Recommended model training methods, media, devices and computing equipment

Abstract

The embodiments of the present disclosure provide a recommendation model training method, medium, apparatus and computing device. The method includes: generating training samples based on the push records of pushed content and the user's interactive operation records on the pushed content, the training samples including the popularity corresponding to the pushed content and the label value of whether the user has an interactive operation on each pushed content; inputting the training samples into the recommendation model for training, outputting user feature parameters and content feature parameters, and the recommendation model is a matrix decomposition model. By incorporating popularity into the training goal, the popular and non-popular factors in the obtained user feature parameters and content feature parameters are decoupled, reducing the dominant role of popular content in the recommendation model in determining the content to be pushed, thereby reducing the Matthew effect of the pushed content being biased towards popularity, and improving the coverage of user interest preference areas other than popularity, bringing a better experience to users.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of computer technology. More specifically, the embodiments of the present disclosure relate to a recommendation model training method, medium, apparatus, and computing device. Background Art

[0002] This section is intended to provide a background or context to the embodiments of the present disclosure that are recited in the claims. No statement herein is admitted to be prior art by virtue of its inclusion in this section.

[0003] With the development of Internet technology, content recommendation functions are used in more and more applications. By pushing corresponding content based on the characteristics of user traffic, the user experience can be improved.

[0004] Current recommendation systems tend to recommend popular content to users. This popular content, due to its greater exposure, becomes even more popular. Meanwhile, unrecommended content, due to its lack of exposure, becomes less popular. This is the Matthew effect in recommendation systems.

[0005] Therefore, due to the existence of the Matthew effect, the recommended content is too biased towards popular content, resulting in insufficient user satisfaction with the pushed content. Summary of the Invention

[0006] The present disclosure provides a recommendation model training method, medium, apparatus, and computing device to solve the problem that content recommendation systems in the prior art tend to recommend content with high popularity, resulting in insufficient user satisfaction with the pushed content.

[0007] In a first aspect of the embodiments of the present disclosure, a recommendation model training method is provided, comprising:

[0008] Generate training samples based on push records of pushed content and user interaction records on pushed content. The training samples include the popularity of the pushed content and the label value indicating whether the user has performed an interaction operation on each pushed content.

[0009] The training samples are input into the recommendation model for training, and user feature parameters and content feature parameters are output. The recommendation model is a matrix decomposition model.

[0010] In one embodiment of the present disclosure, the recommendation model training method includes: inputting training samples into the recommendation model for training, and outputting user feature parameters and content feature parameters, including:

[0011] Based on the user feature parameters and content feature parameters corresponding to the recommendation model, the predicted interactive operation probability value and predicted popularity corresponding to the training sample are determined;

[0012] Based on the difference between the label value in the training sample and the predicted interactive operation probability value, and the difference between the popularity and the predicted popularity, the user feature parameters and content feature parameters are adjusted until the training end conditions are met.

[0013] In an exemplary embodiment of the present disclosure, both user feature parameters and content feature parameters include popularity factors and preference factors; based on the user feature parameters and content feature parameters corresponding to the recommendation model, the predicted interaction operation probability value and predicted popularity corresponding to the training sample are determined, including: determining the predicted interaction operation probability value according to the sum of the inner product of the popularity factors of the user feature parameters and content feature parameters and the inner product of the preference factors of the user feature parameters and content feature parameters; determining the predicted popularity according to the difference between the inner product of the popularity factors of the user feature parameters and content feature parameters and the inner product of the preference factors of the user feature parameters and content feature parameters.

[0014] In an exemplary embodiment of the present disclosure, the predicted popularity is determined based on the difference between the inner product of the popularity factors of the user feature parameters and the content feature parameters and the inner product of the preference factors of the user feature parameters and the content feature parameters, including: when the label value in the training sample indicates that the user has performed at least one interactive operation on the pushed content, the predicted popularity is determined based on the result of subtracting the inner product of the preference factors of the user feature parameters and the content feature parameters from the inner product of the popularity factors of the user feature parameters and the content feature parameters; when the label value in the training sample indicates that the user has not performed any interactive operation on the pushed content, the predicted popularity is determined based on the result of subtracting the inner product of the popularity factors of the user feature parameters and the content feature parameters from the inner product of the preference factors of the user feature parameters and the content feature parameters.

[0015] In an exemplary embodiment of the present disclosure, the model parameters of the recommendation model also include a preset bias coefficient; based on the user feature parameters and content feature parameters corresponding to the recommendation model, the predicted interaction operation probability value and predicted popularity corresponding to the training sample are determined, including: based on the user feature parameters, content feature parameters and bias coefficient corresponding to the recommendation model, the adjusted predicted interaction operation probability value and predicted popularity are determined; accordingly, based on the difference between the label value and the predicted interaction operation probability value, and the difference between the popularity and the predicted popularity in the training sample, the user feature parameters and content feature parameters are adjusted, including: based on the difference between the adjusted predicted interaction operation probability value and the label value, and the difference between the adjusted predicted popularity and the popularity, the user feature parameters and content feature parameters are adjusted until the training end conditions are met.

[0016] In another embodiment of the present disclosure, the recommendation model training method includes: generating training samples based on push records of pushed content and user interaction operation records on the pushed content, including:

[0017] Based on the number of pushes of each pushed content in the push record or the number of interactive operations in the interactive operation record, the popularity of the pushed content at both ends of the sorting and the preset sorting position is determined as the set popularity;

[0018] Based on the push content and the set popularity as a benchmark, the popularity of each push content in the training sample is determined.

[0019] In an exemplary embodiment of the present disclosure, the popularity of each pushed content in the training sample is determined based on the pushed content and the set popularity, including: normalizing the number of pushes or the number of interactive operations of each pushed content in the sorting based on the pushed content and the corresponding set popularity, and using the processed result as the popularity of the pushed content.

[0020] In an exemplary embodiment of the present disclosure, after inputting training samples into a recommendation model for training and outputting user feature parameters and content feature parameters, it also includes: obtaining a predicted interactive operation probability value of each user for each pushed content based on the user feature parameters and content feature parameters; and determining the push content to be pushed to the target user based on the ranking of the predicted interactive operation probability values.

[0021] In a second aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, comprising:

[0022] The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the recommendation model training method in the first aspect of the present disclosure.

[0023] In a third aspect of the embodiments of the present disclosure, a recommendation model training device is provided, comprising:

[0024] A generation module is used to generate training samples based on push records of pushed content and user interaction operation records on the pushed content. The training samples include the popularity of the pushed content and the label value of whether the user has performed an interaction operation on each pushed content;

[0025] The training module is used to input training samples into the recommendation model for training and output user feature parameters and content feature parameters. The recommendation model is a matrix decomposition model.

