candidate recommendation object contained in the candidate recommendation sequence and the position of each candidate recommendation object, and specifically includes the following steps:

determining the priority of each candidate recommendation object contained in the candidate recommendation sequence;

and inputting the candidate recommendation sequence as input into a feature extraction layer of a pre-trained click rate estimation model to obtain candidate recommendation features which are output by the feature extraction layer and are determined according to the content of each candidate recommendation object, the position of each candidate recommendation object and the corresponding priority of each candidate recommendation object in the candidate recommendation sequence.

Optionally, the click rate pre-estimation model is trained in the following manner:

for each user, determining sequences browsed by the user in history as target sequences, determining labels of the target sequences according to the historical behaviors of the user, and determining sequences of designated operations executed by the user in history from the target sequences as designated sequences;

inputting the target sequences and the designated sequences as input into a feature extraction layer of a click rate estimation model to be trained to obtain target features and designated features, which are output by the feature extraction layer and determined based on the content of each object contained in the target sequence and the position of each object, of each target sequence;

inputting each target characteristic and each specified characteristic as input into a prediction layer of the click rate prediction model to obtain click rates respectively corresponding to each target sequence output by the prediction layer;

and determining loss according to the click rate and the mark of each target sequence, and adjusting model parameters of the click rate estimation model based on the loss.

inputting the candidate recommended features and the designated features as input into a prediction layer of the click rate prediction model;

for each operation type, determining the probability that the candidate recommended features correspond to the operation type according to the similarity of the candidate recommended features corresponding to each specified sequence of the operation type;

and determining the click rate of the candidate recommendation sequence according to the probability that the candidate recommendation features respectively correspond to each operation type and the preset weight of each operation type.

The present specification provides a recommendation device comprising:

the receiving module is used for determining a specified number of candidate recommendation objects and each specified sequence of specified operations executed by the user in history according to a recommendation request sent by the user;

the combination module is used for combining the candidate recommendation objects according to different positions in the sequence to determine candidate recommendation sequences;

the characteristic determining module is used for inputting the candidate recommendation sequence and each appointed sequence as input aiming at each candidate recommendation sequence, inputting a characteristic extraction layer of a pre-trained click rate estimation model, and obtaining candidate recommendation characteristics and each appointed characteristic which are output by the characteristic extraction layer and determined according to the content of each candidate recommendation object and the position of each candidate recommendation object contained in the candidate recommendation sequence;

the prediction module is used for inputting the candidate recommendation characteristics and the designated characteristics into a prediction layer of the click rate prediction model to obtain the click rate of the candidate recommendation sequence output by the prediction layer;

and the determining module is used for determining the recommended sequences according to the click rates respectively corresponding to the candidate recommended sequences and recommending the candidate recommended sequences to the user.

The present specification provides a computer-readable storage medium storing a computer program which, when executed by a processor, implements the recommendation method described above.

The present specification provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the above recommended method when executing the program.

The technical scheme adopted by the specification can achieve the following beneficial effects:

in the recommendation method provided by the present specification, each candidate recommendation sequence determined by combining a specified number of candidate recommendation objects according to different positions in the sequence and a specified sequence of a user historically executing specified operations are determined, and then candidate recommendation features determined based on the content and the position of each candidate recommendation object in the candidate recommendation sequences and each specified feature are obtained through a feature extraction layer of a click rate estimation model, and then the click rate of each candidate recommendation sequence is determined through a prediction layer of the click rate estimation model, so as to determine a recommendation sequence recommended to the user based on each click rate.

According to the method, when the click rate of each candidate recommendation sequence is predicted, the click rate is determined based on the content of the candidate recommendation object contained in the candidate recommendation sequence and the position of each candidate recommendation object, so that the determined click rate is more accurate, and the recommendation efficiency is higher.

Drawings

The accompanying drawings, which are included to provide a further understanding of the specification and are incorporated in and constitute a part of this specification, illustrate embodiments of the specification and together with the description serve to explain the principles of the specification and not to limit the specification in a limiting sense. In the drawings:

FIG. 1 is a schematic flow chart of a recommendation method provided herein;

FIG. 2 is a schematic flow chart of a training method of a click rate estimation model provided in the present specification;

FIG. 3 is a schematic diagram of a training process of a click rate estimation model provided in the present specification;

FIG. 4 is a schematic view of a recommendation device provided herein;

fig. 5 is a schematic diagram of an electronic device corresponding to fig. 1 provided in the present specification.

Detailed Description

In order to make the objects, technical solutions and advantages of the present disclosure more clear, the technical solutions of the present disclosure will be clearly and completely described below with reference to the specific embodiments of the present disclosure and the accompanying drawings. It is to be understood that the embodiments described are only a few embodiments of the present disclosure, and not all embodiments. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments in the present specification without any creative effort belong to the protection scope of the present specification.

Generally, in the recommendation field, the click rate of a user for different objects (e.g., contents) can be determined through a click rate estimation model, a recommendation object to be presented to the user is determined from candidate objects based on the click rate of each object, and the determined recommendation objects are presented in a page in a sequence form to be recommended to the user.

