Recommended method, device, computing device and computer storage medium

Abstract

The present invention discloses a recommendation method, device, computing equipment and computer storage medium. The method includes: obtaining user data of a user to be recommended; inputting the user data into a pre-trained recommendation model, the recommendation model is obtained by training the first sample data and the second sample data of a sample user, the first sample data includes the object features of a first recommended object, and the sample user has feedback behavior on the first recommended object; the second sample data includes the object features of a second recommended object, and the sample user has no feedback behavior on the second recommended object, the recommendation model includes a first objective function for processing the first sample data, and a second objective function for processing the second sample data; obtaining the output result of the recommendation model, and obtaining the target recommended object based on the output result. The adoption of this solution can avoid sample bias in the recommendation model during training, and improve the prediction accuracy of the recommendation model and the acquisition accuracy of the recommended object.

Description

Technical Field

[0001] The present invention relates to the technical field of information recommendation, and in particular to a recommendation method, device, computing equipment and computer storage medium. Background Art

[0002] With the continuous development of science and technology and society, the emergence of various Internet products has greatly enriched people's work and life. In order to improve user experience and increase user retention rate, many Internet products usually make corresponding recommendations for users, such as recommending videos, information, and / or e-books to users.

[0003] At present, some Internet products will make recommendations to users based on recommendation models. However, the inventors found in the implementation process that the prior art has the following defects: when training the recommendation model, the prior art often generates positive samples based on the information of the recommended objects for which users have given positive feedback, and generates negative samples based on the information of the recommended objects for which users have given negative feedback. However, since the recommended objects for which users have given positive and negative feedback are only a small part of the global recommended object set, the use of this training method will cause a bias in the training samples, reduce the prediction accuracy of the recommendation model, and thus cause the technical disadvantage of low recommendation accuracy. Summary of the invention

[0004] In view of the above problems, the present invention is proposed to provide a recommended method, apparatus, computing device and computer storage medium that overcome the above problems or at least partially solve the above problems.

[0005] According to a first aspect of the present invention, there is provided a recommendation method, comprising:

[0006] Acquire user data of the user to be recommended; wherein the user data includes feedback behavior data of the user to be recommended on the historical recommendation object within a preset historical time window;

[0007] Input the user data into a pre-trained recommendation model; wherein the recommendation model is obtained by training first sample data and second sample data of a sample user, the first sample data includes object features of a first recommended object, and the sample user has feedback behavior on the first recommended object; the second sample data includes object features of a second recommended object, and the sample user has no feedback behavior on the second recommended object; and the recommendation model includes a first objective function for processing the first sample data, and a second objective function for processing the second sample data;

[0008] The output result of the recommendation model is obtained, and based on the output result, a target recommendation object matching the user to be recommended is obtained.

[0009] According to a second aspect of the present invention, there is provided a recommendation device, comprising:

[0010] The first acquisition module is used to acquire user data of the user to be recommended; wherein the user data includes feedback behavior data of the user to be recommended on the historical recommendation object within a preset historical time window;

[0011] An input module, used for inputting the user data into a pre-trained recommendation model; wherein the recommendation model is obtained by training first sample data and second sample data of a sample user, the first sample data includes object features of a first recommended object, and the sample user has feedback behavior on the first recommended object; the second sample data includes object features of a second recommended object, and the sample user has no feedback behavior on the second recommended object; and the recommendation model includes a first objective function for processing the first sample data, and a second objective function for processing the second sample data;

[0012] The second acquisition module is used to acquire the output result of the recommendation model, and obtain the target recommendation object matching the to-be-recommended user based on the output result.

[0013] According to a third aspect of the present invention, there is provided a computing device, comprising: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus;

[0014] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to execute operations corresponding to the above-mentioned recommended method.

[0015] According to a fourth aspect of the present invention, a computer storage medium is provided, wherein the storage medium stores at least one executable instruction, and the executable instruction enables a processor to execute operations corresponding to the above-mentioned recommended method.

[0016] In the present invention, user data of the user to be recommended is obtained; the user data is input into a pre-trained recommendation model, the recommendation model is obtained by training the first sample data and the second sample data of the sample user, the first sample data includes the object features of the first recommended object, the sample user has feedback behavior on the first recommended object; the second sample data includes the object features of the second recommended object, the sample user has no feedback behavior on the second recommended object, the recommendation model includes a first objective function for processing the first sample data, and a second objective function for processing the second sample data; the output result of the recommendation model is obtained, and the target recommended object is obtained based on the output result. The adoption of this solution can avoid sample bias, improve the prediction accuracy of the recommendation model and improve the accuracy of obtaining the recommended object.