[0026] In an exemplary embodiment of the present disclosure, the training module is specifically used to: determine the predicted interaction operation probability value and predicted popularity corresponding to the training sample based on the user feature parameters and content feature parameters corresponding to the recommendation model; and adjust the user feature parameters and content feature parameters based on the difference between the label value in the training sample and the predicted interaction operation probability value, and the difference between the popularity and the predicted popularity, until the training end conditions are met.

[0027] In an exemplary embodiment of the present disclosure, the training module is specifically used to: when both the user feature parameters and the content feature parameters include popularity factors and preference factors, determine the predicted interaction operation probability value and predicted popularity corresponding to the training sample based on the user feature parameters and content feature parameters corresponding to the recommendation model, including: determining the predicted interaction operation probability value according to the sum of the inner product of the popularity factors of the user feature parameters and the content feature parameters and the inner product of the preference factors of the user feature parameters and the content feature parameters; determining the predicted popularity according to the difference between the inner product of the popularity factors of the user feature parameters and the content feature parameters and the inner product of the preference factors of the user feature parameters and the content feature parameters.

[0028] In an exemplary embodiment of the present disclosure, the training module is specifically used to: when the label value in the training sample indicates that the user has performed at least one interactive operation on the pushed content, determine the predicted popularity based on the result of subtracting the inner product of the preference factor of the user feature parameter and the content feature parameter from the inner product of the popularity factor of the user feature parameter and the content feature parameter; when the label value in the training sample indicates that the user has not performed any interactive operation on the pushed content, determine the predicted popularity based on the result of subtracting the inner product of the popularity factor of the user feature parameter and the content feature parameter from the inner product of the preference factor of the user feature parameter and the content feature parameter.

[0029] In an exemplary embodiment of the present disclosure, the training module is specifically used to: when the model parameters of the recommendation model also include a preset bias coefficient, determine the adjusted predicted interaction operation probability value and predicted popularity based on the user feature parameters, content feature parameters and bias coefficient corresponding to the recommendation model; and, based on the difference between the adjusted predicted interaction operation probability value and the label value, and the difference between the adjusted predicted popularity and the popularity, adjust the user feature parameters and content feature parameters until the training end conditions are met.

[0030] In an exemplary embodiment of the present disclosure, the generation module is specifically used to: determine the popularity of the push content at both ends of the sorting and the preset sorting position as the set popularity based on the number of pushes of each push content in the push record or the number of interaction operations in the interaction operation record; determine the popularity of each push content in the training sample based on the push content and the set popularity as a benchmark.

[0031] In an exemplary embodiment of the present disclosure, the generation module is specifically used to: normalize the number of pushes or interactive operations of each pushed content in the sorting based on the pushed content and the corresponding set popularity, and use the processed result as the popularity of the pushed content.

[0032] In an exemplary embodiment of the present disclosure, the training module is also used to: after inputting the training samples into the recommendation model for training and outputting the user feature parameters and content feature parameters, obtain the predicted interactive operation probability value of each user for each pushed content based on the user feature parameters and content feature parameters; and determine the push content to be pushed to the target user based on the ranking of the predicted interactive operation probability values.

[0033] In a fourth aspect of the embodiments of the present disclosure, there is provided a computing device comprising: at least one processor;

[0034] and a memory communicatively coupled to the at least one processor;

[0035] The memory stores instructions that can be executed by at least one processor, and the instructions are executed by at least one processor to enable the computing device to perform the recommendation model training method as in the first aspect of the present disclosure.

[0036] According to the recommendation model training method, medium, apparatus, and computing device of the embodiment of the present disclosure, a training sample is generated based on the push record of the pushed content and the user's interactive operation record on the pushed content, and the training sample is input into the recommendation model for training, and user feature parameters and content feature parameters are output. In this way, the recommendation model can be trained based on the two goals of popularity and label value contained in the training sample. By incorporating popularity into the training goal, the popular and non-popular factors in the obtained user feature parameters and content feature parameters are decoupled, reducing the dominant role of popular content in the recommendation model in determining the content to be pushed, thereby reducing the Matthew effect of pushed content biased towards popularity, improving the coverage of user interest preference areas other than popularity, and bringing a better experience to users. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The above and other objects, features and advantages of the exemplary embodiments of the present disclosure will become readily understood by reading the detailed description below with reference to the accompanying drawings, in which several embodiments of the present disclosure are shown by way of example and not limitation, wherein:

[0038] Figure 1 Schematically shows a schematic diagram of an application scenario according to an embodiment of the present disclosure;

[0039] Figure 2 The following schematically shows a flow chart of a recommendation model training method according to an embodiment of the present disclosure;

[0040] Figure 3 The following schematically shows a flow chart of a recommendation model training method according to another embodiment of the present disclosure;

[0041] Figure 4The following schematically shows a flow chart of a recommendation model training method according to another embodiment of the present disclosure;

[0042] Figure 5 Schematically shows a structural diagram of a computer-readable storage medium according to yet another embodiment of the present disclosure;

[0043] Figure 6 The following schematically shows a structural diagram of a recommendation model training device according to another embodiment of the present disclosure;

[0044] Figure 7 The figure schematically shows a structural diagram of a computing device according to another embodiment of the present disclosure.

[0045] In the drawings, the same or corresponding reference numerals denote the same or corresponding parts. DETAILED DESCRIPTION

[0046] The principles and spirit of the present disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present disclosure, and are not intended to limit the scope of the present disclosure in any way. Rather, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0047] Those skilled in the art will appreciate that the embodiments of the present disclosure may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present disclosure may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software.

[0048] According to an embodiment of the present disclosure, a recommendation model training method, medium, apparatus and computing device are proposed.

[0049] In this document, it should be understood that the terms involved are only used to facilitate understanding and do not represent any limitation on the meaning. In addition, the number of any element in the drawings is for illustration and not limitation, and any naming is only for distinction and does not have any limiting meaning.

[0050] The following is a description of the terms used in this disclosure:

[0051] Popularity bias: During the operation of the recommendation system, unpopular content has fewer chances to be recommended or displayed, and the recommendation system prefers to recommend popular content.

[0052] Decoupling: By designing modeling methods, the influence of different factors in the data can be separated from each other, reducing the coupling between them, which can improve the accuracy of modeling.

[0053] Recall: The process by which a recommendation system initially selects items to recommend to users, equivalent to a rough ranking. Existing recommendation systems are broadly divided into two stages: recall and ranking. Popular recall methods in the industry are primarily based on matrix factorization models, but also include content-based recall and nearest neighbor models.