However, when a user clicks on an object in a sequence, the reason for clicking on the object may be factors such as the position of the object in the sequence, the influence of other objects in the sequence on the object, and the like, in addition to the content of the object itself. The object may be a product, a service, a merchant, a business object, and the like, and the type of the object may be set according to a need, which is not limited in this specification.

Taking the page containing A, B, C three products as an example, where the price of the product A and the product C is high and the price of the product B is moderate, the reason that the user clicks on the product B may be because the product B is the lowest price in the page and not because the user likes the product with moderate price.

In summary, the method of determining similarity only by using the object clicked by the user and the candidate object may result in an inaccurate click rate.

The technical solutions provided by the embodiments of the present description are described in detail below with reference to the accompanying drawings.

Fig. 1 is a schematic flow chart of a recommendation method provided in this specification, specifically including the following steps:

s100: according to the received recommendation request and the recommendation request sent by the user, determining a specified number of candidate recommendation objects and each specified sequence of specified operations executed by the user in history.

At present, in the recommendation field, user data can be determined based on a user identifier in a received recommendation request through a recommendation method, and then a specified number of candidate recommendation objects according with user preferences are selected from the candidate recommendation objects based on the user data to form a recommendation sequence to be displayed to a user in a page so as to recommend the user. That is, what the service provider actually recommends to the user is a recommendation sequence containing a specified number of candidate recommendation objects. Compared with the method that the click rate of the user on the candidate recommended object is only determined, the click rate of the user on the recommended sequence containing the recommended object is determined, and the effect is more accurate.

The method is different from the prior art that the click rate of the user on the candidate recommended object is determined by adopting a technical means based on the similarity between the candidate recommended object and each object clicked historically by the user, and the influence of other candidate recommended objects in the recommended sequence on the candidate recommended object and the influence of the position of the candidate recommended object in the recommended sequence are not considered, so that the determined click rate is not accurate enough. The present specification provides a new recommendation method, so that the click rate of a candidate recommendation sequence can be determined based on the content and position of each candidate recommendation object in the candidate recommendation sequence, and the determined click rate is more accurate.

In one or more embodiments provided in this specification, the recommendation method is directed to a scenario in which candidate recommendation objects are presented to a user in units of candidate recommendation sequences, and the recommendation method may be specifically executed by a server.

Based on the method, the specified number of candidate recommendation objects and user data can be determined according to the received recommendation request sent by the user and the number of the candidate recommendation objects contained in the candidate recommendation sequence.

Specifically, the server may receive a recommendation request sent by a user, and determine the specified number carried in the recommendation request and a user identifier corresponding to the user.

Then, the server may determine candidate recommended objects designated as the user amount from among the predetermined candidate recommended objects.

And finally, the server can acquire each sequence of the user which has performed appointed operation historically as each appointed sequence according to the user identification, wherein each appointed sequence is user data representing user preference.

Of course, each candidate recommended object may be determined according to a keyword carried in the recommendation request, or may be determined in advance according to user data for the server, and how to determine each candidate recommended object may be set as needed, which is not limited in this specification.

S102: and combining the candidate recommendation objects according to different positions in the sequence to determine candidate recommendation queues.

In one or more embodiments provided in this specification, even if a candidate recommendation sequence is composed of identical candidate recommendations, if the content of each candidate recommendation in the sequence is different, the click rate of the user for the candidate recommendation is also different, so the server may combine the candidate recommendations into different candidate recommendation queues according to different positions in the sequence.

Specifically, the server may rank and combine the recommendation objects, sort the candidate recommendation objects according to different orders in the sequence, and use the sorting results as the candidate recommendation sequences.

S104: and for each candidate recommendation sequence, inputting the candidate recommendation sequence and each specified sequence as input, inputting a feature extraction layer of a pre-trained click rate estimation model, and obtaining candidate recommendation features and each specified feature, which are output by the feature extraction layer and are determined according to the content of each candidate recommendation object and the position of each candidate recommendation object in the candidate recommendation sequence.

In one or more embodiments provided in the present specification, after determining the candidate recommendation sequence, the server may determine a feature of the candidate recommendation sequence, so as to perform the subsequent step of determining the click rate based on the feature of the candidate recommendation sequence. As described above, unlike the method of determining the click rate by using the similarity between the object clicked by the user and each candidate recommended object, the recommendation method provided in the present specification may determine the click rate not only based on the content of the candidate recommended object, but also based on the position of the candidate recommended object in the candidate recommended sequence. Based on this, the server may determine the candidate recommendation features of the candidate recommendation sequence based on the candidate recommendation objects included in the candidate recommendation sequence and the positions of the candidate recommendation objects in the candidate recommendation sequence.

Specifically, for each candidate recommendation sequence, the server may input the candidate recommendation sequence as an input to a feature extraction layer of a pre-trained click rate estimation model.

