CN110209922B

CN110209922B - Object recommendation method and device, storage medium and computer equipment

Info

Publication number: CN110209922B
Application number: CN201810601391.0A
Authority: CN
Inventors: 吴书; 黄婷婷; 杨敏
Original assignee: Tencent Technology Shenzhen Co Ltd; Institute of Automation of Chinese Academy of Science
Current assignee: Tencent Technology Shenzhen Co Ltd; Institute of Automation of Chinese Academy of Science
Priority date: 2018-06-12
Filing date: 2018-06-12
Publication date: 2023-11-10
Anticipated expiration: 2038-06-12
Also published as: CN110209922A

Abstract

The embodiment of the invention provides an object recommending method, an object recommending device, a storage medium and computer equipment. Therefore, the embodiment enriches the expression content of the feature vector, considers the operation sequence of each target object in the model calculation process, and improves the accuracy of the prediction result; in addition, as a plurality of objects output by the application platform share attribute information, the parameter number of the training prediction model is greatly reduced, the model training difficulty is reduced, and the method and the device are applicable to a large number of application scenes.

Description

Object recommendation method and device, storage medium and computer equipment

Technical Field

The present invention relates to the field of data processing technologies, and in particular, to an object recommendation method, an object recommendation device, a storage medium, and a computer device.

Background

In recent years, the popularization of the Internet brings a large amount of information to users, and the information utilization efficiency is reduced because the users cannot quickly find useful part of information due to the overlarge information amount while meeting the information demands of the users. In the fields of news, business, entertainment and the like, the recommendation system is used for analyzing behavior data of the user on the corresponding application platform, predicting preference of the user, screening information possibly interested in the user in the future according to the preference, pushing the information to the user, and helping the user to quickly and accurately select the required information.

In the existing recommendation system, the behavior data of a plurality of sample users in an application platform are generally trained based on a cyclic neural network, so that a corresponding prediction model is obtained and used for predicting the probability that the current user is interested in an object on the application platform. In the model training process, corresponding word vectors are generally allocated as training data for each object on an application platform, so that the training of a prediction model is realized.

However, the number of objects actually provided by the application platform tends to be very large, which causes a rapid increase in the number of parameters of the prediction model, greatly increases the training difficulty of the prediction model, and since one word vector tends to not accurately represent the object content, this affects the accuracy of the prediction result actually obtained by the prediction model, so that the object pushed to the user may not be the object of interest to the user.

It can be seen that how to improve the accuracy of the recommended objects pushed to the user is one of the important research directions for technicians.

Disclosure of Invention

In view of the above, the embodiments of the present invention provide an object recommendation method, apparatus, storage medium, and computer device, which reduce model training difficulty and improve accuracy of prediction results.

In order to achieve the above object, the embodiment of the present invention provides the following technical solutions:

an object recommendation method, the method comprising:

acquiring feature vectors respectively corresponding to a plurality of target objects, wherein each feature vector represents various attribute information of the corresponding target object, and the attribute information is shared by the plurality of target objects;

sequentially inputting corresponding feature vectors into a prediction model according to the operation sequence of the plurality of target objects to obtain predicted feature vectors;

calculating the similarity between the predicted feature vector and the feature vector corresponding to each candidate object;

and screening at least one candidate object as a recommended object based on the similarity corresponding to each candidate object.

An object recommendation device, the device comprising:

the characteristic vector acquisition module is used for acquiring characteristic vectors corresponding to a plurality of target objects respectively, each characteristic vector characterizes various attribute information of the corresponding click object, and the attribute information is shared by the plurality of target objects;

the prediction feature vector calculation module is used for sequentially inputting corresponding feature vectors into a prediction model according to the operation sequence of the plurality of target objects to obtain prediction feature vectors;

The similarity calculation module is used for calculating the similarity between the predicted feature vector and the feature vector corresponding to each candidate object;

and the recommended object screening module is used for screening at least one candidate object as a recommended object based on the similarity corresponding to each candidate object.

A storage medium having stored thereon a computer program for execution by a processor for performing the steps of the object recommendation method as described above.

A computer device, the computer device comprising:

a communication interface;

a memory for storing a program for implementing the object recommendation method as described above;

a processor for loading and executing a program stored in the memory, the program for:

acquiring feature vectors respectively corresponding to a plurality of target objects, wherein each feature vector represents various attribute information of a corresponding click object, and the attribute information is shared by the plurality of target objects;

Based on the above technical solution, in the embodiment of the present invention, by obtaining multiple attribute information of a target object, a feature vector of the target object is constructed, and a prediction model is input according to an operation sequence of each target object to obtain a predicted feature vector of the user, so that the expression content of the feature vector of the target object is enriched, and the operation sequence of each target object is considered in the model calculation process, thereby improving the accuracy of a prediction result; in addition, as a plurality of objects output by the application platform share attribute information, the parameter number of the training prediction model is greatly reduced, the model training difficulty is reduced, and the method and the device are applicable to a large number of application scenes.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are required to be used in the embodiments or the description of the prior art will be briefly described below, and it is obvious that the drawings in the following description are only embodiments of the present invention, and that other drawings can be obtained according to the provided drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic diagram of an object recommendation system according to an embodiment of the present invention;

fig. 2 is a flow chart of an object recommendation method according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of a prediction model used for implementing an object recommendation method according to the present invention;

fig. 4 is an application schematic diagram of an object recommendation method according to an embodiment of the present invention;

FIG. 5 is a flowchart illustrating another object recommendation method according to an embodiment of the present invention;

FIG. 6 is a schematic application diagram of another object recommendation method according to an embodiment of the present invention;

fig. 7 is an application schematic diagram of another object recommendation method according to an embodiment of the present invention;

FIG. 8 is a flowchart illustrating another object recommendation method according to an embodiment of the present invention;

fig. 9 is a schematic structural diagram of an object recommendation apparatus according to an embodiment of the present invention;

FIG. 10 is a schematic structural diagram of another object recommendation apparatus according to an embodiment of the present invention;

FIG. 11 is a schematic structural diagram of another object recommendation apparatus according to an embodiment of the present invention;

fig. 12 is a schematic hardware structure of a computer device according to an embodiment of the present invention.