[0017] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0019] Figure 1 A schematic diagram of a process flow of a recommended method provided by an embodiment of the present invention is shown;

[0020] Figure 2 A schematic diagram of a flow chart of a training method for a recommendation model provided by an embodiment of the present invention is shown;

[0021] Figure 3 A schematic diagram of a process of determining a similar user provided by an embodiment of the present invention is shown;

[0022] Figure 4 A schematic diagram showing a flow chart of a method for obtaining similarity of interest points provided by an embodiment of the present invention;

[0023] Figure 5 A schematic diagram of interest degree change trend provided by an embodiment of the present invention is shown;

[0024] Figure 6 A data flow diagram of a recommendation model provided by an embodiment of the present invention is shown;

[0025] Figure 7 A schematic diagram of the structure of a recommended device provided by an embodiment of the present invention is shown;

[0026] Figure 8 A schematic diagram of the structure of a computing device provided by an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0027] The exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present invention and to enable the scope of the present invention to be fully communicated to those skilled in the art.

[0028] Figure 1FIG. 1 is a flow chart of a recommended method provided by an embodiment of the present invention. Figure 1 As shown, the method comprises the following steps:

[0029] Step S110, obtaining user data of the user to be recommended; wherein the user data includes feedback behavior data of the user to be recommended on the historical recommendation object within a preset historical time window.

[0030] The user to be recommended is a user to whom information is to be recommended later. The embodiment of the present invention can use the pre-generated recommendation model to determine the recommended objects that match each user to be recommended in real time. The recommended object specifically refers to the information recommended to the user, and the recommended object can be video, information, music, and / or electronic books, etc.

[0031] In the specific implementation process, for any user to be recommended, the user data of the user to be recommended is obtained. The user data contains the feedback behavior data of the user to be recommended on the historical recommendation object within the preset historical time window. Among them, the preset historical time window is a certain historical period. Optionally, in order to improve the recommendation accuracy, the preset historical time window is the most recent historical period, such as the most recent day, the most recent week, etc. The historical recommendation object of the user to be recommended refers to the recommendation object that has been recommended to the user to be recommended. The feedback behavior data is specifically the relevant information of the feedback behavior on the recommendation object, and the feedback behavior can be positive feedback behavior or negative feedback behavior. In addition, the user data may also include user attribute data such as user identity information.

[0032] This step can obtain the latest behavior data of the user to be recommended in real time, so as to facilitate the subsequent determination of the recommendation object currently matching the user to be recommended based on the latest behavior data of the user to be recommended.

[0033] Step S120, inputting the user data into a pre-trained recommendation model; wherein the recommendation model is trained by first sample data and second sample data of sample users, the first sample data includes object features of a first recommended object, and the sample user has feedback on the first recommended object; the second sample data includes object features of a second recommended object, and the sample user has no feedback on the second recommended object; and the recommendation model includes a first objective function for processing the first sample data, and a second objective function for processing the second sample data.

[0034] The embodiment of the present invention uses a pre-trained recommendation model to learn and process user data in real time. Unlike the prior art that only uses the recommended objects for which sample users have given feedback for training, the embodiment of the present invention uses both the first recommended object for which sample users have given feedback and the second recommended object for which sample users have not given feedback when training the recommendation model, thereby avoiding sample bias and improving the prediction accuracy of the recommendation model.

[0035] In the specific training process, for any sample user, the first recommended object and the second recommended object of the sample user are determined. The sample user has feedback on the first recommended object, which can be positive feedback or negative feedback; and the sample user has no feedback on the second recommended object. Then, the first sample data is generated based on the object features of the first recommended object of the sample user, and the second sample data is generated based on the object features of the second recommended object of the sample user. The first sample data and the second sample data together constitute the training samples of the recommendation model.

[0036] After obtaining the training samples, the training samples are used for model training. In the training process, the first sample data is processed using the first objective function of the recommendation model, and the second sample data is processed using the second objective function, thereby improving the processing accuracy of the recommendation model.

[0037] Step S130, obtaining the output result of the recommendation model, and obtaining the target recommendation object matching the user to be recommended based on the output result.

[0038] In an optional implementation, the recommendation model can directly output at least one recommended object that matches the user to be recommended through deep learning of user data, and the recommended object is the target recommended object that matches the user to be recommended.

[0039] In another optional implementation, after obtaining a trained recommendation model, in order to improve the efficiency of user recommendation using the recommendation model, the object vector of each recommended object can be extracted from the recommendation model, and the object vector of the recommended object can be stored in a preset retrieval system. Therefore, after processing user data using the recommendation model, the embodiment of the present invention can obtain a user vector matching the user data, and the user vector reflects the user's latest recommendation object preference information, thereby using the user vector output by the recommendation model to perform similarity matching with each object vector in the retrieval system to obtain at least one recommended object with high similarity, and the obtained recommended object is the target recommended object matched by the user to be recommended.

[0040] It can be seen that the embodiment of the present invention can recommend highly matched recommendation objects to the recommended users through the trained recommendation model, thereby realizing fast and accurate recall of the recommended objects; and the recommendation model in the embodiment of the present invention is trained using the information of the recommended objects for which the sample users have had feedback and the recommended objects for which they have not had feedback, thereby avoiding sample bias and improving the prediction accuracy of the recommendation model, thereby improving the accuracy of obtaining the recommended objects; and the embodiment of the present invention uses different objective functions to process the sample data generated for the recommended objects for which the sample users have had feedback and the recommended objects for which they have not had feedback, thereby further improving the prediction accuracy of the recommendation model and improving the accuracy of determining the recommended objects.