[0054] The principles and spirit of the present disclosure are explained in detail below with reference to several representative embodiments of the present disclosure. SUMMARY OF THE INVENTION

[0056] The inventors have discovered that in current technology, when applications with content recommendation functions recommend content to users, the recommendation system usually first recalls some content to be pushed that the user may be interested in, then sorts it based on certain rules, and determines the final content recommended to the user based on the sorting results.

[0057] To best meet the personalized interests of different users, recommendation systems must ensure both high-quality content and tailor it to the personalized interests of different users when recalling content to be pushed. However, existing recommendation systems typically favor more popular content over less popular content. When the popularity of a consumer's personalized interests is low, the existing content to be pushed often fails to fully satisfy the user, impacting their long-term user experience. Furthermore, content producers face challenges with the reasonable distribution of their low-popularity, high-quality content, and thus, insufficient revenue. For platforms, the lack of fair content distribution and the inability to fully satisfy users' personalized interests directly leads to the loss of both content producer and consumer users, ultimately degrading the platform ecosystem.

[0058] Therefore, the content recommended by the recommendation system is too biased towards popular content, which will lead to insufficient user satisfaction with the pushed content.

[0059] In this solution, the recommendation model is trained based on the two objectives of popularity and label value contained in the training samples, so that the popular and non-popular factors in the obtained user feature parameters and content feature parameters are decoupled, reducing the dominant role of popular content in the recommendation model in determining the content to be pushed, thereby reducing the Matthew effect of pushing content towards popular content, and improving the coverage of user interest preference areas other than popular content, bringing a better experience to users and improving user satisfaction.

[0060] After introducing the basic principles of the present disclosure, various non-limiting embodiments of the present disclosure are described in detail below.

[0061] Application Scenario Overview

[0062] First reference Figure 1As shown, in content recommendation, the recommendation system 101 of the server side 100 recalls a set number of high-quality content to be pushed that meets the user's personalized interests from the database 102 according to the preferences of the user 110, and after sorting and screening the content to be pushed, determines the final push content recommended to the user 110, and pushes it to the client 120, completing the content recommendation process.

[0063] It should be noted that Figure 1 In the scenario shown, only one server, user, content to be pushed, pushed content, and client are used as an example for illustration, but the present disclosure is not limited to this. That is, the number of servers, users, content to be pushed, pushed content, and clients can be arbitrary.

[0064] Exemplary Methods

[0065] The following combination Figure 1 For application scenarios, refer to Figures 2 to 5 To describe a method for training a recommendation model according to an exemplary embodiment of the present disclosure. It should be noted that the above application scenarios are only shown to facilitate understanding of the spirit and principles of the present disclosure, and the embodiments of the present disclosure are not limited in this respect. On the contrary, the embodiments of the present disclosure can be applied to any applicable scenario.

[0066] Figure 2 This is a flowchart of a recommendation model training method provided by one embodiment of the present disclosure. Figure 2 As shown, the recommendation model training method provided in this embodiment includes the following steps:

[0067] Step S201: Generate training samples based on push records of pushed content and user interaction operation records on the pushed content.

[0068] The training samples include the popularity of the pushed content and the label value of whether the user has an interactive operation on each pushed content.

[0069] Specifically, push content refers to content pushed to the user by the recommendation system, such as recommended music, videos, playlists, reviews, events, news, etc. Push content can be multiple, such as recommending multiple songs at once, or just one, such as recommending a short video that switches to playing on the user interface.

[0070] Push content can be generated based on the user's active selection operation, such as the recommended songs returned by the server after the user selects the "Daily Recommended Songs" function; it can also be generated based on configuration, such as when the user opens the homepage of the video application, the recommendation column corresponding to the homepage will automatically generate recommended videos for the corresponding user.

[0071] Furthermore, when the server pushes content to the user, it also generates a corresponding push record. This record includes the time of push, the content, and the corresponding user. Therefore, the server can determine the number of times a particular content has been pushed based on the push records. For example, if a comment has 1,000 push records, the server can determine that the comment has been pushed 1,000 times.

[0072] At the same time, after receiving the pushed content, the user can perform interactive operations on the pushed content, such as clicking (including opening, playing, displaying the full text, clicking to view the large image, etc.), commenting, collecting (including liking, etc.), forwarding, etc. After the server receives the user's interactive operations on the pushed content, it will generate corresponding interactive operation records.

[0073] Similar to push records, interaction records typically include the time, user, and corresponding push content of the interaction. The server can also determine the number of interactions for a piece of push content based on the interaction records. For example, if a video is clicked 50 times and forwarded 10 times, its number of interactions is 60.

[0074] In this solution, the recalled content to be pushed is determined by the recommendation model. To ensure the accuracy of the recommendation model and the fairness of the pushed content, it is necessary to provide training samples and use the principles of deep learning in artificial intelligence to train the recommendation model.

[0075] The training samples are determined based on the aforementioned push records and interactive operation records.

[0076] Furthermore, in the training samples, the server can only record whether the user has interacted with a piece of pushed content, rather than specifically recording the number of interactions. That is, when a user has an interaction record corresponding to a piece of pushed content, regardless of the number of interaction records, such as 3 or 20 (for example, a user repeatedly played a song 20 times), the server only records that the user has interacted with the pushed content. Conversely, when there is no interaction record corresponding to the user's interaction with the pushed content, the server records that the user has not interacted with the pushed content.

[0077] Among them, for the record of no interactive operation between the user and the pushed content, the server will record that no interactive operation exists between the user and the pushed content only when the pushed content is pushed to the user and the user does not interact with the pushed content.

[0078] The tag value is a parameter value determined by the server based on whether the user has interacted with the pushed content. Depending on the value of the tag value, it can be used to indicate that the user has interacted with the pushed content (at least once) or that the user has not interacted with the pushed content.

[0079] When the tag value indicates that the user has not performed any interactive operation with the pushed content, it may mean that the pushed content is not attractive enough to the user, or the user is not interested in the pushed content.

[0080] Popularity is used to quantify the popularity of a piece of pushed content. This indicator is determined based on interaction operation records or push records, because pushed content with a large number of pushes and content with a large number of user interaction operations can be considered as highly popular content.

[0081] Each training sample can be represented as a user, a pushed content, the popularity of the content, and the label value corresponding to the user and the pushed content.

[0082] Step S202: Input the training samples into the recommendation model for training, and output user feature parameters and content feature parameters.

[0083] Among them, the recommended model is a matrix decomposition model.

[0084] Specifically, the recommendation system in this solution uses a matrix decomposition model. By representing user preferences for content as a matrix, it can intuitively reflect each user's preference for each piece of content. In the matrix decomposition model, by representing parameters related to user characteristics and content characteristics as latent vectors, the corresponding preference value can be calculated by taking the inner product of the two latent vectors. Therefore, as long as the two latent vectors of user characteristic parameters and content characteristic parameters can be determined, each user's preference value for each piece of content can be determined, and then the content to be recommended to the user can be determined.