The feature extraction layer of the click-through rate estimation model can determine the object features of the candidate recommendation objects according to the content of the candidate recommendation objects and the positions of the candidate recommendation objects in the candidate recommendation sequence aiming at each candidate recommendation object in the candidate recommendation sequence.

Then, the feature extraction layer may fuse the determined candidate recommendation objects to determine candidate recommendation features of the candidate recommendation sequence. The fusion means can include various means such as splicing, adding and the like, and the means of fusing a plurality of characteristics into one characteristic is a mature technical scheme at present, so the scheme is not repeated.

In addition, in the present specification, the click rate of each candidate recommended sequence needs to be determined based on each designated sequence in which the user has performed a designation operation historically. Based on the above, the server may determine each candidate recommended feature, and simultaneously input the specified sequence as an input to the feature extraction layer of the click rate estimation model for each specified sequence, so as to obtain the content of the object output by the feature extraction layer and determined according to the position of the object in the specified sequence.

Furthermore, each candidate recommendation object in the same candidate recommendation sequence is influenced by other candidate recommendation objects in the candidate recommendation sequence, so that in order to accurately determine the object characteristics to determine more accurate candidate recommendation characteristics to ensure the accuracy of the estimated click rate, when determining the object characteristics, the object characteristics of each candidate recommendation object can be enhanced by using the object characteristics of other candidate recommendation objects in the candidate recommendation sequence where the candidate recommendation object is located.

Specifically, the server may first input the candidate recommendation sequence into a feature extraction layer of the click-through rate estimation model, and the feature extraction layer determines, for each candidate recommendation object in the candidate recommendation sequence, an object feature of the candidate recommendation object according to the content of the candidate recommendation object and the position of the candidate recommendation object.

Then, the feature extraction layer may determine similarity corresponding to each of the other candidate recommended objects and the candidate recommended object according to the object feature of the candidate recommended object and the object features of the other candidate recommended objects included in the candidate recommended sequence.

And finally, enhancing the object characteristics of the candidate recommendation object according to the determined similarity and the object characteristics of other candidate recommendation objects corresponding to the similarities.

Finally, the server may use the enhancement result as the object feature of the candidate recommendation object.

S106: and inputting the candidate recommendation characteristics and the designated characteristics as input into a prediction layer of the click rate prediction model to obtain the click rate of the candidate recommendation sequence output by the prediction layer.

In one or more embodiments provided in this specification, the click rate of the user for each candidate recommendation sequence is predicted based on each specified sequence in which the user has performed a specified operation historically and the similarity between each candidate recommendation sequence composed of each candidate recommendation object. Based on the above, after the candidate recommendation features and the designated features are determined, the server can input the determined candidate recommendation features and the determined designated features as inputs to a prediction layer of the estimation model for the click rate to determine the click rate of each candidate recommendation sequence.

Specifically, the server may input, for each candidate recommended sequence, the candidate recommended features of the candidate recommended sequence and the determined specified features into the prediction layer of the click rate prediction model.

The prediction layer can determine similarity corresponding to the candidate recommended features and each designated feature according to the determined candidate recommended features and each designated feature, and determine the click rate of the candidate recommended features according to the type of the designated operation and the weight corresponding to each similarity.

Taking the designated operation as the "click" type as an example, if the similarity is 80%. The determined weight may be 0.8, etc. Taking the example of designating the operation as the "leave" type, if the similarity is 80%, the determined weight may be 0.2, etc. How to determine the weight based on the similarity and the type of the specified operation can be set as required, and this specification does not limit this.

S108: and determining a recommendation sequence according to the click rate corresponding to each candidate recommendation sequence, and recommending to the user.

In one or more embodiments provided in this specification, after determining the click rate corresponding to each candidate recommendation sequence, the server may sort the plurality of candidate recommendation sequences according to the determined click rates, and determine the recommendation sequence to be presented to the user according to the sort.

Based on the recommendation method shown in fig. 1, each candidate recommendation sequence determined by combining the specified number of candidate recommendation objects according to different positions in the sequence and the specified sequence of the user which has performed the specified operation historically are determined, and then the candidate recommendation features determined based on the content and the positions of the candidate recommendation objects in the candidate recommendation sequence and each specified feature are obtained through the feature extraction layer of the click rate estimation model, and further the click rate of each candidate recommendation sequence is determined through the prediction layer of the click rate estimation model, so that the recommendation sequence recommended to the user is determined based on each click rate. When the click rate of each candidate recommendation sequence is predicted, the method is determined based on the content of the candidate recommendation objects contained in the candidate recommendation sequence and the positions of the candidate recommendation objects, so that the determined click rate is more accurate, and the recommendation efficiency is higher.

In addition, the click rate estimation model used in the recommendation method can be generally trained based on the following ways:

first, the server may determine, for each user, sequences historically browsed by the user as target sequences, determine labels of the target sequences according to historical behaviors of the user, and determine sequences historically executed by the user as designated sequences from the target sequences.