[0041] Figure 2 A flow chart of a training method for a recommendation model provided by an embodiment of the present invention is shown.

[0042] The training method provided by the embodiment of the present invention can be applied to Figure 1 In the preferred method embodiment shown.

[0043] like Figure 2 As shown, the method comprises the following steps:

[0044] Step S210: identifying a first recommended object and a second recommended object for any sample user from the global recommended object set.

[0045] When Internet products make user recommendations, they usually construct a global recommendation object set. The global recommendation object set contains a large number of recommendation objects that can be recommended to users. The recommendation objects can be videos, information, music, and / or e-books, etc. Taking the Internet product as a video platform as an example, the global recommendation object set contains all videos that can be recommended to users, and the video is the recommendation object.

[0046] When using Internet products, users will generate corresponding interaction data with recommended objects. For example, when using video platforms, users will click, watch, collect, forward, set as like, set as dislike, block, etc., thereby generating interaction data between users and recommended objects. In the prior art, only the relevant data of recommended objects with which users have interacted are often used to train recommendation models. However, the recommended objects with which users have interacted are only a small part of the exposed recommended object set, and the exposed recommended object set is only a small part of the global recommended object set. Therefore, the prior art will cause serious sample bias when training recommendation models, which will then result in low prediction accuracy of the recommendation models obtained through training.

[0047] The embodiment of the present invention is different from the prior art in that the embodiment of the present invention uses relevant information of a large amount of recommended objects in the global recommended object set to train the recommendation model, thereby making the corresponding recommended object set when training the recommendation model and subsequently using the recommendation model for recommendation the same, avoiding sample bias and improving the prediction accuracy of the recommendation model.

[0048] In a specific implementation process, for any sample user, the first recommended object and the second recommended object of the sample user are determined from the global recommended object set. The sample user specifically refers to an existing user in the Internet platform, and the Internet platform usually contains relevant user feature data of the user, etc. As an optional implementation, the embodiment of the present invention can identify the first recommended object and the second recommended object of each sample user by whether the interaction data between the sample user and the recommended object is stored in the Internet platform.

[0049] The first recommended object refers to the recommended object to which the sample user has provided feedback, that is, the Internet platform contains the interaction data between the sample user and the first recommended object. The corresponding feedback includes but is not limited to the following feedback: click, watch / read / play, favorite, forward, set as like, set as dislike, block, and / or delete, etc. That is, the corresponding feedback includes positive feedback and / or negative feedback.

[0050] The second recommended object refers to the recommended object to which the sample user has no feedback behavior, that is, there is no interaction data between the sample user and the second recommended object in the Internet platform. For example, the recommended object C1 has not been exposed in the terminal of the sample user U1, so there is no interaction data between the sample user U1 and the recommended object C1 on the Internet platform, then the recommended object C1 is the second recommended object of the sample user U1; or, the recommended object C2 has been exposed in the terminal of the sample user U1, but the sample user U1 has not performed any interactive operation on the recommended object C2, then the recommended object C2 is the second recommended object of the sample user U1.

[0051] It should be understood here that, since the historical interaction data of each sample user is different, the first recommended object and the second recommended object of each sample user are also different.

[0052] Step S220 , generating first sample data according to the user characteristics of the sample user and the object characteristics of the corresponding first recommended object; and generating second sample data according to the user characteristics of the sample user and the object characteristics of the corresponding second recommended object.

[0053] When generating sample data, the embodiment of the present invention can pre-extract user features of each sample user, and the user features are specifically related data that can characterize user characteristics, such as user identity information, behavior information, etc. In addition, object features of each recommended object in the global recommended object set are also extracted, and the object features are specifically related data that can characterize the characteristics of the recommended object, such as object category, object duration, etc.

[0054] After the first recommended object and the second recommended object of each sample user are identified in step S210, for each sample user, first sample data is generated based on the user feature of the sample user and the object feature of the first recommended object corresponding to the sample user, that is, the first sample data has the corresponding sample user and the recommended object, and the recommended object is the first recommended object of the sample user. And, for each sample user, second sample data is generated based on the user feature of the sample user and the object feature of the second recommended object corresponding to the sample user, that is, the second sample data has the corresponding sample user and the recommended object, and the recommended object is the second recommended object of the sample user.

[0055] Step S230, inputting the first sample data and the second sample data into a pre-built recommendation model for model training of the recommendation model; wherein the optimization objective function of the recommendation model is composed of the first objective function and the second objective function.

[0056] A recommendation model is constructed in advance based on a machine learning algorithm. The embodiment of the present invention does not limit the specific structure of the recommendation model. For example, the recommendation model can be constructed based on a convolutional neural network (such as a DNN multi-classification model) and the like.