[0085] To determine user and content feature parameters, deep learning is required using training samples. Because training samples contain both label values reflecting the user's preference for the pushed content and the popularity of the pushed content itself, the user and content feature parameters output through training can be decoupled from the influence of content popularity. For example, user feature parameters can reflect the user's different preferences for popular and unpopular content, while content feature parameters can reflect the popularity of content. This can avoid over-recommending popular content to users, resulting in a failure to meet user preferences for unpopular content.

[0086] By outputting user feature parameters and content feature parameters after deep learning of training samples, it is possible to be more fair when determining the specific content to be pushed and meet the preferences of users with different popularity levels.

[0087] According to the recommendation model training method of the embodiment of the present disclosure, a training sample is generated based on the push record of the pushed content and the user's interactive operation record on the pushed content, and the training sample is input into the recommendation model for training, outputting user feature parameters and content feature parameters. In this way, the recommendation model can be trained based on the two goals of popularity and label value contained in the training sample. By incorporating popularity into the training goal, the popular and non-popular factors in the obtained user feature parameters and content feature parameters are decoupled, reducing the dominant role of popular content in the recommendation model in determining the content to be pushed, thereby reducing the Matthew effect of pushed content biased towards popularity, improving the coverage of user interest preference areas other than popularity, and bringing a better experience to users.

[0088] Figure 3 This is a flowchart of a recommendation model training method provided by one embodiment of the present disclosure. Figure 3 As shown, the recommendation model training method provided in this embodiment includes the following steps:

[0089] Step S301: Generate training samples based on push records of pushed content and user interaction operation records on the pushed content.

[0090] The training samples include the popularity of the pushed content and the label value of whether the user has an interactive operation on each pushed content.

[0091] This step and Figure 2 The content of step S201 in the illustrated embodiment is the same and will not be repeated here.

[0092] Step S302: Determine a predicted interactive operation probability value based on the sum of the inner product of the popularity factor of the user characteristic parameter and the content characteristic parameter and the inner product of the preference factor of the user characteristic parameter and the content characteristic parameter.

[0093] Among them, both user characteristic parameters and content characteristic parameters include popularity factors and preference factors.

[0094] In deep learning based on training samples, the user feature parameters and content feature parameters corresponding to the recommendation model are determined through the training samples. It is necessary to first determine the predicted interaction operation probability value and predicted popularity corresponding to the training sample based on the user feature parameters and content feature parameters corresponding to the recommendation model, and then continuously adjust the user feature parameters and content feature parameters according to the difference between the predicted interaction operation probability value and the label value in the training sample, and the difference between the predicted popularity and the popularity in the training sample, and finally determine the user feature parameters and content feature parameters.

[0095] In some embodiments, the user feature parameters and content feature parameters trained through training samples can be in a state where the recommendation system has been trained through an existing scheme. In this case, the user feature parameters and content feature parameters are in a state that is affected by popularity and can be used to determine the content to be pushed, that is, training is performed when the recommendation model needs to be updated; or they can be in an original state that has not been trained. In this case, the user feature parameters and content feature parameters only include all users and all content, but the values of each specific parameter are all 0 or any initial value, that is, the recommendation model is trained when the recommendation system is initialized.

[0096] Specifically, in order to respectively reflect the user's preference for content with high popularity and content with low popularity in the user characteristic parameters, the user characteristic parameters need to be split into a popularity factor and a preference factor (for content with low popularity). Similarly, in order to respectively reflect the situation in which the user prefers content in the content characteristic parameters when the content is highly popular and when the content is less popular, the content characteristic parameters are also split into a popularity factor and a preference factor.

[0097] Among them, the sum of the inner product of the popularity factors of the user characteristic parameters and the content characteristic parameters and the inner product of the preference factors of the user characteristic parameters and the content characteristic parameters is consistent with the result of directly calculating the inner product of the user characteristic parameters and the content characteristic parameters.

[0098] Step S303: Determine the predicted popularity based on the difference between the inner product of the popularity factor of the user characteristic parameter and the content characteristic parameter and the inner product of the preference factor of the user characteristic parameter and the content characteristic parameter.

[0099] Specifically, the value obtained by the inner product of the popularity factors of the user characteristic parameters and the content characteristic parameters can be used to represent the degree of preference of a user when a certain pushed content has high popularity; the inner product of the preference factors of the user characteristic parameters and the content characteristic parameters can be used to represent the degree of preference of the user when the push content has low popularity; therefore, when the two are subtracted, the influence of the user on the preference of the content can be eliminated, and then used to represent popularity.

[0100] Since different content has different popularity, users have different preferences for the same content. When a piece of content has low popularity or a user has a high preference for it, the inner product of the popularity factor of the user feature parameters and the content feature parameters may be smaller than the inner product of the preference factor of the user feature parameters and the content feature parameters. To ensure the accuracy of the adjustment of the user feature parameters and the content feature parameters during deep learning, the specific calculation process also includes two different cases:

[0101] Case 1: When the label value in the training sample indicates that the user has performed at least one interactive operation on the pushed content, the predicted popularity is determined based on the result of subtracting the inner product of the preference factor of the user feature parameters and the content feature parameters from the inner product of the popularity factor of the user feature parameters and the content feature parameters.

[0102] The label value indicating that the user has performed at least one interactive operation on the pushed content can take any value. To facilitate parameter training, it can usually take a positive integer value, such as 1, 2, 10, 100, etc.

[0103] At this time, the user has a high degree of preference for the content, that is, the inner product of the preference factors of the user characteristic parameters and the content characteristic parameters plays a dominant role; at this time, if the popularity of the pushed content itself is high, the inner product of the popularity factors of the user characteristic parameters and the content characteristic parameters should be made larger. Therefore, the value of the inner product of the popularity factors of the user characteristic parameters and the content characteristic parameters minus the inner product of the preference factors of the user characteristic parameters and the content characteristic parameters will not be too small, that is, the predicted popularity obtained can also be larger.

[0104] Case 2: When the label value in the training sample indicates that the user has not performed any interactive operation on the pushed content, the predicted popularity is determined based on the result of subtracting the inner product of the popularity factor of the user feature parameters and the content feature parameters from the inner product of the preference factor of the user feature parameters and the content feature parameters.

[0105] Among them, the label value indicating that the user has not performed any interactive operation on the pushed content can take any value different from the label value corresponding to at least one interactive operation. To facilitate parameter training, it can usually take 0 or 1 (when the label value of at least one interactive operation is not 1).