Detailed Description

The inventor of the present invention found that when a user selects an object output by an application platform, the user often refers to factors of multiple aspects of the object, and takes a shopping platform as an example, when the user selects a commodity, the user usually considers parameters of multiple aspects such as a class and a style of the commodity, and then selects the commodity of the heart instrument, and does not only see parameters of one aspect of the commodity. Based on this, the inventor proposes to express the rich features of the object more carefully, that is, to express an object by adopting a multi-semantic expression mode, so as to improve the prediction accuracy.

In order to reduce the training difficulty of the prediction model, a semantic expression vector is provided for representing one type of parameters, a plurality of objects share the semantic expression vector of the one type of parameters, the number of constructed semantic expression vectors is greatly reduced, the number of parameters of the prediction model is greatly reduced, and the concept proposed by the inventor is known through actual calculation, so that the number of parameters of the prediction model is reduced to the opening order of the original number. Therefore, the inventor proposes the idea, not only improves the accuracy of the prediction result, but also improves the model training efficiency, and achieves the aim of improving the prediction efficiency.

The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

Referring to fig. 1, for implementing a system structure diagram of an object recommendation method provided by the present invention, the system may include an application server 10 and an application client 11, where:

The application server 10 may be a service device that provides a service to a user, and may be a server that matches a client, such as the application client being shopping software, the application server being a server that provides shopping services, the application client being news browsing software, the application server being a server that provides news services, etc. The object recommendation method provided by the invention can be executed by the application server in the embodiment, so that the needed recommended objects can be pushed to the online user, the user can be assisted to quickly and accurately select the needed objects, or the user can conveniently know other information related to the current browsed objects, and the business service efficiency is improved.

The application server 10 may be an independent application service device, or may be a service cluster formed by a plurality of servers, and the structure of the server is not limited in this embodiment.

The application client 11 may be an application program installed on a device such as a mobile phone, a notebook computer, an iPad, an industrial personal computer, etc., and may specifically be an independent application program, such as a software installed by downloading from an application store, or may be a web application program, that is, without downloading, the application program such as a browser directly starts the client, and establishes a communication connection with a corresponding application server.

In practical application of the embodiment, a user enters an application platform of an application server through an application client, and can browse various objects output by the application platform, in the browsing process, the server can push at least one recommended object to the application client used by the user according to the method provided by the embodiment, the recommended object is often an object of interest to the user, such as an object similar to the object operated by the user, and the like, so that the user can conveniently and quickly know the information of interest or select the object of interest, and the like, the user is not required to output the corresponding one-to-one view to the application platform, and the selection efficiency is improved.

It should be noted that, in the system provided in this embodiment, the application server and the application client are not limited to the application server and the application client listed above, but may also include other computer devices such as a multimedia server, a session server, etc., and the application server 10 may generally include an application database, etc., and the system composition may be selected according to actual needs, and this embodiment will not be described in detail here.

Referring to the system structure schematic diagram shown in fig. 1, as shown in fig. 2, an embodiment of the present invention provides a flowchart of an object recommendation method, where the method provided by the embodiment may be executed by an application server, and may specifically include the following steps:

Step S101, obtaining attribute categories of all objects output by an application platform;

in this embodiment, for convenience in describing attribute information of each type of attribute, each type of attribute may be referred to as a first attribute, a second attribute, etc., and attribute information of a corresponding type of attribute may be referred to as first attribute information, second attribute information, etc., where each first attribute information and each second attribute information may include attribute information of different contents, and this embodiment is not listed here one by one, and it should be noted that the "first" and "second" are not meant to be used only to distinguish the attribute of each type and the attribute information corresponding to each type of attribute.

Therefore, when the object is expressed, the embodiment does not use a single attribute to express, but adopts a multi-element expression mode, namely a multi-category attribute to describe the content of the object, so that the expression of the content of the object is more accurate, and the prediction accuracy is further improved.

Taking the application platform as a shopping platform as an example, the object on the application platform can be a commodity, the commodity can be divided into commodities with different styles such as sports, gentlewomen, white collars and the like according to the styles of the commodity, and multi-level category information of the commodity can be obtained according to the types of the commodity, the styles of the commodity can be marked as first attributes, the categories of the commodity can be marked as second attributes, and accordingly, the first attribute information can be the style information of the commodity and the second attribute information can be the category information of the commodity, but the description of the first attribute and the second attribute in the embodiment is not limited, namely the attribute category of the obtained object is not limited to the two attributes given by the embodiment, and the attribute information of each category can comprise different contents such as style attribute and can comprise sports style, gentlewoman style, white collar style and the like; the category attributes may include shoes, hats, shirts, skirts, and the like.

The commodities are often not displayed on the same interface of the shopping platform and are usually displayed in different levels, so that the categories of the commodities can be multi-level categories which can be generated based on the display levels of the commodities, and therefore, the information of the categories of the commodities can be obtained and can be multi-level category information composition, so that the display condition of the shopping platform on each commodity is determined, namely the display level relation of each commodity is determined.

Optionally, in this embodiment, the attribute types of all objects on the application platform may be determined by detecting the attributes of all the commodities on the application platform, and the specific detection method is not limited, and then all the attribute types of all the objects included in the application platform may be obtained, or at least two attributes may be selected from the determined attribute types, so that the construction of the attribute list of the subsequent objects is completed, and the specific selection of which attribute types is not limited.