Secondly, the server can take the target sequence and each appointed sequence as input aiming at each target sequence, and inputs the target sequence and each appointed sequence into a feature extraction layer of the click rate estimation model to be trained to obtain target features which are output by the feature extraction layer and determined based on the content of each object and the position of each object in the target sequence, and each appointed feature;

then, the server takes the target characteristic and each designated characteristic as input, inputs the prediction layer of the click rate prediction model, and obtains the click rate of the target sequence output by the prediction layer;

and finally, the server can determine loss according to the click rate and the label of each target sequence, and adjust the model parameters of the click rate estimation model based on the loss.

It should be noted that the server for training the model and the server for executing the recommendation method provided in this specification may be the same server or different servers.

Further, the accuracy of the result determined only by using positive feedback or negative feedback is obviously lower than that determined by using both positive feedback and negative feedback, so that the server can also set various types of specified sequences in order to more accurately determine the click rate of each candidate recommended sequence.

Specifically, the server may determine, according to a recommendation request sent by a user, a user identifier carried in the recommendation request, and determine, according to the user identifier, each sequence that the user has historically browsed, as each browsing sequence.

Then, after each browsing sequence is determined, the server can also determine the historical behavior of the user, and determine the operation of the user on the browsing sequence according to the determined historical behavior for each browsing sequence, and determine the operation type to which the operation belongs, as the label of the browsing sequence.

Finally, after determining the annotations of the browsing sequences, the server may determine, for each operation type, each browsing sequence labeled as the operation type from the browsing sequences as each specified sequence corresponding to the operation type.

Further, when determining the click rate corresponding to the candidate recommendation sequence, the server may determine the click rate of the user for the candidate recommendation sequence based on multiple types of operations performed by the user at the same time.

Specifically, first, the server may input the candidate recommended features and each specified feature as input, and input the prediction layer of the click-through rate prediction model.

Then, for each operation type, the prediction layer may determine, according to the similarity of each specified sequence of the operation type corresponding to the candidate recommended feature, a probability that the candidate recommended feature corresponds to the operation type.

And finally, according to the probability that the candidate recommendation features respectively correspond to each operation type and the preset weight of each operation type, the server can determine the click rate of the candidate recommendation sequence.

In addition, when recommending to a user, the objects included in the candidate recommendation sequence often correspond to different priorities, for example, the objects include an advertisement and a natural result, the priority corresponding to the natural result is high, and the priority corresponding to the advertisement is low, or the priority corresponding to the natural result is low, and the priority corresponding to the advertisement is high. The service provider is more inclined to click the advertisement on the user, so that in determining the candidate recommendation feature of the candidate recommendation sequence, the server may also determine the priority of each candidate recommendation object contained in the candidate recommendation sequence.

The server may use the candidate recommendation sequence as an input, and input the feature extraction layer of the pre-trained click rate estimation model to obtain candidate recommendation features output by the feature extraction layer and determined according to the content of each candidate recommendation object, the position of each candidate recommendation object, and the priority corresponding to each candidate recommendation object included in the candidate recommendation sequence.

Based on the same idea, the specification further provides a training method of the click rate estimation model, as shown in fig. 2.

Fig. 2 is a schematic flow chart of a training method of a click rate estimation model provided in this specification, in which:

s200: and determining sequences historically browsed by the user as target sequences for each user, determining labels of the target sequences according to the historical behaviors of the user, and determining sequences historically executed with designated operations by the user from the target sequences as designated sequences.

Generally, the click-through rate prediction model is obtained by a server for training the model and is trained in advance based on training samples. The present specification provides a training method of a click-through rate estimation model, and as such, the process of training the click-through rate estimation model may be performed by a server for training the model.

The training model can be divided into a sample generation stage and a training model stage, and in the sample generation stage, samples for training the model can be determined according to model requirements and training requirements. In this specification, the server may first determine training samples for training a click through rate estimation model, and since the click through rate estimation model in this specification is used for click through rate estimation of candidate recommended sequences, the server may first determine sequences historically browsed by the user to determine the training samples.

Based on this, the server may determine, for each user, sequences that the user has historically browsed. The sequences may be pre-stored, and the server may obtain from a pre-stored history browsing log corresponding to each user according to a user identifier of the user, or may directly obtain from a terminal corresponding to the user, where the history browsing log includes each sequence browsed by the user in history, an object included in each sequence, an operation performed by the user on each sequence, and the like, and specifically, the content included in the history browsing log and how to obtain the history browsing log may be set as needed, which is not limited in this specification.

After determining the historical browsing log of the user, the server may determine, based on the historical browsing log, sequences that the user has historically browsed as target sequences of the user.

Similarly, the server may also determine the labels corresponding to the target sequences respectively based on the operations performed by the user in the target sequences, that is, the historical behaviors of the user, which are included in the historical browsing log. The label may include two types, click and un-click.

Of course, for each target sequence of the user, the operation performed by the user in each target sequence reflects the preference of the user corresponding to each target sequence to a large extent, and thus, each target sequence on which the user performed the specified operation may be determined as each specified sequence based on the type of the operation performed by the user in each target sequence. The specified operation may be a "click" operation, a "collection" operation, a "leave" operation, and the like, and a specific type of the specified operation may be set as needed, which is not limited in this specification.