[0106] At this time, the user's preference for the content is relatively low, so the inner product of the popularity factors of the user characteristic parameters and the content characteristic parameters should play a dominant role. At this time, if the popularity of the pushed content itself is low, the inner product of the preference factors of the user characteristic parameters and the content characteristic parameters minus the inner product of the popularity factors of the user characteristic parameters and the content characteristic parameters can be guaranteed not to be too large, that is, the predicted popularity obtained can also be smaller.

[0107] For example, both the user feature parameter and the content feature parameter are composed of two vectors. The first part is the popularity factor, where the popularity factor of the user feature parameter is represented by user_vec con , the popularity factor of the content feature parameter is expressed as item_vec con , the latter part is the preference factor, where the preference factor of the user feature parameter is expressed as user_vec int , the preference factor of the content feature parameter is expressed as item_vec int, then the inner product s of the popularity factor of the user characteristic parameter and the content characteristic parameter con It can be expressed as:

[0108] s con = <user_vec con ,item_vec con >

[0109] The inner product s of the preference factor of the user feature parameter and the content feature parameter int It can be expressed as:

[0110] s int = <user_vec int ,item_vec int >

[0111] Then the sum of the inner product of the popularity factor of the user feature parameters and the content feature parameters and the inner product of the preference factor of the user feature parameters and the content feature parameters is the predicted interactive operation probability value s click (Click operation is used to represent interactive operation) can be expressed as:

[0112] s click =s con +s int ;

[0113] The difference between the inner product of the popularity factor of the user feature parameters and the content feature parameters and the inner product of the preference factor of the user feature parameters and the content feature parameters is the predicted popularity s popularity It can be expressed as:

[0114] s popularity =s con -s int When the label value label indicates that the user has performed at least one interactive operation on the pushed content; or

[0115] s popularity =s int -s con When the label value label indicates that the user has no interactive operation on the pushed content.

[0116] Step S304: Adjust the user feature parameters and content feature parameters based on the difference between the label value in the training sample and the predicted interactive operation probability value, and the difference between the popularity and the predicted popularity, until the training end condition is met.

[0117] Among them, the training end condition can be the difference between the label value and the predicted interactive operation probability value, the difference between the popularity and the predicted popularity is zero, both are less than 5% or less than any set threshold, or the number of adjustments reaches a set number, such as 5 times, 10 times or any number, or other set conditions, such as training time condition (such as the total training time is less than or equal to the set value), adjustment amplitude condition (such as the amplitude value of the most recent adjustment is less than or equal to the set value), parameter convergence condition (such as user feature parameters and content feature parameters converge during the adjustment process).

[0118] Specifically, the method for adjusting user and content feature parameters can use the alternating least squares method (ALS). The specific process is to first fix one parameter (denoted as the first parameter), then optimize the other parameter (denoted as the second parameter) once, then fix the second parameter and optimize the first parameter once, and repeat this alternating process until the first and second parameters converge. The first parameter can be any one of the user and content feature parameters, and the second parameter is another parameter that has not been selected.

[0119] The specific optimization method is the same as the alternating least squares optimization process in the existing matrix decomposition model and will not be repeated here.

[0120] Step S305 : Determine the adjusted predicted interactive operation probability value and predicted popularity based on the user characteristic parameters, content characteristic parameters, and bias coefficient corresponding to the recommendation model.

[0121] Among them, the model parameters of the recommendation model also include the preset bias coefficient

[0122] In some embodiments, based on the needs of the application service provider, a bias coefficient will be added to the recommendation model to make the recommendation results biased towards the direction expected by the application service provider. For example, if the service provider of a music application hopes to push electronic music activities, the recommendation model needs to be updated to make users prefer electronic music more (the bias coefficient for adjusting user feature parameters) or the probability of electronic music being preferred by users is higher (the bias coefficient for adjusting content feature parameters). Therefore, it is necessary to adjust the predicted interaction operation probability value and the predicted popularity based on the preset bias coefficient to retrain the user feature parameters and content feature parameters.

[0123] This step is an optional step parallel to step S304, and the corresponding step can be selected for execution according to actual needs.

[0124] Step S306: Based on the difference between the adjusted predicted interactive operation probability value and the label value, and the difference between the adjusted predicted popularity and the popularity, adjust the user feature parameters and the content feature parameters until the training end condition is met.

[0125] This step is based on step S305 and is an optional step parallel to step S304. The corresponding step can be selected for execution according to actual needs.

[0126] Since bias coefficients are added to the user feature parameters and / or content feature parameters used for training, the inner product of the trained user feature parameters and content feature parameters will be more inclined towards the results expected by the application service provider, compared to the results without the addition of bias coefficients, such as increasing the predicted probability of users interacting with a certain type of recommended content.

[0127] The specific training process and step S304 are not described again here.

[0128] Step S307: Output the trained user feature parameters and content feature parameters.

[0129] Specifically, the user feature parameters and content feature parameters after training can be used by the recommendation system to determine the content to be pushed for recall.

[0130] According to the recommendation model training method of the embodiment of the present disclosure, a training sample is generated based on the push records of the pushed content and the user's interactive operation records on the pushed content, and then the user feature parameters and content feature parameters are respectively split into popularity factors and preference factors, and the predicted interactive operation probability value and predicted popularity corresponding to the training sample are calculated accordingly, and compared with the label value and popularity in the training sample. Based on the difference in the comparison, the user feature parameters and content feature parameters are adjusted and optimized, and after the training end conditions are met, the user feature parameters and content feature parameters are output. As a result, the user feature parameters and content feature parameters used to recall the content to be pushed can clearly include factors related to high popularity and factors related to low popularity, thereby decoupling the popular and non-popular factors in the obtained user feature parameters and content feature parameters, and can effectively avoid the dominant role of the popularity factor when determining the content to be pushed, thereby reducing the Matthew effect of the pushed content being biased towards popularity, improving the coverage of user interest preference areas other than popularity, and bringing a better experience to users.

[0131] Figure 4 This is a flowchart of a recommendation model training method provided by one embodiment of the present disclosure. Figure 4 As shown, the recommendation model training method provided in this embodiment includes the following steps:

[0132] Step S401: Based on the order of push times of each push content in the push record, the popularity of the push content at both ends of the order and the preset order position is determined as a set popularity.

[0133] Specifically, the server determines the training samples based on the push records and interactive operation records related to the user and content in the database corresponding to the application.

[0134] Since popularity is a parameter value based on the content, it is necessary to combine push records based on the same push content and then calculate the number of pushes.

[0135] The specific method for calculating the number of push times can be determined based on the total number of push records for the same push content (each push record corresponds to one push of the push content from the server to the user side).