Step S102, generating an object attribute list according to the acquired attribute category;

in this embodiment, taking the number of the acquired objects as two, that is, the first attribute and the second attribute as examples, a two-dimensional list may be generated by using the acquired binary attribute, the list may be denoted as an object attribute list, all rows thereof represent one attribute, all columns thereof represent another attribute, for example, each row thereof may represent the first attribute of the object, each column thereof may represent the second attribute of the object, that is, a row header of the object attribute list is represented by different attribute information of the first attribute, and a column header is represented by different attribute information of the second attribute.

In the description of the shopping platform, the first attribute is a style of the commodity, and the second attribute is a category of the commodity, and thus the object attribute list may be configured in a form shown in table 1, but is not limited to the form shown in table 1.

TABLE 1

In table 1, the article nm represents the article in the nth row and the mth column in the object attribute list, and it is known from the object attribute list that each article may have two attribute descriptions, the corresponding first attribute information of the articles in the same row is the same in content, the corresponding second attribute information of the articles in the same column is the same in content, that is, the articles in the same row may share the first attribute information of the same content, and the articles in the same column may share the second attribute information of the same content.

Based on this, in practical application, after determining the position of the commodity in the object attribute list, the attribute information of the line and the column where the commodity is located may describe the content of the commodity together, for example, the commodity 24 may be a piece of lady skirt, the commodity 11 is a pair of sports shoes, and the like, and according to the above manner, each object on the application platform may be mapped into the object attribute list.

The number of rows and columns of the object attribute list may be determined according to the number of objects of various attributes, in the example, the maximum number of objects corresponding to various attributes is taken as the number of rows of the two-dimensional list, and the number of categories is taken as the number of columns, so as to construct the two-dimensional list, i.e. the object attribute list of the application platform.

It should be noted that, the representation of the object attribute list of the application platform is not limited to the form shown in table 1, but may be a two-dimensional matrix, and each object on the application platform may be represented by a two-dimensional array, and the content of the two-dimensional array is not described in detail in this embodiment. Besides the binary attribute representation mode, the object can also adopt other multi-element attribute representation modes, and the specific implementation method is similar, and the detailed description of the embodiment is omitted here.

Optionally, taking the shopping platform with the first attribute as style and the second attribute as category as an example, based on the multi-level category information of the commodity, the embodiment may use a tree structure to represent word vectors of the multi-level category information, when the feature vector of the object needs to be obtained, the word vectors corresponding to the multi-level category information corresponding to the object may be obtained, and then the word vectors are spliced to obtain the word vectors corresponding to the total category attribute information of the object, so as to perform subsequent processing.

Similarly, the word vectors of the second attribute information of each column of the objects in the object attribute list may also be obtained by stitching in the manner described above, that is, the word vectors of the category information corresponding to each column of the objects in the object attribute list are obtained, and then the word vectors corresponding to each object in the same column are stitched to obtain the second word vector shared by the column of objects, but the generation manner of the word vector corresponding to each column of the attribute information in the object attribute list is not limited to the manner described herein

It should be noted that, for the object attribute list of the application platform in this embodiment, the object attribute list may be configured on-line, so that the application server of the application platform may use the object attribute list to predict the preference of the current user.

Step S103, acquiring historical behavior data of a user;

in an embodiment, a user operates an object output by an application platform, and generated behavior data can be stored as a piece of historical behavior data in a database of the application platform for subsequent call, so that the database of the application platform can store a plurality of pieces of historical behavior data generated at different operation times, and the content contained in each piece of historical behavior data is not limited in this embodiment.

Optionally, each historical behavior data may be stored in association with a user identifier of a corresponding user, and the behavior data generated by the object operation of each user on the application platform is distinguished by each user identifier, where the user identifier may be a user account number, a terminal identifier, etc., and the storage form of the content of the user identifier and the historical behavior data is not limited in this embodiment.

Step S104, analyzing the historical behavior data to obtain a plurality of target objects and corresponding operation sequences of the user;

Still taking the shopping platform as an example for illustration, the historical behavior data may be the historical behavior data generated when the user purchases the commodity, and the generation time of the historical behavior data is the time when the user purchases the commodity, so that in this embodiment, each piece of historical behavior data may be sorted according to the time sequence of the user purchasing the commodity, and then each piece of historical behavior data is analyzed to obtain the corresponding purchased commodity, that is, the target object.

Therefore, according to the method and the device, through analyzing the historical behavior data of the user on the application platform, the target objects of the user can be determined, and the access to the target objects is realized according to the operation sequence, so that the access sequence of the user to the objects on the application platform can be considered when the object interested by the user is predicted, and further the interest transition and interest accumulation of the user can be accurately positioned.

Step 105, obtaining row attribute information and column attribute information of each target object in the object attribute list according to the object attribute list;

step S106, according to the row attribute information and the column attribute information, obtaining the feature vector of the corresponding target object;

Optionally, in combination with the description about the object attribute list, for any target object of the user, attribute information, that is, row attribute information and column attribute information, corresponding to the row and column where the target object is located may be obtained from the object attribute list, so that feature vectors of the corresponding target object are generated by using the two attribute information, and a specific generating method is not limited, and may refer to the description of the corresponding embodiment below, but is not limited to the implementation method listed below.

It can be seen that the feature vector of each target object obtained in this embodiment characterizes multiple attribute information of the corresponding target object, that is, the content of the target object in this embodiment is described by multiple attribute information, and each attribute information corresponds to a type of attribute.

It should be noted that, the method for acquiring the corresponding feature vector of the target object is not limited to the implementation method using the object attribute list described in the embodiment, and may be implemented by a matrix, an array, or the like, and the implementation method is similar, which is not described in detail herein.

Step S107, according to the operation sequence of each target object, inputting the corresponding feature vectors into a prediction model in turn to obtain predicted feature vectors;

Optionally, the prediction model of this embodiment may be obtained by training feature vectors of a plurality of sample target objects based on a cyclic neural network, where the cyclic neural network may include a plurality of Long Short-Term Memory (LSTM) layers, that is, a hidden layer of the cyclic neural network is formed by a plurality of LSTM layers, and the calculation of the hidden layer is implemented by using the LSTM principle, which is not described in detail herein.