[0057] During the training process of the recommendation model, the first sample data and the second sample data are input into the pre-built recommendation model, and the error between the predicted result and the true label is calculated through the optimization objective function of the recommendation model. Then, the parameters of the recommendation model are updated through the error back propagation algorithm, thereby achieving the purpose of training the recommendation model.

[0058] Different from the prior art method of processing all sample data with the same objective function, the embodiment of the present invention specifically processes the first sample data with the first objective function and processes the second sample data with the second objective function, and generates the overall optimization objective function of the recommendation model based on the first objective function and the second objective function.

[0059] It can be seen that the embodiment of the present invention uses the recommended objects in the global recommended object set to train the recommendation model, thereby avoiding sample bias and improving the prediction accuracy of the recommendation model; and the embodiment of the present invention generates first sample data according to the user characteristics of the sample user and the object characteristics of the corresponding first recommended object, and generates second sample data according to the user characteristics of the sample user and the object characteristics of the corresponding second recommended object, and processes the first sample data using the first objective function, and processes the second sample data using the second objective function, so that different objective functions are used for different sample data to perform optimization processing, thereby further improving the prediction accuracy of the recommendation model.

[0060] As an optional implementation of this embodiment, the method further includes: assigning a positive feedback label or a negative feedback label to the first sample data. Specifically, after the first sample data is generated, the feedback behavior type of the sample user to the first recommended object is determined based on the interaction data between the sample user corresponding to the sample data and the corresponding first recommended object, and the feedback behavior type includes positive feedback and negative feedback, and then the first sample data is assigned a positive feedback label or a negative feedback label based on the feedback type of the sample user to the first recommended object. Among them, if the sample user has positive feedback behavior on the first recommended object, a positive feedback label is assigned to the first sample data; and / or, if the sample user has negative feedback behavior on the first recommended object, a negative feedback label is assigned to the first sample data. For example, if the first sample data D1 corresponds to the sample user U1 and the first recommended object U3, and if the interaction data between the sample user U1 and the first recommended object U3 includes collection behavior data, it is determined that the sample user U1 has positive feedback on the first recommended object U3, and a positive feedback label is assigned to the first sample data D1; if the first sample data D2 corresponds to the sample user U1 and the first recommended object U4, and if the interaction data between the sample user U1 and the first recommended object U4 includes blocking behavior data, it is determined that the sample user U1 has negative feedback on the first recommended object U4, and a negative feedback label is assigned to the first sample data D2.

[0061] Furthermore, the method further includes: assigning a similar user preference label to the second sample data; wherein the similar user preference label includes feedback behavior information of similar users of the sample user; and the feedback behavior information includes positive feedback behavior information or negative feedback behavior information of similar users on the recommended object. Specifically, the sample user corresponding to the second sample data has no feedback behavior on the corresponding second recommended object, and thus the second sample data is not assigned a positive feedback label or a negative feedback label, but the similar users of the sample user corresponding to the second sample data are determined, and then the similar user preference label of the second sample data is generated according to the feedback behavior of the similar users on the recommended object corresponding to the second sample data.

[0062] As an optional implementation of this embodiment, before assigning similar user preference labels to the second sample data, the method further includes: Figure 3 Steps S310 to S330 are shown.

[0063] Step S310: determining the interest points of any sample user and candidate similar users.

[0064] For any sample user, the sample user's points of interest (i.e., POI) are identified based on the sample user's historical data. The point of interest can be a recommended object (for example, the point of interest can be a popular video), or an overall attribute of the recommended object (for example, the point of interest can be a video type), etc. Each sample user corresponds to one or more points of interest.

[0065] And, for any sample user, determine the candidate similar users of the sample user. In an optional manner, to simplify the operation of determining the candidate similar users, the candidate similar users of each sample user are the same, and can be other sample users in the sample user set except the current sample user. In another optional manner, to simplify subsequent operations, users who have the same interest point as the sample user can be used as candidate similar users of the sample user.

[0066] Step S320, obtaining the interest point similarity between the candidate similar user and the sample user for the same interest point.

[0067] For the same point of interest, the similarity of the point of interest between each candidate similar user and the sample user is calculated. The similarity of the point of interest indicates the degree of similarity between the candidate similar user and the sample user in the point of interest. For example, the similarity of the point of interest can be determined based on the difference in interest between the candidate similar user and the sample user for the point of interest, or based on the trend of interest change between the candidate similar user and the sample user for the point of interest, etc.

[0068] Optionally, in order to improve the accuracy of obtaining the similarity of interest points, the embodiment of the present invention may specifically adopt Figure 4 Steps S321 to S323 are shown to determine the similarity of corresponding interest points based on the interest change trend:

[0069] Step S321, obtaining the interest change trend of the sample users for the interest point, and obtaining the interest change trend of the candidate similar users for the interest point.