S202: and aiming at each target sequence, inputting the target sequence and each appointed sequence as input into a feature extraction layer of the click rate estimation model to be trained, and obtaining target features and each appointed feature, which are output by the feature extraction layer and determined based on the content of each object contained in the target sequence and the position of each object.

In one or more embodiments provided herein, after determining the target sequences, the server may determine characteristics of each target sequence to facilitate subsequent steps of determining click rates based on the characteristics of the target sequences.

Different from the method for determining the click rate of the candidate recommended object by using the object clicked by the user at present, the click rate estimation model provided by the specification not only determines the click rate of the candidate recommended sequence based on the content of each candidate recommended object, but also determines the click rate of the candidate recommended sequence based on the position of each candidate recommended content in the candidate recommended sequence.

Based on this, the server may determine the target feature based on the objects contained in the target sequence and the position of each object in the target sequence.

Specifically, for each target sequence, the server may input the target sequence as an input to the feature extraction layer of the click rate estimation model.

The feature extraction layer of the click rate pre-estimation model can determine the object features corresponding to each object contained in the target sequence according to the content of the object and the position of the object in the target sequence.

Then, the feature extraction layer may fuse the object features corresponding to the determined objects to determine the target features corresponding to the target sequence. The fusion means can include various means such as splicing, adding and the like, and the means of fusing a plurality of characteristics into one characteristic is a mature technical scheme at present, so the scheme is not repeated.

Meanwhile, the server can also input each designated sequence into the feature extraction layer to obtain the designated features which are output by the feature extraction layer and determined according to each designated sequence and the position of each object in the designated sequence, wherein each object is contained in the designated sequence.

Of course, when determining the target feature and the specified feature, the server may also determine a user feature corresponding to the user, and update the target feature and the specified feature based on the user feature. The updating method can adopt the technical means adopted in the fusion.

Further, for each object, in order to more accurately reflect the influence of other objects in the target sequence on the object, when determining the target feature, the server may also enhance the object feature of the object with the object feature of the other objects to determine the target feature.

Specifically, the server may first input the target sequence into the feature extraction layer, and the feature extraction layer determines, for each object included in the target sequence, an object feature of the object.

Then, the feature extraction layer may determine similarity between the object and each of the other objects according to the object feature and object features of the other objects included in the target sequence, and enhance the object feature of the object based on the determined similarities and object features of the other objects corresponding to the similarities. Wherein, the technical means adopted during the enhancement can be the technical means adopted during the updating.

Finally, the server can fuse the enhancement results of the object features of the objects to determine the target features of the target sequence.

Of course, the enhanced manner described above may also be employed in determining each of the specified characteristics to determine a more accurate click rate.

S204: and inputting the target characteristics and the specified characteristics as input into a prediction layer of the click rate prediction model to obtain the click rate of the target sequence output by the prediction layer.

In one or more embodiments provided in this specification, because the relationship between the sequence including different features and different positions of the different features and the click rate of the user on the sequence is predicted based on the target sequence historically browsed by the user and the relationship between each designated sequence in which the user performs the designated operation, after the target feature and each designated feature are determined, the server may predict the click rate corresponding to the target sequence based on the target feature and each designated feature.

Specifically, the server may input the target feature and each of the designated features as input to a prediction layer of the click-through rate prediction model.

The prediction layer may determine, according to the obtained target feature and each specified feature, a similarity that each specified feature corresponds to the target feature, and determine, according to the type of the specified operation and the weight corresponding to each similarity, a click rate corresponding to the target sequence.

In addition, when the click rate corresponding to the target sequence is determined, the prediction layer can also fuse the target feature and each designated feature, and input the fusion result into network structures such as a full-link layer and a multilayer perceptron to determine the click rate corresponding to the fusion feature. Of course, the specific network structure of the prediction layer may be set as needed, which is not limited in this specification.

Further, in order to further enhance the characteristics of the target sequence to determine a more accurate click rate, the server may further input the target characteristics of the target sequence and each designated characteristic into a fusion layer of the click rate prediction model for fusion before inputting the target characteristics of the target sequence into the prediction layer.

The fusion layer can respectively determine the similarity between each target feature and the designated feature, determine the weight based on the similarity, and enhance the target feature according to the determined weight and each designated feature, that is, enhance the target feature. The fusion layer may output the enhancement result as a fusion feature.

The server can input the determined fusion characteristics into a prediction layer of the click rate prediction model, and the click rate corresponding to the target sequence is output by the prediction layer.

S206: and determining loss according to the click rate and the mark of each target sequence, and adjusting model parameters of the click rate estimation model based on the loss.

In one or more embodiments provided in this specification, after determining the click through rate of the target sequence, the server may train the click through rate estimation model based on the label and the click through rate of each target sequence.