The LSTM is used for solving the problem of long-term dependence of the traditional circulating neural network, solves the problems of memory/forgetting, input degree and output degree of the memory unit by introducing the gating unit, and can learn when to open each door to what degree by a certain learning.

Referring to the schematic structure of the prediction model shown in fig. 3, it can be divided into an input layer, a hidden layer and an output layer, wherein the input layer inputs feature vectors corresponding to each momentAs X in FIG. 3 _t-1 、X _t、 X _t+1 Etc., the hidden layer may include a plurality of LSTM layers (such as structures in the blocks in fig. 3), and as known from the structures shown in fig. 3, in the recurrent neural network, the output at the previous time is added to the input at the current time, and the output at the next time is obtained through the tanh activation function, and the addition of LSTM adds three gates to the hidden layer, which specifically adds the output h at the previous time _t-1 Input X to the current time _t As a whole, as the input of the current hidden layer, the input of three gates is controlled simultaneously, the obtained residual part continues to move forward through the forgetting gate, and the input information in the state is added (as in the middle box in FIG. 3, the tanh operation is performed on the input, and the value range is controlled to be [ -1, 1)]In the range), the input signal processed by the input gate is added up with the residual department of the last gate to obtain the cell state at the current moment, namely, the updating from the last moment to the current moment is completed. Finally, the function of the output gate is to multiply the hidden layer state information after the tanh operation to obtain the output at the current moment. In fig. 3, σ represents a coefficient in the corresponding gate operation process in LSTM, the numerical value of the coefficient is not limited in this embodiment, and h represents an output of the hidden layer at the corresponding time.

Based on the above description, the output of the hidden layer at the current moment may be formed by the current input and the output of the hidden layer at the previous moment, specifically, the prediction feature vector obtained by the current calculation may be respectively subjected to inner products with word vectors of various attributes, and the larger the inner product is, the more preference is given to the object of the corresponding attribute, so that the weights of the two attributes may be determined.

It can be seen that the expression of the hidden layer at each moment may be that, according to the behavior sequence information of the target object before the moment (such as the feature vector of each target object), the preference expression of the target object is obtained, and it should be noted that the obtained preference expression of the user does not necessarily correspond to the existing object, and may be a sparse approximation expression.

For example, for the user a, the historical behavior data is processed in the above manner, and the predicted feature vector obtained by the prediction model can express that the user a wants to buy shoes and sportswind clothes (the shopping platform corresponding to table 1 is described as an example) at the next moment, so that the user's current preference is closest to the sportsshoes through the hidden layer.

In practical application of this embodiment, before the application server predicts the object for the user, the prediction model for implementing the prediction calculation may be trained in advance, in this embodiment, the feature vectors of the target objects of the plurality of sample users may be obtained by referring to the above manner, these feature vectors are used as training data, and based on the recurrent neural network algorithm, these training data are continuously trained until the model is completely converged, that is, the prediction result is substantially the same as the actual result, the error is within a certain range, and the training may be stopped, and the finally obtained model is used as the prediction model. The model training process described above will not be described in detail in this embodiment.

In addition, in combination with the above description of the LSTM principle, in the training process of the prediction model, the embodiment may utilize the obtained training data of multiple sample users to perform gradient back-propagation learning on parameters in the prediction model, that is, a random gradient descent method is used, according to residuals (such as the residual portions described above) from a target result, from a functional relationship, the residuals are transferred to each layer of the recurrent neural network through gradients, so as to modify each parameter of the recurrent neural network, and thus the output result is closer to a real result.

In the model training process, the feature vector of each sample target object may be obtained by referring to the above method for obtaining the feature vector of the target object, which is not described in detail in this embodiment.

In addition, it should be noted that, if the data amount of the object output by the obtained application platform is insufficient and the calculation amount is limited, the present invention may also use other modes to train the training data to obtain prediction models with other structures, for example, markov models, collaborative filtering, matrix decomposition and other algorithms to implement model training on the training data, and the specific training process may be determined based on the principles of each algorithm, which is not described in detail herein. As can be seen from this, the above-described prediction model is not limited to the model structure shown in fig. 3, and this embodiment will be described by way of example only.

Step S108, calculating the similarity between the feature vector corresponding to each candidate object contained in the object attribute list and the predicted feature vector;

in this embodiment, each object on the application platform may be used as a candidate object, and the object attribute list has corresponding attribute information, and the feature vector corresponding to each candidate object may be obtained according to the above-described manner, which is not described herein in detail.

Optionally, the target object in this embodiment may be a click object that the user performs a clicking operation, so that, based on factors such as a click rate of an object output by the application platform and a click rate on another application platform, all the objects output by the application platform may be screened to obtain multiple candidate objects, and then feature vectors of the candidate objects are obtained according to the above manner. It should be noted that the candidate object may include a target object that is not currently output, and the operation of the user on the object is not limited to clicking, and for other operations, the basis for screening the candidate object may be adjusted accordingly, which is not limited to the screening manner described in this embodiment.

In addition, the similarity calculation method between vectors in this embodiment is not limited, such as a cosine similarity calculation method, a distance calculation method, and the like, and may be implemented according to the principle of the corresponding similarity calculation method, which is not described in detail herein.

Step S109, screening out at least one candidate object as a recommended object based on the similarity of each candidate object.

The higher the calculated similarity is, the greater the probability that the corresponding candidate object is screened as the recommended object is, and the screening method for realizing the recommended object based on the similarity is not limited in this embodiment.