[0136] The sorting may be in descending order or ascending order. In the subsequent description of this embodiment, the descending order will be described.

[0137] The preset ranking position is used to represent the predetermined ordinal number after sorting the number of pushes of all pushed content. For example, if there are 100 pushed content, the preset ranking position is the 10% position, which means the position of the pushed content ranked 90th is predetermined.

[0138] Furthermore, the set popularity is used to represent a preset popularity value, such as a maximum value, a minimum value, and a set middle value (such as a median value of popularity).

[0139] For example, for all 100 pieces of content, the preset ranking position is 80%, then the preset popularity of the pushed content ranked 1st is the maximum value (such as 10), the preset popularity of the pushed content ranked 100th is the minimum value (such as 1), and the preset popularity of the pushed content ranked 20th (that is, the preset ranking position) is the set middle value (such as 5).

[0140] Step S402: Based on the order of the number of interactive operations for each pushed content in the interactive operation record, the popularity of the pushed content at both ends of the order and at a preset order position is determined as a set popularity.

[0141] The calculation of the number of interactive operations is similar to the calculation of the number of pushes, and is based on the number of all interactive operations corresponding to the same push content.

[0142] The specific method for calculating the number of interactive operations can be determined based on the total number of interactive operation records for the same pushed content (each interactive operation record corresponds to an interactive operation of a certain user on the pushed content).

[0143] Step S403: Based on the pushed content and the corresponding set popularity, the push count or interactive operation count of each pushed content in the ranking is normalized, and the processed result is used as the popularity of the pushed content.

[0144] After determining the set popularity, it is necessary to determine the popularity of each pushed content in the training sample based on the pushed content and the set popularity to determine the training sample.

[0145] Since the popularity of the pushed content at two endpoint positions and one preset sorting position has been determined in step S402, the popularity of other pushed content can be determined based on the relative sorting position of each pushed content and the three pushed content with determined popularity or the corresponding number of pushes (or number of interactive operations).

[0146] The specific method for determining the popularity of other pushed content can be implemented by using the Max-Min normalization method (or Max-Min standardization method).

[0147] For example, when the maximum popularity is 1, the minimum popularity is 0, the ranking position is set to 80%, and the corresponding popularity is set to 0.5, the specific algorithm can be expressed as:

[0148]

[0149] Among them, popularity(x) is used to represent the popularity of the pushed content with a push count (or interactive operation count) of x, x is the push count (or interactive operation count) of the pushed content, exp_min, exp_mid, and exp_max are the push counts (or interactive operation counts) corresponding to the pushed content with the minimum, 0.5, and maximum popularity, respectively. Therefore, the popularity value of each pushed content can be directly determined based on the push count (or interactive operation count) of each pushed content.

[0150] Step S404: Determine the label value corresponding to each training sample.

[0151] Specifically, the tag value is a parameter value used to indicate whether the user has performed any interactive operation on the pushed content after each push is generated. That is, the tag value corresponds to the push record one by one.

[0152] The calculation method is to determine whether there is a corresponding interactive operation record within the set condition range (such as the set duration range) of each push record when each push record only contains the process of pushing push content to a certain user once.

[0153] For example, the push record of an application records that content m was pushed to user A at 12:00 on January 1. If the corresponding interaction operation record of the application records that user A clicked, collected, shared or commented on the pushed content m before 12:10, then it is considered that the label value for the push record is the parameter value indicating that the user performed an interactive operation on the pushed content; otherwise, the label value should be the parameter value indicating that the user did not perform an interactive operation on the pushed content.

[0154] Step S405: Input the training samples into the recommendation model for training, and output user feature parameters and content feature parameters.

[0155] The content of this step can refer to Figure 2 Step S202 or Figure 3 Steps S302 to S306 in the corresponding embodiment are not described again here.

[0156] Step S406: Obtain a predicted interactive operation probability value of each user for each pushed content based on the user characteristic parameters and the content characteristic parameters.

[0157] Specifically, by taking the inner product of the output user feature parameters and content feature parameters, we can obtain a matrix of predicted interaction operation probability values corresponding to each user and each pushed content. This allows us to directly determine the degree of preference of each user for all pushed content (i.e., the predicted interaction operation probability value), and further determine the content to be pushed to the user.

[0158] Step S407: Determine the push content for recall corresponding to the target user based on the ranking of the predicted interactive operation probability values.

[0159] Specifically, according to the needs of the server, after determining the target user to be pushed, the predicted interactive operation probability value between the user and all pushed contents can be queried.

[0160] Then, based on the descending order of the predicted interaction probability values, the pushed content that the user is most likely to interact with is determined and recalled, thus completing the recall process of the content recommendation system.

[0161] According to the recommendation model training method of the embodiment of the present disclosure, the popularity of the pushed content at a specific location is determined based on the push record or interactive operation record of each pushed content, and then the popularity of other pushed content is normalized to obtain the popularity of all pushed content. Then, the label value of the corresponding push record is determined, and then the training sample is obtained. Based on the trained user feature parameters and content feature parameters, the pushed content for recall is determined. In this way, it can be ensured that the training sample intuitively reflects the popularity characteristics of each pushed content, so that the dominant role of the popularity factor can be effectively avoided when determining the content to be pushed, thereby reducing the Matthew effect of the pushed content being biased towards popularity, improving the coverage of user interest preference areas other than popularity, and bringing a better experience to users.

[0162] Exemplary media

[0163] After introducing the method of the exemplary embodiment of the present disclosure, next, refer to Figure 5 A storage medium according to an exemplary embodiment of the present disclosure is described.

[0164] refer to Figure 5 As shown, a program product 50 for implementing the above method according to an embodiment of the present disclosure is described, which can be a portable compact disc read-only memory (CD-ROM) and includes program code and can be run on a terminal device, such as a personal computer. However, the program product of the present disclosure is not limited thereto.

[0165] The program product may be implemented in any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or component, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0166] A readable signal medium may include a data signal transmitted in baseband or as part of a carrier wave, which carries readable program code. Such a transmitted data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium.

[0167] Program code for performing the operations disclosed herein can be written in any combination of one or more programming languages, including object-oriented programming languages such as Java, C++, and the like, as well as conventional procedural programming languages such as "C" or similar programming languages. The program code can be executed entirely on the user computing device, partially on the user device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving a remote computing device, the remote computing device can be connected to the user computing device via any type of network, including a local area network (LAN) or a wide area network (WAN).

[0168] Exemplary devices

[0169] After introducing the medium of the exemplary embodiment of the present disclosure, next, reference is made to Figure 6 The recommendation model training device of the exemplary embodiment of the present disclosure is described, which is used to implement the method in any of the above method embodiments. Its implementation principles and technical effects are similar and will not be repeated here.