Specifically, the server may determine a loss based on the click rate and the label of each target sequence, and adjust model parameters of the click rate estimation model based on the loss to complete training of the click rate estimation model.

In addition, based on the same idea of the training method of the click rate estimation model shown in fig. 2, the present specification further provides a schematic diagram of a training process of the click rate estimation model, as shown in fig. 3.

Fig. 3 is a schematic diagram of a training process of a click rate estimation model provided in this specification. The server may first obtain a historical browsing log of the user, determine each target sequence of the user according to the historical browsing log, and determine each designated sequence in which the user has performed a designated operation. And then, aiming at each target sequence, taking the target sequence and each appointed sequence as input, and inputting a feature extraction layer of the click rate estimation model to obtain the target feature and each appointed feature of the target sequence. And then inputting the target characteristic and each designated characteristic as input, inputting a fusion layer of the click rate estimation model, and determining the fusion characteristic of the target sequence. The server can input the fusion characteristics into a prediction layer of the click rate prediction model to obtain the click rate of the target sequence output by the prediction layer. And determining loss based on the click rate and the label of the target sequence, and training the click rate estimation model according to the loss.

The history browsing log is only an example description of a manner of obtaining a target sequence of a user, and specifically how to obtain the target sequence and a specified sequence of the user may be set as required, which is not limited in this specification.

Based on the recommendation method shown in fig. 2, based on each target sequence of the user and each designated sequence of the user executing designated operation, determining target features determined according to each object contained in the target sequence and the position of each object in the target sequence and each designated feature through a feature extraction layer of a click rate prediction model, determining a click rate determined according to the target features and each designated feature through a prediction layer of the click rate prediction model, and finally determining a loss according to the click rate and a label of each target sequence to train the model. According to the scheme, when the click rate of the page is predicted, the click rate is determined not only based on the contained objects, but also based on the positions of the objects, and the trained model is more accurate in prediction.

Furthermore, in order to train a more accurate click rate estimation model, when determining the labels and each designated sequence of the target sequence, multiple types of labels and designated sequences can be set.

Specifically, the server may determine, for each target sequence, an operation executed by the user in the target sequence according to the historical behavior of the user, and determine an operation type to which the operation belongs, as a label of the target sequence.

Then, after determining the label of each target sequence, the server may determine, for each operation type, each target sequence labeled as the operation type from the target sequences as each designated sequence corresponding to the operation type.

Furthermore, when recommending to the user, the objects included in the candidate recommendation sequence often correspond to different priorities, for example, the objects include an advertisement and a natural result, the priority corresponding to the natural result is high, and the priority corresponding to the advertisement is low, or the priority corresponding to the natural result is low, and the priority corresponding to the advertisement is high. The service provider is more inclined to have the user click on the advertisement and therefore, in determining the targeting characteristics of the targeted sequence, the server may also determine the priority of each object that the targeted sequence contains.

The server may use the target sequence as an input, input the feature extraction layer of the click rate pre-estimation model to be trained, obtain the content corresponding to each object included in the target sequence, the position of each object in the target sequence, and the priority corresponding to each object, which are output by the feature extraction layer, and determine the target feature of the target sequence.

It should be noted that, in order to ensure uniformity of model input dimensions and facilitate better practice of the model, when determining a specified sequence of a user, a specified number of sequences need to be determined as the specified sequence of the user, and if the number of sequences included in the specified sequence of the user is less than the specified number, a sequence in which the user performs the specified operation and a null sequence can be combined into the specified sequence.

It should be noted that all actions of acquiring signals, information or data in the present application are performed under the premise of complying with the corresponding data protection regulation policy of the country of the location and obtaining the authorization given by the owner of the corresponding device.

Based on the same idea, the recommendation method provided above for one or more embodiments of the present specification further provides a corresponding recommendation device, as shown in fig. 4.

Fig. 4 is a recommendation device provided in the present specification, including:

the receiving module 300 is configured to determine, according to a recommendation request sent by a user, a specified number of candidate recommendation objects and each specified sequence of specified operations that have been executed by the user historically.

And the combining module 302 is configured to combine the candidate recommendation objects according to different positions in the sequence to determine each candidate recommendation sequence.

And the feature determination module 304 is configured to, for each candidate recommendation sequence, input the candidate recommendation sequence and each designated sequence as inputs into a feature extraction layer of a pre-trained click rate estimation model, and obtain candidate recommendation features and each designated feature, which are output by the feature extraction layer and are determined according to the content of each candidate recommendation object included in the candidate recommendation sequence and the position of each candidate recommendation object.

And the prediction module 306 is configured to input the candidate recommendation features and the specified features as inputs to a prediction layer of the click rate prediction model, so as to obtain a click rate of the candidate recommendation sequence output by the prediction layer.

And the recommending module 308 is configured to determine a recommending sequence according to the click rate corresponding to each candidate recommending sequence, and recommend the candidate recommending sequence to the user.