Alternatively, the ranking may be performed according to the calculated similarity of each candidate object, so that a specific number of candidate objects with the highest similarity are selected as recommended objects of the user, that is, a specific number of candidate objects are selected as recommended objects according to the order of the similarity from high to low, and of course, a similarity threshold may be preset, so that candidate objects with a similarity greater than the similarity threshold are selected as recommended objects, and so on.

In practical application of the embodiment, the screened recommended objects can be pushed to the user client for display, so that the user is assisted in quickly selecting appropriate objects.

In summary, referring to the schematic diagram shown in fig. 4, in this embodiment, multiple attribute information of the target object is obtained, a feature vector of the target object is constructed, and is used as an input of a prediction model, so that a predicted feature vector of a user is obtained by calculation, not only is the expression content of the feature vector enriched, but also because a plurality of objects output by the application platform of this embodiment share one attribute information, attribute information required by the feature vector forming each target object can be reused, the number of parameters of the prediction model is greatly reduced, the problems of serious over-expression and sparse data cold start are avoided, so that model training becomes easier, and the application platform is applicable to a large number of application scenes; in addition, the embodiment carries out model training based on the cyclic neural network, also considers the data time sequence, and improves the accuracy of the prediction result.

As an alternative embodiment of the present invention, another process of obtaining feature vectors corresponding to each target object according to the object attribute list is proposed, which may also be considered as a specific implementation method of the above step S105 and step S106, but the implementation method of the steps S105 and S106 is not limited to the implementation manner described in this embodiment, and regarding the generating process of the object attribute list of the application platform, the description of the corresponding part of the above embodiment may be referred to, and this embodiment mainly describes how to obtain the feature vectors of the target object, as shown in the flowcharts of fig. 5 and 6, where the method may include, but is not limited to, the following steps:

Step S201, obtaining the position information of each target object in an object attribute list;

step S202, based on the position information, acquiring a first word vector of a row where a corresponding target object is located and a second word vector of a column where the corresponding target object is located;

taking the object attribute list shown in table 1 as an example, if the target object is the commodity 24, the position of the commodity in the object attribute list, that is, the 2 nd row and the 4 th column, may be searched, and then the attribute information corresponding to the 2 nd row of the object attribute list (that is, the row attribute information of the gentlewoman wind), that is, the row attribute information shared by each commodity in the row where the commodity 24 is located, and the attribute information of the 4 th column (that is, the column attribute information shared by each commodity in the column where the commodity 24 is located) may be obtained.

Then, in this embodiment, the first word vector may be generated from the obtained row attribute information of the target object, and the second word vector may be generated from the column attribute information, which is not limited in the specific generation process. As can be seen from the object attribute list, the objects in the same row share one row of attribute information, and further share the same first word vector, and the objects in the same column share one column of attribute information, and further share the same second word vector, so that the first word vector and the second word vector of the object do not need to be built for each object, and the step of initializing semantic expression for each object in the object attribute list is simplified. The word vector in this embodiment may be a semantic expression manner of the corresponding attribute information of the object, and the method of how to generate the corresponding word vector by using the attribute information in this embodiment is not limited.

Taking the object attribute list shown in table 1 as an example, the first word vector of the line of commodities of the moving wind (i.e. a line of attribute information) can be obtained by performing semantic analysis on the motion; similarly, a first word vector corresponding to a line of goods such as gentlewoman wind (i.e. another line of attribute information) can be obtained, a second word vector corresponding to a line of goods such as shoes (i.e. another line of attribute information) and a second word vector corresponding to a line of goods such as skirt (i.e. another line of attribute information) can be obtained, and the like, and each of the goods can be expressed together by one first word vector and one second word vector, and the first word vectors of the goods positioned in the same line are the same, and the second word vectors of the goods positioned in the same line are the same.

In this embodiment, the specific implementation method of the first word vector of each row of objects and the second word vector of each column of objects in the object attribute list is not limited.

Step S203, fusion is carried out on the first word vector and the second word vector corresponding to the same target object, and feature vectors corresponding to the target objects are obtained.

The method of how to realize the fusion of the first word vector and the second word vector corresponding to the same target object is not limited in this embodiment, for example, the first word vector and the second word vector of the same target object are directly spliced, and the obtained vector is used as the feature vector of the click vector of the user, so that the feature vector can simultaneously represent two types of attribute information of the target object.

Optionally, the present embodiment may use a gating unit to implement fusion of the first word vector and the second word vector corresponding to each target object, where the gating unit is intended to process and select a specific gravity of the row attribute information and the column attribute information in the next selection action of the user under the current object expression, and to fuse word vectors corresponding to the two attribute information according to such specific gravity, so as to obtain an optimal recommended feature expression vector, and this embodiment marks the optimal recommended feature expression vector as a feature vector. Therefore, the two word vectors are fused through the gating unit, so that the thinking characteristics of the user in the process of accessing the object are simulated, the flexibility is higher, and the personalized recommendation requirements of different users are met.

In summary, in the embodiment, when the feature vector of the target object is constructed, multiple attribute information constructions of the target object are obtained, object contents expressed by the feature vector are enriched, and because the objects located in the same row or the same column in the object attribute list share the same attribute information, the feature vector located in the same row or the same column can be shared when the feature vector of the object is clicked, so that the parameter number of the prediction model is greatly reduced, and the prediction efficiency and accuracy are improved.

Optionally, in the process of training the prediction model, the embodiment may also construct the feature vector according to the above manner, that is, construct the feature vector in a manner of sharing word vectors, so as to greatly reduce the number of parameters of the training model and greatly reduce the training difficulty of the prediction model.

Optionally, in order to improve the prediction accuracy of the prediction model, in the training process of the prediction model, the position of each object in the object attribute list may be optimized according to the following manner, which is not limited to the method provided in this embodiment, as shown in fig. 7, and the method may include:

step S301, acquiring row attribute information of an actual object;

the actual object may be a recommended object selected as described above, such as a commodity actually purchased by the user.