[0170] The recommendation model training device 600 provided in the present disclosure includes:

[0171] A generation module 610 is configured to generate training samples based on push records of pushed content and user interaction records on the pushed content. The training samples include a popularity value corresponding to the pushed content and a label value indicating whether the user has performed an interaction operation on each pushed content.

[0172] The training module 620 is used to input training samples into the recommendation model for training, and output user feature parameters and content feature parameters. The recommendation model is a matrix decomposition model.

[0173] In an exemplary embodiment of the present disclosure, the training module 620 is specifically used to: determine the predicted interaction operation probability value and predicted popularity corresponding to the training sample based on the user feature parameters and content feature parameters corresponding to the recommendation model; and adjust the user feature parameters and content feature parameters based on the difference between the label value in the training sample and the predicted interaction operation probability value, and the difference between the popularity and the predicted popularity, until the training end conditions are met.

[0174] In an exemplary embodiment of the present disclosure, the training module 620 is specifically used to: when both the user feature parameters and the content feature parameters include popularity factors and preference factors, determine the predicted interaction operation probability value and predicted popularity corresponding to the training sample based on the user feature parameters and content feature parameters corresponding to the recommendation model, including: determining the predicted interaction operation probability value according to the sum of the inner product of the popularity factors of the user feature parameters and the content feature parameters and the inner product of the preference factors of the user feature parameters and the content feature parameters; determining the predicted popularity according to the difference between the inner product of the popularity factors of the user feature parameters and the content feature parameters and the inner product of the preference factors of the user feature parameters and the content feature parameters.

[0175] In an exemplary embodiment of the present disclosure, the training module 620 is specifically used to: when the label value in the training sample indicates that the user has performed at least one interactive operation on the pushed content, determine the predicted popularity based on the result of subtracting the inner product of the preference factor of the user feature parameter and the content feature parameter from the inner product of the popularity factor of the user feature parameter and the content feature parameter; when the label value in the training sample indicates that the user has not performed any interactive operation on the pushed content, determine the predicted popularity based on the result of subtracting the inner product of the popularity factor of the user feature parameter and the content feature parameter from the inner product of the preference factor of the user feature parameter and the content feature parameter.

[0176] In an exemplary embodiment of the present disclosure, the training module 620 is specifically used to: when the model parameters of the recommendation model also include a preset bias coefficient, determine the adjusted predicted interaction operation probability value and predicted popularity based on the user feature parameters, content feature parameters and bias coefficient corresponding to the recommendation model; and, based on the difference between the adjusted predicted interaction operation probability value and the label value, and the difference between the adjusted predicted popularity and the popularity, adjust the user feature parameters and content feature parameters until the training end conditions are met.

[0177] In an exemplary embodiment of the present disclosure, the generation module 610 is specifically used to: determine the popularity of the push content at both ends of the sorting and the preset sorting position as the set popularity based on the number of pushes of each push content in the push record or the number of interaction operations in the interaction operation record; determine the popularity of each push content in the training sample based on the push content and the set popularity as a benchmark.

[0178] In an exemplary embodiment of the present disclosure, the generation module 610 is specifically used to: based on the pushed content and the corresponding set popularity, normalize the number of pushes or interactive operations of each pushed content in the sorting, and use the processed result as the popularity of the pushed content.

[0179] In an exemplary embodiment of the present disclosure, the training module 620 is also used to: after inputting the training samples into the recommendation model for training and outputting the user feature parameters and content feature parameters, obtain the predicted interactive operation probability value of each user for each pushed content based on the user feature parameters and content feature parameters; based on the ranking of the predicted interactive operation probability values, determine the pushed content for recall corresponding to the target user.

[0180] Exemplary Computing Devices

[0181] After introducing the method, medium and apparatus of the exemplary embodiments of the present disclosure, Figure 7 A computing device according to an exemplary embodiment of the present disclosure is described.

[0182] Figure 7 The computing device 70 shown is only an example and should not bring any limitation to the functionality and scope of use of the embodiments of the present disclosure.

[0183] like Figure 7 As shown, computing device 70 is implemented as a general-purpose computing device. Components of computing device 70 may include, but are not limited to, at least one processing unit 701, at least one storage unit 702, and a bus 703 connecting various system components (including processing unit 401 and storage unit 702).

[0184] The bus 703 includes a data bus, a control bus, and an address bus.

[0185] The storage unit 702 may include a readable medium in the form of a volatile memory, such as a random access memory (RAM) 7021 and / or a cache memory 7022 , and may further include a readable medium in the form of a non-volatile memory, such as a read-only memory (ROM) 4023 .

[0186] The storage unit 702 may also include a program having a set (at least one) of program modules 7024, such program modules 7024 including but not limited to: an operating system, one or more application programs, other program modules and program data, each of which or some combination may include an implementation of a network environment.

[0187] The computing device 70 may also communicate with one or more external devices 704 (e.g., a keyboard, a pointing device, etc.). Such communication may be performed via an input / output (I / O) interface 705. Furthermore, the computing device 70 may also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network such as the Internet) via a network adapter 706. Figure 7As shown, network adapter 706 communicates with other modules of computing device 70 via bus 703. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with computing device 70, including but not limited to microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0188] It should be noted that although several units / modules or sub-units / modules of the recommendation model training device are mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided to be embodied by multiple units / modules.

[0189] Furthermore, although the operations of the disclosed method are described in a particular order in the accompanying drawings, this does not require or imply that the operations must be performed in this particular order, or that all illustrated operations must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0190] Although the spirit and principles of the present disclosure have been described with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the specific embodiments disclosed, and the division into various aspects does not mean that the features in these aspects cannot be combined to benefit. Such division is only for the convenience of expression. The present disclosure is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the appended claims.

Claims

1. A recommendation model training method, comprising the following steps: Generate training samples based on push records of pushed content and user interaction operation records on the pushed content, wherein the training samples include the popularity of the pushed content and the label value of whether the user has performed an interaction operation on each pushed content; Input the training samples into the recommendation model for training, and output user feature parameters and content feature parameters, wherein the recommendation model is a matrix decomposition model; The step of inputting the training samples into the recommendation model for training and outputting user feature parameters and content feature parameters includes: Determining a predicted interactive operation probability value and a predicted popularity corresponding to the training sample based on user feature parameters and content feature parameters corresponding to the recommendation model; Adjusting the user feature parameters and content feature parameters based on the difference between the label value in the training sample and the predicted interactive operation probability value, and the difference between the popularity and the predicted popularity, until a training end condition is met; The user characteristic parameters and content characteristic parameters both include popularity factors and preference factors; The determining, based on the user characteristic parameters and content characteristic parameters corresponding to the recommendation model, the predicted interactive operation probability value and the predicted popularity corresponding to the training sample includes: determining the predicted interactive operation probability value according to the sum of the inner product of the popularity factor of the user characteristic parameter and the content characteristic parameter and the inner product of the preference factor of the user characteristic parameter and the content characteristic parameter; The predicted popularity is determined according to the difference between the inner product of the popularity factor of the user characteristic parameter and the content characteristic parameter and the inner product of the preference factor of the user characteristic parameter and the content characteristic parameter.