Optionally, the method further includes:

a training module 310, configured to determine, for each user, sequences historically browsed by the user as target sequences, determine labels of the target sequences according to historical behaviors of the user, determine, from the target sequences, sequences historically executed by the user as designated sequences, and, for each target sequence, input the target sequence and the designated sequences as inputs into a feature extraction layer of a click rate prediction model to be trained, obtain target features output by the feature extraction layer and determined based on contents of objects included in the target sequence and positions of the objects, and each designated feature, input the target features and the designated features as inputs into a prediction layer of the click rate prediction model, and obtain a click rate of the target sequence output by the prediction layer, and determining loss according to the click rate and the mark of each target sequence, and adjusting model parameters of the click rate estimation model based on the loss.

Optionally, the feature determining module 304 is configured to, for each candidate recommendation sequence, use the candidate recommendation sequence as an input, input the input to a feature extraction layer of a pre-trained click rate prediction model, for each candidate recommendation object in the candidate recommendation sequence, determine object features of the candidate recommendation object according to the content of the candidate recommendation object and the position of the candidate recommendation object in the candidate recommendation sequence, determine similarity degrees corresponding to the candidate recommendation object and other candidate recommendation objects in the candidate recommendation sequence, enhance the object features of the candidate recommendation object according to the similarity degrees and the object features of the other candidate recommendation objects corresponding to the similarity degrees, and determine the candidate recommendation features of the candidate recommendation sequence according to the enhancement result of the object features.

Optionally, the predicting module 306 is configured to input the candidate recommendation feature and the specified features as inputs, input a fusion layer of the click rate prediction model, obtain an enhancement result that is output by the fusion layer and enhances the candidate recommendation feature according to the specified features, and a fusion result that is output by the fusion layer and fused according to the specified features, and input the fusion feature as an input to the predicting layer of the click rate prediction model, so as to obtain the click rate of the candidate recommendation sequence output by the predicting layer.

Optionally, the designated operation includes multiple operation types, and the receiving module 300 is configured to determine, according to a recommendation request sent by a user, a user identifier carried in the recommendation request, determine, according to the user identifier, each sequence historically browsed by the user as each browsing sequence, determine a historical behavior of the user, determine, for each browsing sequence, an operation of the user on the browsing sequence according to the historical behavior, determine an operation type to which the operation belongs, serve as a label of the browsing sequence, and determine, for each operation type, each browsing sequence labeled as the operation type from each browsing sequence as each designated sequence corresponding to the operation type.

Optionally, for each target sequence, the priorities of the objects included in the target sequence are not completely the same, the feature determination module 304 is configured to determine the priorities of the candidate recommendation objects included in the candidate recommendation sequence, and input the candidate recommendation sequence as an input to a feature extraction layer of a pre-trained click rate estimation model to obtain candidate recommendation features output by the feature extraction layer and determined according to the content of the candidate recommendation objects included in the candidate recommendation sequence, the positions of the candidate recommendation objects, and the priorities of the candidate recommendation objects, which correspond to the priorities respectively.

Optionally, the recommending module 308 is configured to take the candidate recommended features and the specified features as input, input the candidate recommended features into a prediction layer of the click rate prediction model, determine, for each operation type, a probability that the candidate recommended features correspond to the operation type according to similarity between each specified sequence of the operation type and each of the candidate recommended features, and determine, according to the probability that the candidate recommended features correspond to each of the operation types and preset weight of each of the operation types, the click rate of the candidate recommended sequence.

The present specification also provides a computer-readable storage medium storing a computer program, which can be used to execute the recommendation method provided in fig. 1.

This specification also provides a schematic block diagram of the electronic device shown in fig. 5. As shown in fig. 5, at the hardware level, the electronic device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, and may also include hardware required by other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs the computer program to implement the recommended method described in fig. 1. Of course, besides the software implementation, the present specification does not exclude other implementations, such as logic devices or a combination of software and hardware, and the like, that is, the execution subject of the following processing flow is not limited to each logic unit, and may be hardware or logic devices.

In the 90 s of the 20 th century, improvements in a technology could clearly distinguish between improvements in hardware (e.g., improvements in circuit structures such as diodes, transistors, switches, etc.) and improvements in software (improvements in process flow). However, as technology advances, many of today's process flow improvements have been seen as direct improvements in hardware circuit architecture. Designers almost always obtain the corresponding hardware circuit structure by programming an improved method flow into the hardware circuit. Thus, it cannot be said that an improvement in the process flow cannot be realized by hardware physical modules. For example, a Programmable Logic Device (PLD), such as a Field Programmable Gate Array (FPGA), is an integrated circuit whose Logic functions are determined by programming the Device by a user. A digital system is "integrated" on a PLD by the designer's own programming without requiring the chip manufacturer to design and fabricate application-specific integrated circuit chips. Furthermore, nowadays, instead of manually making an Integrated Circuit chip, such Programming is often implemented by "logic compiler" software, which is similar to a software compiler used in program development and writing, but the original code before compiling is also written by a specific Programming Language, which is called Hardware Description Language (HDL), and HDL is not only one but many, such as abel (advanced Boolean Expression Language), ahdl (alternate Hardware Description Language), traffic, pl (core universal Programming Language), HDCal (jhdware Description Language), lang, Lola, HDL, laspam, hardward Description Language (vhr Description Language), vhal (Hardware Description Language), and vhigh-Language, which are currently used in most common. It will also be apparent to those skilled in the art that hardware circuitry for implementing the logical method flows can be readily obtained by a mere need to program the method flows with some of the hardware description languages described above and into an integrated circuit.