[0070] The changing trend of the user's interest in the point of interest is obtained in the following manner: the user's interest in the point of interest in each time unit in the preset time window is obtained respectively, and the interest in each time unit is arranged according to the temporal relationship of the time unit, thereby generating the changing trend of the user's interest in the point of interest. For example, based on the snapshot data of each user in the last 10 days, the user's interest in the point of interest every day in the last 10 days can be determined, and then the determined interest is arranged in chronological order to obtain the changing trend of the interest. The specific method of obtaining the interest can be obtained by using an existing interest evaluation algorithm, which is not limited in the embodiment of the present invention.

[0071] Specifically, the interest change trend can be presented in a sequence. As shown in Table 1, Table 1 shows the interest of users user1, user2, and user3 for the interest point Tagk every day in the last 10 days.

[0072] Table 1. Interestingness sequence

[0073] day1 day2 day3 day4 day5 day6 day7 day8 day9 day10 user1 43.96 34.43 82.82 62.46 32.01 33.89 77.35 64.41 30.93 33.38 user2 22.17 41.94 13.23 12.99 24.49 14.09 36.88 37.47 88.45 36.67 user3 11.22 4.36 55.55 22.66 6.05 7.40 40.65 31.75 5.63 6.29

[0074] In addition, the interest change trend can also be expressed as Figure 5 The line chart shown. Figure 5 In , each broken line corresponds to a user, and the horizontal axis of each plotted point is the corresponding unit time, and the vertical axis is the interest degree.

[0075] Step S322, calculating the similarity of the interest point change trends of the candidate similar users and the sample users for the same interest point.

[0076] The embodiment of the present invention does not limit the specific algorithm for calculating the similarity of the change trend of interest points. For example, the calculation can be performed using a Euclidean distance algorithm or a DTW (Dynamic Time Warping) algorithm.

[0077] Preferably, since the sequence lengths of interest point change trends corresponding to different users may be different and the time units may be different, if the Euclidean distance algorithm is used to directly calculate the Euclidean distance of the interest point change trend sequence, there will be a large error. Based on this, the embodiment of the present invention preferably uses the DTW algorithm to calculate the similarity between the interest point change trends corresponding to any two users, thereby improving the scope of application of the method and improving the accuracy of determining the similarity of interest point change trends.

[0078] Specifically, the DTW algorithm calculates the distance matrix between each point in the two sequences and finds a path from the upper left corner to the lower right corner of the matrix so that the sum of the elements on the path is minimized. For example, the negative correlation parameter of the similarity of the change trend of the interest point of useri and userj for the interest point Tagk can be calculated by the following formula 1: in, The smaller it is, the higher the similarity of the change trend of interest points of useri and userj for the interest point Tagk is:

[0079]

[0080] In formula 1, It represents the Euclidean distance between the interest levels of users useri and userj for the point of interest Tagk in T unit time.

[0081] By recursively applying the above formula 1, the shortest path of the interest change trend sequence of useri and userj for the interest point Tagk can be obtained, and then the interest change trend similarity of useri and userj for the interest point Tagk can be obtained.

[0082] Step S323: The similarity of interest point change trends between the candidate similar user and the sample user for the same interest point is used as the interest point similarity between the candidate similar user and the sample user for the interest point.

[0083] The similarity of the interest point change trends between the candidate similar users and the sample users for the same interest point calculated in step S322 can characterize the trend similarity between users for the same interest point, and then the interest point change trend similarity can characterize the similarity of users' behaviors for the same interest point. Therefore, the interest point change trend similarity between the candidate similar users and the sample users for the same interest point is used as the interest point similarity between the candidate similar users and the sample users for the interest point.

[0084] Step S330, determining candidate similar users whose interest point similarity with the sample user for the same interest point is greater than a preset threshold as similar users of the sample user.

[0085] Specifically, if the interest point similarity between a candidate similar user and a sample user for the same interest point is greater than a preset threshold, the candidate similar user is determined as a similar user of the sample user.

[0086] That is, the similar user set of the sample user can be expressed by the following formula 2:

[0087]

[0088] In formula 2, represents the set of similar users of the sample user, The negative correlation parameter representing the similarity of interest points between the sample user and similar users. The smaller the negative correlation parameter, the greater the similarity of interest points. Threshold is the negative correlation parameter of the preset threshold. The smaller the threshold, the greater the preset threshold.

[0089] In addition, in order to improve the training speed of the recommendation model, similar users to each sample user can be determined in advance.

[0090] As an optional implementation of this embodiment, the first objective function may be a maximum likelihood function for binary classification. Optionally, the first objective function may be specifically as shown in Formula 3.

[0091]

[0092] In Formula 3, L1 is the first objective function, p(y|x; θ) is the conditional probability, which is specifically shown in Formula 4, θ is the classification model parameter, σ is the Sigmoid function, L(y) is the label indicator function, L(y)=1 for the positive feedback label, and L(y)=0 for the negative feedback label.