Step S302, for a plurality of objects in any column of the object attribute list, calculating the similarity between the row attribute information of the plurality of objects and the row attribute information of the actual object;

In this embodiment, the method for calculating the similarity of the line attribute information between two objects is not limited, for example, after generating corresponding word vectors, the similarity between the corresponding word vectors is calculated to represent the similarity of the line attribute of the two objects, and if the line attribute is the style attribute of the commodity, the similarity may represent the similarity of the two objects in style.

Step S303, based on the calculated similarity, the positions of the objects in the object attribute list are adjusted according to a bipartite graph matching algorithm.

For example, if the recommended object is three pairs of shoes, respectively denoted as q1, q2 and q3, and q1 is purchased by the user in the training process using the prediction model, the similarity of the three pairs of shoes on the style attributes s1, s2 and s3 is (0.1,0.5,0.1), (0.3,0.1,0.2) and (0.3,0.3,0.7), respectively, and then, according to the bipartite graph matching algorithm, it can be known that the global similarity is the highest when q1 matches s2, q2 matches s1 and q3 matches s3, and therefore, q1 can be allocated to the style attributes s2, q2 can be allocated to the style attributes s1 and q 3. The bipartite graph matching algorithm can adopt Minimum-cost flow problem and uses a graph analysis algorithm package of google OR-tools to optimize the positions of all objects in the obtained object attribute list so as to improve the accuracy of object prediction realized according to the positions.

In combination with the above analysis, the application of the object recommendation method is described by taking a commodity recommendation scene as an example, an application platform under the scene can be a shopping platform, a target object can be a commodity purchased by a user, based on the fact, referring to a flow diagram shown in fig. 8, the user can send an access request to an application server through a client, log in the shopping platform and browse various commodities output by the shopping platform, at the moment, the corresponding application server can acquire historical behavior data of the commodity purchased by the user based on a user identifier carried in the access request after receiving the access request, and can acquire the commodity purchased by the user through analysis of the historical behavior data, the commodity can be recorded as a historical commodity, then, for each historical commodity, the position of the historical commodity in an object attribute list constructed by the shopping platform can be determined, so that style information and category information of the historical commodity in the line can be acquired, and corresponding first word vectors and second word vectors can be generated, and feature vectors of the historical commodity can be obtained through fusion processing.

And then, according to the purchasing sequence of each historical commodity, inputting the feature vector corresponding to each historical commodity into a trained prediction model, wherein the output vector is the prediction feature vector used for predicting the purchasing preference of the user, and calculating the similarity between the feature vector of each candidate commodity and the prediction feature vector by taking the feature vector as the standard for selecting the recommended commodity, wherein the higher the similarity is, the greater the likelihood that the user purchases the candidate commodity is indicated.

In the method, the feature vectors of the commodities purchased by the plurality of sample users are obtained by constructing the prediction model, namely, latent semantic expressions of modeling objects are refined, so that the recommendation prediction accuracy of the prediction model is improved. And the hierarchical category attribute and the preference attribute (namely, style attribute) are decoupled and used as independent attributes to determine corresponding semantic expressions, so that the prediction model can clearly analyze the habit of a user, and other works, such as collocation of commodities and the like, can be known based on the obtained prediction category and preference attribute,

referring to fig. 9, a block diagram of an object recommendation apparatus according to an embodiment of the present invention may be applied to an application server, and specifically may include, but is not limited to, the following composition structures:

a feature vector obtaining module 91, configured to obtain feature vectors corresponding to a plurality of target objects respectively;

wherein each feature vector characterizes a plurality of attribute information of a corresponding target object, and the attribute information is shared by a plurality of target objects.

The predicted feature vector calculation module 92 is configured to sequentially input corresponding feature vectors into a prediction model according to the operation sequence of the plurality of target objects to obtain predicted feature vectors;

In this embodiment, the prediction model may be obtained by training feature vectors corresponding to a plurality of sample target objects, and may specifically be implemented based on a cyclic neural network, collaborative filtering, matrix decomposition, and other modes, where a specific training mode of the prediction model is not limited in this embodiment.

According to the analysis, when model training is carried out, the training data comprise various attribute information, so that training features are richer, one type of object on the application platform can share one attribute information, special attribute information is not required to be set for each object, the number of model parameters is reduced, model training difficulty is reduced, and model prediction accuracy is improved.

A similarity calculation module 93, configured to calculate a similarity between the predicted feature vector and the feature vector corresponding to each candidate object;

and a recommended object screening module 94, configured to screen at least one candidate object as a recommended object based on the similarity corresponding to each candidate object.

Wherein, the higher the similarity of the candidate object, the greater the probability of being screened as a recommended object.

It should be noted that, the method of how to screen the recommended objects based on the similarity is not limited in this embodiment, for example, a specific number of candidate objects may be selected as recommended objects according to the order of the similarity of the candidate objects from high to low, or candidate objects whose similarity reaches a preset threshold may be screened as recommended objects, and so on.

Optionally, as shown in fig. 10, the apparatus may further include:

an attribute obtaining module 95, configured to obtain attribute categories of the objects;

the attribute category of the object may include at least a first attribute and a second attribute.

An object attribute list generating module 96, configured to generate an object attribute list according to the acquired attribute types;

wherein, the row header of the object attribute list is represented by the different attribute information of the first attribute of each object, the column header is represented by the different attribute information of the second attribute of each object, and the specific construction method of the object attribute list can refer to the description of the embodiment of the method, as shown in the above table 1, but is not limited to the form shown in the table 1.

Accordingly, the feature vector obtaining module 91 may include:

the attribute information acquisition unit is used for acquiring row attribute information and column attribute information of each target object in the object attribute list according to the object attribute list;

and the characteristic vector acquisition unit is used for acquiring the characteristic vector of the corresponding target object according to the row attribute information and the column attribute information.