The controller may be implemented in any suitable manner, for example, the controller may take the form of, for example, a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro) processor, logic gates, switches, an Application Specific Integrated Circuit (ASIC), a programmable logic controller, and an embedded microcontroller, examples of which include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320, the memory controller may also be implemented as part of the control logic for the memory. Those skilled in the art will also appreciate that, in addition to implementing the controller as pure computer readable program code, the same functionality can be implemented by logically programming method steps such that the controller is in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers and the like. Such a controller may thus be considered a hardware component, and the means included therein for performing the various functions may also be considered as a structure within the hardware component. Or even means for performing the functions may be conceived to be both a software module implementing the method and a structure within a hardware component.

The systems, devices, modules or units illustrated in the above embodiments may be implemented by a computer chip or an entity, or by a product with certain functions. One typical implementation device is a computer. In particular, the computer may be, for example, a personal computer, a laptop computer, a cellular telephone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

For convenience of description, the above devices are described as being divided into various units by function, respectively. Of course, the functions of the various elements may be implemented in the same one or more software and/or hardware implementations of the present description.

As will be appreciated by one skilled in the art, embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

In a typical configuration, a computing device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.

The memory may include forms of volatile memory in a computer readable medium, Random Access Memory (RAM) and/or non-volatile memory, such as Read Only Memory (ROM) or flash memory (flash RAM). Memory is an example of a computer-readable medium.

Computer-readable media, including both non-transitory and non-transitory, removable and non-removable media, may implement information storage by any method or technology. The information may be computer readable instructions, data structures, modules of a program, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), Read Only Memory (ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), Digital Versatile Discs (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device. As defined herein, a computer readable medium does not include a transitory computer readable medium such as a modulated data signal and a carrier wave.

It should also be noted that the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.

As will be appreciated by one skilled in the art, embodiments of the present description may be provided as a method, system, or computer program product. Accordingly, the description may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the description may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

This description may be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media including memory storage devices.

The embodiments in the present specification are described in a progressive manner, and the same and similar parts among the embodiments are referred to each other, and each embodiment focuses on the differences from the other embodiments. In particular, for the system embodiment, since it is substantially similar to the method embodiment, the description is simple, and for the relevant points, reference may be made to the partial description of the method embodiment.

The above description is only an example of the present specification, and is not intended to limit the present specification. Various modifications and alterations to this description will become apparent to those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present specification should be included in the scope of the claims of the present specification.

Claims

1. A recommendation method, comprising:

2. The method according to claim 1, wherein the candidate recommendation sequence is used as an input, a feature extraction layer of a pre-trained click rate estimation model is input, and candidate recommendation features determined according to the content of each candidate recommendation object included in the candidate recommendation sequence and the position of each candidate recommendation object and output by the feature extraction layer are obtained, and the method specifically comprises:

3. The method of claim 1, wherein the step of inputting the candidate recommendation features and the specified features as inputs into a prediction layer of the click-through rate prediction model to obtain the click-through rate of the candidate recommendation sequence output by the prediction layer comprises:

4. The method of claim 1, wherein the specified operation comprises a plurality of operation types;

and aiming at each operation type, determining each browsing sequence marked as the operation type from each browsing sequence as each specified sequence corresponding to the operation type.

5. The method of claim 1, wherein for each target sequence, the priority of objects contained in the target sequence is not identical;

6. The method of claim 1, wherein the click-through rate prediction model is trained by:

for each target sequence, inputting the target sequence and each designated sequence as input into a feature extraction layer of a click rate estimation model to be trained, and obtaining target features and each designated feature, which are output by the feature extraction layer and determined based on the content of each object contained in the target sequence and the position of each object;

inputting the target characteristics and the designated characteristics as input into a prediction layer of the click rate prediction model to obtain the click rate of the target sequence output by the prediction layer;

7. The method of claim 4, wherein the step of inputting the candidate recommendation features and the specified features as inputs into a prediction layer of the click-through rate prediction model to obtain the click-through rate of the candidate recommendation sequence output by the prediction layer comprises:

inputting the candidate recommendation characteristics and the designated characteristics as input into a prediction layer of the click rate prediction model;

8. A recommendation device, comprising:

and the recommending module is used for determining the recommending sequences according to the click rates respectively corresponding to the candidate recommending sequences and recommending the candidate recommending sequences to the user.

9. A computer-readable storage medium, characterized in that the storage medium stores a computer program which, when executed by a processor, implements the method of any of the preceding claims 1 to 7.

10. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method of any of claims 1 to 7 when executing the program.