2. The recommendation model training method according to claim 1, wherein determining the predicted popularity based on the difference between the inner product of the popularity factor of the user characteristic parameter and the content characteristic parameter and the inner product of the preference factor of the user characteristic parameter and the content characteristic parameter comprises: When the label value in the training sample indicates that the user has performed at least one interactive operation on the pushed content, determining the predicted popularity based on the result of subtracting the inner product of the preference factor of the user characteristic parameter and the content characteristic parameter from the inner product of the popularity factor of the user characteristic parameter and the content characteristic parameter; When the label value in the training sample indicates that the user has not performed any interactive operation on the pushed content, the predicted popularity is determined based on the result of subtracting the inner product of the popularity factor of the user feature parameter and the content feature parameter from the inner product of the preference factor of the user feature parameter and the content feature parameter.

3. The recommendation model training method according to claim 1, wherein the model parameters of the recommendation model further include a preset bias coefficient; The determining, based on the user characteristic parameters and content characteristic parameters corresponding to the recommendation model, the predicted interactive operation probability value and the predicted popularity corresponding to the training sample includes: Determining an adjusted predicted interactive operation probability value and predicted popularity based on the user characteristic parameters, content characteristic parameters, and the bias coefficient corresponding to the recommendation model; Accordingly, adjusting the user characteristic parameters and the content characteristic parameters based on the difference between the label value in the training sample and the predicted interactive operation probability value, and the difference between the popularity and the predicted popularity, includes: Based on the difference between the adjusted predicted interaction operation probability value and the label value, and the difference between the adjusted predicted popularity and the popularity, the user feature parameters and the content feature parameters are adjusted until a training end condition is met.

4. The recommendation model training method according to any one of claims 1 to 3, wherein generating training samples based on push records of pushed content and user interaction operation records on the pushed content comprises: Based on the number of pushes of each pushed content in the push record or the number of interactive operations in the interactive operation record, the popularity of the pushed content at both ends of the sorting and the preset sorting position is determined as the set popularity; The popularity of each pushed content in the training sample is determined based on the pushed content and the set popularity.

5. The recommendation model training method according to claim 4, wherein determining the popularity of each pushed content in the training sample based on the pushed content and the set popularity as a benchmark comprises: Based on the pushed content and the corresponding set popularity, the number of pushes or the number of interactive operations of each pushed content in the ranking is normalized, and the processed result is used as the popularity of the pushed content.

6. The recommendation model training method according to any one of claims 1 to 3, further comprising: Obtaining a predicted interactive operation probability value for each user on each pushed content based on the user characteristic parameters and the content characteristic parameters; Based on the ranking of the predicted interactive operation probability values, push content for recall corresponding to the target user is determined.

7. A computer-readable storage medium comprising: The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the recommendation model training method according to any one of claims 1 to 6.

8. A recommendation model training device comprising: A generation module is used to generate training samples based on push records of pushed content and user interaction operation records on the pushed content. The training samples include the popularity of the pushed content and the label value of whether the user has performed an interaction operation on each pushed content; A training module, configured to input the training samples into a recommendation model for training and output user feature parameters and content feature parameters, wherein the recommendation model is a matrix decomposition model; The training module is specifically configured to determine a predicted interactive operation probability value and a predicted popularity corresponding to a training sample based on user feature parameters and content feature parameters corresponding to the recommendation model; and adjust the user feature parameters and content feature parameters based on the difference between the label value and the predicted interactive operation probability value, and the difference between the popularity and the predicted popularity, in the training sample, until a training end condition is met; The training module is specifically used to: when both the user feature parameters and the content feature parameters include popularity factors and preference factors, determine the predicted interactive operation probability value based on the sum of the inner product of the popularity factors of the user feature parameters and the content feature parameters and the inner product of the preference factors of the user feature parameters and the content feature parameters; determine the predicted popularity based on the difference between the inner product of the popularity factors of the user feature parameters and the content feature parameters and the inner product of the preference factors of the user feature parameters and the content feature parameters.

9. The recommendation model training device according to claim 8, wherein the training module is specifically used for: when the label value in the training sample indicates that the user has performed at least one interactive operation on the pushed content, the predicted popularity is determined based on the result of subtracting the inner product of the preference factor of the user characteristic parameter and the content characteristic parameter from the inner product of the popularity factor of the user characteristic parameter and the content characteristic parameter; when the label value in the training sample indicates that the user has not performed any interactive operation on the pushed content, the predicted popularity is determined based on the result of subtracting the inner product of the popularity factor of the user characteristic parameter and the content characteristic parameter from the inner product of the preference factor of the user characteristic parameter and the content characteristic parameter.

10. The recommendation model training device according to claim 8, wherein the training module is specifically used for: when the model parameters of the recommendation model also include a preset bias coefficient, determining the adjusted predicted interaction operation probability value and predicted popularity based on the user feature parameters, content feature parameters and bias coefficient corresponding to the recommendation model; and adjusting the user feature parameters and content feature parameters based on the difference between the adjusted predicted interaction operation probability value and the label value, and the difference between the adjusted predicted popularity and the popularity, until the training end conditions are met.

11. According to the recommendation model training device according to any one of claims 8 to 10, the generation module is specifically used to: based on the number of pushes of each pushed content in the push record or the number of interaction operations in the interaction operation record, determine the popularity of the pushed content at both ends of the sorting and the preset sorting position as the set popularity; based on the push content and the set popularity as a benchmark, determine the popularity of each pushed content in the training sample.

12. The recommendation model training device according to claim 11, wherein the generation module is specifically used to: normalize the number of pushes or interactive operations of each pushed content in the sorting based on the pushed content and the corresponding set popularity, and use the processed result as the popularity of the pushed content.

13. According to the recommendation model training device according to any one of claims 8 to 10, the training module is also used to: after inputting the training sample into the recommendation model for training and outputting the user feature parameters and content feature parameters, obtain the predicted interactive operation probability value of each user for each pushed content based on the user feature parameters and content feature parameters; based on the ranking of the predicted interactive operation probability values, determine the push content to be pushed to the target user.

14. A computing device comprising: at least one processor; and a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the computing device to perform the recommendation model training method according to any one of claims 1 to 6.

Images