[0093]

[0094] As an optional implementation of this embodiment, the second objective function incorporates the similarity of interest points between the sample user and the corresponding similar user, and the second objective function incorporates the positive feedback behavior information or negative feedback behavior information of the similar user on the recommended object. That is, the second objective function in the embodiment of the present invention can be based on the first objective function, and further utilizes the similarity of interest points between the sample user and the corresponding similar user, and the positive feedback behavior information or negative feedback behavior information of the similar user on the recommended object to modify the first objective function to obtain the second objective function.

[0095] Optionally, the second objective function can be constructed based on the maximum likelihood estimation function, which can be specifically as follows:

[0096] As shown in formula 5:

[0097]

[0098] In Formula 5, L2 is the second objective function, p(y|x; θ) is the conditional probability, θ is the classification model parameter, σ is the Sigmoid function, Neighbor(u) is the set of similar users of sample user u, and L(y i ) is the label indicator function of similar user i. If similar user i gives positive feedback to the recommended object, then L(y i )=1, if similar user i gives negative feedback to the recommended object, then L(y i )=0,r i is the similarity of interest points between sample user u and similar user i, which is specifically shown in Formula 6.

[0099]

[0100] In formula 6, is the negative correlation parameter in Formula 1.

[0101] As an optional implementation of this embodiment, the sum of the first objective function and the second objective function can be used as the optimization objective function of the recommendation model, as shown in Formula 7. In addition, the weighted sum of the first objective function and the second objective function can also be used as the optimization objective function of the recommendation model, and so on.

[0102] L=L1+L2 (Formula 7)

[0103] In Formula 7, L is the optimization objective function of the recommendation model, L1 is the first objective function, and L2 is the second objective function.

[0104] As an optional implementation of this embodiment, Figure 6 FIG. 2 shows a data flow diagram of a recommendation model provided by an embodiment of the present invention. Figure 6 As shown in the figure, the recommendation model adopts a DNN multi-classification model. During the training process, the input data includes: user features (User Feature), object features (Item Feature), context features (ContextFeature), and labels including similar user preference labels (User All Neighbor's Labels). Among them, context features specifically include context and other information that can assist the recommendation model in making decisions. For example, context features can include information such as the device model of the terminal used by the user and the time when the behavior occurred. User features, object features, and context features enter the shared layer (Share Embedding Layer) of the recommendation model, and then pass through the pooling layer (Poling Layer) and three relu layers. Finally, during training, the results of the relu layer are fused and calculated with the labels in SoftMax, and the classification probability (ClassProbabilities) is obtained, and then the model parameters are adjusted through negative feedback to complete the model training.

[0105] As an optional implementation of this embodiment, Figure 6 As shown, after the recommendation model training is completed, in order to improve the recommendation efficiency of user recommendation using the recommendation model, the object vector (video vector v) of each recommended object can be extracted from the recommendation model, such as obtaining the Embedding of the Item collection, and stored in the ANN retrieval system (Index), for example, it can be stored in FAISS, which is convenient for subsequent query and other processing.

[0106] Figure 7 FIG. 1 is a schematic diagram showing a structure of a recommended device provided by an embodiment of the present invention. Figure 7 As shown, the recommended device includes the following modules:

[0107] The first acquisition module 710 is used to acquire user data of the user to be recommended; wherein the user data includes feedback behavior data of the user to be recommended on the historical recommendation object within a preset historical time window;

[0108] An input module 720 is used to input the user data into a pre-trained recommendation model; wherein the recommendation model is trained by first sample data and second sample data of a sample user, the first sample data includes object features of a first recommended object, and the sample user has feedback on the first recommended object; the second sample data includes object features of a second recommended object, and the sample user has no feedback on the second recommended object; and the recommendation model includes a first objective function for processing the first sample data, and a second objective function for processing the second sample data;

[0109] The second acquisition module 730 is used to acquire the output result of the recommendation model, and obtain the target recommendation object matching the to-be-recommended user based on the output result.

[0110] In an optional implementation, the recommendation device further includes: a training module, configured to identify a first recommended object and a second recommended object of any sample user from the global recommended object set;

[0111] Generate first sample data according to the user characteristics of the sample user and the object characteristics of the corresponding first recommended object; and generate second sample data according to the user characteristics of the sample user and the object characteristics of the corresponding second recommended object;

[0112] Inputting the first sample data and the second sample data into a pre-built recommendation model to perform model training on the recommendation model;

[0113] The optimization objective function of the recommendation model is composed of a first objective function and a second objective function.

[0114] In an optional implementation, the training module is further used to: if the sample user has positive feedback behavior on the first recommended object, assign a positive feedback label to the first sample data;

[0115] And / or, if the sample user has negative feedback behavior on the first recommended object, a negative feedback label is assigned to the first sample data.

[0116] In an optional implementation, the training module is further used to: assign similar user preference labels to the second sample data;

[0117] The similar user preference tags include feedback behavior information of similar users of the sample user; the feedback behavior information includes positive feedback behavior information or negative feedback behavior information of the similar users on the recommended objects.