As another alternative embodiment, as shown in fig. 11, the object recommendation apparatus may further include:

A data acquisition module 97 for acquiring historical behavior data of the user, the historical behavior data being generated based on an operation of the user on the application platform output object;

the data analysis module 98 is configured to analyze the historical behavior data to obtain a plurality of target objects of the user and an operation sequence of the plurality of target objects;

the feature vector acquisition module 91 includes:

a location information obtaining unit 911 for obtaining location information of each target object in the object attribute list;

a word vector obtaining unit 912, configured to obtain, based on the location information, a first word vector of a row where the corresponding target object is located and a second word vector of a column where the corresponding target object is located;

optionally, if the first attribute includes a style of the object, the second attribute includes a category of the object, the second word vector of each column of the object in the object attribute list may be obtained by generating a word vector from category information of each object in the column and fusing the word vector (such as a stitching manner, but not limited to such a fusing manner), and similarly, the first word vector may also be obtained in such a manner, and based on this, the object recommendation apparatus may further include:

the first word vector acquisition module is used for acquiring first attribute information corresponding to each object in the same row of the object attribute list;

The first word vector generation module is used for generating word vectors of corresponding objects according to the first attribute information;

and the first word vector splicing module is used for splicing word vectors corresponding to the objects in the same row to obtain a first word vector shared by the objects in the row of the object attribute list.

The second word vector acquisition module is used for acquiring second attribute information corresponding to each object in the same column of the object attribute list;

the second word vector generation module is used for generating word vectors of the corresponding objects according to the second attribute information;

and the second word vector splicing module is used for splicing word vectors corresponding to the objects in the same column to obtain a second word vector shared by the objects in the column of the object attribute list.

The word vector fusion unit 913 is configured to fuse the first word vector and the second word vector of the same target object to obtain feature vectors corresponding to the target objects.

Optionally, the word vector fusion unit 913 may be specifically configured to splice the first word vector and the second word vector of the same target object to obtain a feature vector of the target object; or fusing the first word vector and the second word vector of the same target object through a gating unit to obtain the feature vector of the corresponding target object, but the method is not limited to the two fusion processing modes.

Optionally, on the basis of the foregoing embodiments, the object recommendation apparatus may further include:

the system comprises an actual object attribute information acquisition module, a recommendation module and a display module, wherein the actual object attribute information acquisition module is used for acquiring row attribute information of an actual object, and the actual object is a screened recommended object;

the line attribute similarity calculation module is used for calculating the similarity between the line attribute information of a plurality of objects and the line attribute information of an actual object for the plurality of objects in any column of the object attribute list;

and the adjusting module is used for adjusting the positions of the objects in the object attribute list according to a bipartite graph matching algorithm based on the calculated similarity.

In practical application of the embodiment, in the model training process based on the generated object attribute list, each training is finished, a prediction result obtained based on the current training is compared with an actual result, and all objects in the object attribute list are mapped to the position with the highest possibility based on the comparison result, so that the object attribute list is updated.

After the actual prediction is performed by using the prediction model, the position of the object in the object attribute list may be continuously adjusted based on the prediction result and the actual operation object, so that the position of each object in the object attribute list is most matched with the corresponding row attribute and the corresponding column attribute, that is, the attribute information of the row where the object is located and the attribute information of the column where the object is located can most accurately express the content of the object, thereby improving the accuracy of the subsequent prediction based on the object attribute list.

In summary, the embodiment acquires various attribute information of the target object, constructs the feature vector of the target object, and uses the feature vector as the input of the prediction model to calculate the predicted feature vector of the user, so that not only the expression content of the feature vector is enriched, but also the attribute information required by the feature vector of each target object can be reused because the plurality of objects output by the application platform of the embodiment share one attribute information, thus greatly reducing the parameter number of the prediction model, avoiding the problems of serious over-expression and sparse data and cold start, facilitating model training, and being applicable to huge application scenes; when predicting the object attribute of interest of the user, the data time sequence is considered, and the accuracy of the prediction result is improved.

The embodiment of the invention also provides a storage medium, on which a computer program is stored, the computer program is executed by a processor, and each step of the object recommendation method is implemented, and the implementation process of the object recommendation method can be described with reference to the embodiment of the method.

As shown in fig. 12, an embodiment of the present invention further provides a hardware structure schematic of a computer device, where the computer device may be an application server implementing the above object recommendation method, and may include a communication interface 121, a memory 122, and a processor 123;

In the embodiment of the present invention, the communication interface 121, the memory 122, and the processor 123 may implement communication between each other through a communication bus, and the number of the communication interface 121, the memory 122, the processor 123, and the communication bus may be at least one.

Alternatively, the communication interface 121 may be an interface of a communication module, such as an interface of a GSM module;

the processor 123 may be a central processing unit CPU, or a specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement embodiments of the present invention.

Memory 122 may comprise high-speed RAM memory or may further comprise non-volatile memory (non-volatile memory), such as at least one disk memory.

Wherein the memory 122 stores a program, and the processor 123 calls the program stored in the memory 122 to implement the steps of the object recommendation method applied to the computer device;

alternatively, the program may be used mainly for:

It should be noted that, regarding other steps of the method for recommending an object by executing a program by a processor, reference may be made to the description of corresponding parts of the above-described method embodiment.

In the present specification, each embodiment is described in a progressive manner, and each embodiment is mainly described in a different point from other embodiments, and identical and similar parts between the embodiments are all enough to refer to each other. The apparatus, the storage medium, and the computer device disclosed in the embodiments are relatively simple to describe, and the relevant parts refer to the description of the method section because they correspond to the methods disclosed in the embodiments.

Those of skill would further appreciate that the various illustrative elements and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both, and that the various illustrative elements and steps are described above generally in terms of functionality in order to clearly illustrate the interchangeability of hardware and software. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.