[0118] In an optional implementation, the device further includes: a similar user determination module, configured to determine the interest points of any sample user and candidate similar users;

[0119] Obtaining the similarity of the point of interest of the candidate similar user and the sample user for the same point of interest;

[0120] Candidate similar users whose interest point similarity with the sample user for the same interest point is greater than a preset threshold are determined as similar users of the sample user.

[0121] In an optional implementation, the similar user determination module is further used to: obtain the interest change trend of the sample user for the interest point, and obtain the interest change trend of the candidate similar user for the interest point;

[0122] Calculate the similarity of interest point change trends of the candidate similar users and the sample users for the same interest point;

[0123] The similarity of interest point change trends between the candidate similar user and the sample user for the same interest point is taken as the interest point similarity between the candidate similar user and the sample user for the interest point.

[0124] In an optional implementation, the second objective function incorporates the similarity of interest points between the sample user and the corresponding similar user, and the second objective function incorporates positive feedback behavior information or negative feedback behavior information of the similar user on the recommended object.

[0125] The specific functions and implementation processes of each module of the present device can refer to the steps in the corresponding method embodiments and will not be described in detail here.

[0126] It can be seen that the embodiment of the present invention can recommend highly matched recommendation objects to the recommended users through the trained recommendation model, thereby realizing fast and accurate recall of the recommended objects; and the recommendation model in the embodiment of the present invention is trained using the information of the recommended objects for which the sample users have had feedback and the recommended objects for which they have not had feedback, thereby avoiding sample bias and improving the prediction accuracy of the recommendation model, thereby improving the accuracy of obtaining the recommended objects; and the embodiment of the present invention uses different objective functions to process the sample data generated for the recommended objects for which the sample users have had feedback and the recommended objects for which they have not had feedback, thereby further improving the prediction accuracy of the recommendation model and improving the accuracy of determining the recommended objects.

[0127] An embodiment of the present invention provides a non-volatile computer storage medium, wherein the computer storage medium stores at least one executable instruction, and the computer executable instruction can execute the recommended method in any of the above method embodiments.

[0128] Figure 8 The schematic diagram of the structure of a computing device provided by an embodiment of the present invention is shown. The specific embodiment of the present invention does not limit the specific implementation of the computing device.

[0129] like Figure 8 As shown, the computing device may include: a processor (processor) 802 , a communication interface (Communications Interface) 804 , a memory (memory) 806 , and a communication bus 808 .

[0130] The processor 802, the communication interface 804, and the memory 806 communicate with each other via the communication bus 808. The communication interface 804 is used to communicate with other devices such as a client or other server network elements. The processor 802 is used to execute the program 810, which can specifically execute the relevant steps in the above method embodiment.

[0131] Specifically, the program 810 may include program codes, which include computer operation instructions.

[0132] The processor 802 may be a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention. The one or more processors included in the computing device may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0133] The memory 806 is used to store the program 810. The memory 806 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory. The program 810 can be specifically used to enable the processor 802 to execute the steps in the above method embodiment.

[0134] The algorithm or display provided herein is not inherently related to any particular computer, virtual system or other equipment. Various general purpose systems can also be used together with the teachings based on this. According to the above description, it is obvious to construct the structure required for this type of system. In addition, the embodiment of the present invention is not directed to any specific programming language yet. It should be understood that various programming languages ​​can be utilized to realize the content of the present invention described herein, and the description made to specific languages ​​above is for disclosing the best mode of the present invention.

[0135] In the description provided herein, a large number of specific details are described. However, it is understood that embodiments of the present invention can be practiced without these specific details. In some instances, well-known methods, structures and techniques are not shown in detail so as not to obscure the understanding of this description.

[0136] Similarly, it should be understood that in order to streamline the present invention and aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of the present invention, the various features of the embodiments of the present invention are sometimes grouped together into a single embodiment, figure, or description thereof. However, this disclosed method should not be interpreted as reflecting the following intention: that the claimed invention requires more features than the features explicitly recited in each claim. More specifically, as reflected in the claims below, inventive aspects lie in less than all the features of the individual embodiments disclosed above. Therefore, the claims that follow the specific embodiment are hereby expressly incorporated into the specific embodiment, with each claim itself serving as a separate embodiment of the present invention.

[0137] Those skilled in the art will appreciate that the modules in the devices in the embodiments may be adaptively changed and arranged in one or more devices different from the embodiments. The modules or units or components in the embodiments may be combined into one module or unit or component, and in addition they may be divided into a plurality of submodules or subunits or subcomponents. Except that at least some of such features and / or processes or units are mutually exclusive, all features disclosed in this specification (including the accompanying claims, abstracts and drawings) and all processes or units of any method or device disclosed in this manner may be combined in any combination. Unless otherwise expressly stated, each feature disclosed in this specification (including the accompanying claims, abstracts and drawings) may be replaced by an alternative feature providing the same, equivalent or similar purpose.

[0138] In addition, those skilled in the art will appreciate that, although some embodiments herein include certain features included in other embodiments but not other features, the combination of features of different embodiments is meant to be within the scope of the present invention and form different embodiments. For example, in the claims below, any one of the claimed embodiments may be used in any combination.

[0139] The various component embodiments of the present invention can be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. It should be understood by those skilled in the art that a microprocessor or digital signal processor (DSP) can be used in practice to implement some or all functions of some or all components according to embodiments of the present invention. The present invention can also be implemented as a device or apparatus program (e.g., computer program and computer program product) for executing part or all of the methods described herein. Such a program implementing the present invention can be stored on a computer-readable medium, or can have the form of one or more signals. Such a signal can be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.

[0140] It should be noted that the above embodiments illustrate the present invention rather than limit it, and that those skilled in the art may devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference symbol between brackets shall not be construed as a limitation on the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "one" or "an" preceding an element does not exclude the presence of a plurality of such elements. The present invention may be implemented by means of hardware comprising a number of different elements and by means of a suitably programmed computer. In a unit claim enumerating a number of devices, several of these devices may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order. These words may be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be understood as limitations on the order of execution.

Claims

1. A recommendation method, characterized in that: include: Acquire user data of the user to be recommended; wherein the user data includes feedback behavior data of the user to be recommended on the historical recommendation object within a preset historical time window; Input the user data into a pre-trained recommendation model; wherein the recommendation model is obtained by training first sample data and second sample data of a sample user, the first sample data includes object features of a first recommended object, and the sample user has feedback behavior on the first recommended object; the second sample data includes object features of a second recommended object, and the sample user has no feedback behavior on the second recommended object; and the recommendation model includes a first objective function for processing the first sample data, and a second objective function for processing the second sample data; The output result of the recommendation model is obtained, and based on the output result, a target recommendation object matching the user to be recommended is obtained.

2. The method according to claim 1, characterized in that: The recommendation model is trained in the following way: Identify a first recommended object and a second recommended object for any sample user from a global recommended object set; Generate first sample data according to the user characteristics of the sample user and the object characteristics of the corresponding first recommended object; and generate second sample data according to the user characteristics of the sample user and the object characteristics of the corresponding second recommended object; Inputting the first sample data and the second sample data into a pre-built recommendation model to perform model training on the recommendation model; The optimization objective function of the recommendation model is composed of a first objective function and a second objective function.

3. The method according to claim 2, characterized in that After generating the first sample data, the method further includes: If the sample user has positive feedback behavior on the first recommended object, assigning a positive feedback label to the first sample data; And / or, if the sample user has negative feedback behavior on the first recommended object, a negative feedback label is assigned to the first sample data.

4. The method according to claim 2 or 3, characterized in that: After generating the second sample data, the method further includes: assigning similar user preference labels to the second sample data; The similar user preference tags include feedback behavior information of similar users of the sample user; the feedback behavior information includes positive feedback behavior information or negative feedback behavior information of the similar users on the recommended objects.

5. The method according to claim 4, characterized in that Before assigning the similar user preference label to the second sample data, the method further includes: Determine the interest points of any sample user and candidate similar users; Obtaining the similarity of the point of interest of the candidate similar user and the sample user for the same point of interest; Candidate similar users whose interest point similarity with the sample user for the same interest point is greater than a preset threshold are determined as similar users of the sample user.

6. The method according to claim 5, characterized in that The obtaining of the similarity of the interest points of the candidate similar users and the sample users for the same interest point further comprises: Obtaining a change trend of the interest of the sample user for the point of interest, and obtaining a change trend of the interest of the candidate similar user for the point of interest; Calculate the similarity of interest point change trends of the candidate similar users and the sample users for the same interest point; The similarity of interest point change trends between the candidate similar user and the sample user for the same interest point is used as the interest point similarity between the candidate similar user and the sample user for the interest point.

7. The method according to claim 5 or 6, characterized in that: The second objective function incorporates the similarity of interest points between the sample user and the corresponding similar user, and the second objective function incorporates positive feedback behavior information or negative feedback behavior information of the similar user on the recommended object.

8. A recommendation device, characterized in that: include: The first acquisition module is used to acquire user data of the user to be recommended; wherein the user data includes feedback behavior data of the user to be recommended on the historical recommendation object within a preset historical time window; An input module is used to input the user data into a pre-trained recommendation model; wherein the recommendation model is obtained by training first sample data and second sample data of a sample user, the first sample data includes object features of a first recommended object, and the sample user has feedback behavior on the first recommended object; the second sample data includes object features of a second recommended object, and the sample user has no feedback behavior on the second recommended object; and the recommendation model includes a first objective function for processing the first sample data, and a second objective function for processing the second sample data; The second acquisition module is used to acquire the output result of the recommendation model, and obtain the target recommendation object matching the to-be-recommended user based on the output result.

9. A computing device, characterized in that include: A processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform an operation corresponding to the recommendation method according to any one of claims 1 to 7.

10. A computer storage medium, characterized in that: The storage medium stores at least one executable instruction, and the executable instruction enables the processor to execute an operation corresponding to the recommendation method according to any one of claims 1 to 7.

Images