The steps of a method or algorithm described in connection with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. The software modules may be disposed in Random Access Memory (RAM), memory, read Only Memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An object recommendation method, the method comprising:

acquiring attribute categories of all objects, wherein the attribute categories at least comprise a first attribute and a second attribute;

Generating an object attribute list according to the acquired attribute category, wherein a row title of the object attribute list is represented by different attribute information of the first attribute, and a column title is represented by different attribute information of the second attribute;

the method for acquiring the feature vectors respectively corresponding to the plurality of target objects specifically comprises the following steps: obtaining row attribute information and column attribute information of each target object in the object attribute list according to the object attribute list, and obtaining feature vectors of the corresponding target objects according to the row attribute information and the column attribute information; each feature vector characterizes various attribute information of a corresponding target object, and the attribute information is shared by a plurality of target objects;

2. The method of claim 1, wherein the predictive model is obtained by training feature vectors corresponding to a plurality of sample target objects, respectively, based on a recurrent neural network, and the recurrent neural network comprises a plurality of long-short-term memory network layers.

3. The method according to claim 1, wherein the method further comprises:

acquiring historical behavior data of a user, wherein the historical behavior data is generated based on the operation of the user on an output object of an application platform;

analyzing the historical behavior data to obtain a plurality of target objects of the user and an operation sequence of the plurality of target objects;

the obtaining feature vectors corresponding to the plurality of target objects respectively specifically includes:

acquiring position information of each target object in an object attribute list;

based on the position information, acquiring a first word vector of a row where a corresponding target object is located and a second word vector of a column where the corresponding target object is located;

and fusing the first word vector and the second word vector of the same target object to obtain feature vectors corresponding to the target objects.

4. The method according to claim 1, wherein the method further comprises:

acquiring row attribute information of an actual object, wherein the actual object is a screened recommended object;

for a plurality of objects in any column of the object attribute list, calculating the similarity between the row attribute information of the plurality of objects and the row attribute information of the actual object;

And based on the calculated similarity, adjusting the position of each object in the object attribute list according to a bipartite graph matching algorithm.

5. The method according to claim 1 or 2, wherein screening at least one candidate object as a recommended object based on the similarity corresponding to each candidate object comprises:

and selecting a specific number of candidate objects as the recommended objects according to the sequence of the similarity of the candidate objects from high to low.

6. The method of claim 3, wherein fusing the first word vector and the second word vector of the same target object to obtain feature vectors corresponding to the target objects includes:

splicing the first word vector and the second word vector to obtain a feature vector corresponding to the target object; or,

and fusing the first word vector and the second word vector through a gating unit to obtain a feature vector corresponding to the target object.

7. A method according to claim 3, characterized in that the method further comprises:

acquiring first attribute information corresponding to each object in the same row of the object attribute list;

Generating word vectors of corresponding objects according to the first attribute information;

and splicing word vectors corresponding to the objects in the same row to obtain a first word vector shared by the objects in the row of the object attribute list.

8. A method according to claim 3, characterized in that the method further comprises:

acquiring second attribute information corresponding to each object in the same column of the object attribute list;

generating word vectors of the corresponding objects according to the second attribute information;

and splicing word vectors corresponding to the objects in the same column to obtain a second word vector shared by the objects in the column of the object attribute list.

9. An object recommendation device, the device comprising:

the attribute acquisition module is used for acquiring attribute categories of all the objects, wherein the attribute categories at least comprise a first attribute and a second attribute;

the object attribute list generation module is used for generating an object attribute list according to the acquired attribute category, wherein the row titles of the object attribute list are represented by different attribute information of the first attribute, and the column titles are represented by different attribute information of the second attribute;

the characteristic vector acquisition module is used for acquiring characteristic vectors corresponding to a plurality of target objects respectively, each characteristic vector characterizes various attribute information of the corresponding target object, and the attribute information is shared by the plurality of target objects;

The prediction feature vector calculation module is used for sequentially inputting corresponding feature vectors into the prediction model according to the operation sequence of the plurality of target objects to obtain prediction feature vectors of the corresponding target objects;

the recommendation object screening module is used for screening at least one candidate object as a recommendation object based on the similarity corresponding to each candidate object;

the feature vector acquisition module includes:

10. The apparatus of claim 9, wherein the predictive model is based on training feature vectors corresponding to a plurality of sample target objects, respectively, on a recurrent neural network, and the recurrent neural network comprises a plurality of long-short-term memory network layers.

11. The apparatus of claim 9, wherein the apparatus further comprises:

The data acquisition module is used for acquiring historical behavior data of a user, wherein the historical behavior data is generated based on the operation of the user on the output object of the application platform;

the data analysis module is used for analyzing the historical behavior data to obtain a plurality of target objects of the user and an operation sequence of the plurality of target objects;

the feature vector acquisition module specifically comprises:

a position information acquisition unit for acquiring position information of each target object in the object attribute list;

the word vector acquisition unit is used for acquiring a first word vector of a row where a corresponding target object is located and a second word vector of a column where the corresponding target object is located based on the position information;

and the word vector fusion unit is used for fusing the first word vector and the second word vector of the same target object to obtain feature vectors corresponding to the target objects.

12. The apparatus of claim 9, wherein the apparatus further comprises:

13. The apparatus according to claim 9 or 10, wherein the recommendation object screening module is specifically configured to:

14. The apparatus according to claim 11, wherein the word vector fusion unit is specifically configured to:

15. The apparatus of claim 11, wherein the apparatus further comprises:

16. The apparatus of claim 11, wherein the apparatus further comprises:

the second word vector generation module is used for generating word vectors of corresponding objects according to the second attribute information;

17. A storage medium having stored thereon a computer program, wherein the computer program is executed by a processor to implement the steps of the object recommendation method according to any one of claims 1-8.

18. A computer device, the computer device comprising:

a communication interface;

a memory for storing a program for implementing the object recommendation method according to any one of claims 1 to 8;