Content recommendation method, device, equipment and medium

Abstract

This application discloses a content recommendation method, apparatus, device, and medium, relating to the field of the internet. The method comprises: obtaining first sample data corresponding to a first account set, the first sample data being used to represent the first account's interest in candidate content to be recommended; performing clustering processing on the first sample data to obtain a first relationship matrix, the first relationship matrix being used to represent the clustering of the candidate content; obtaining second sample data corresponding to a second account set, the second sample data including known interest levels between the second account and the candidate content; performing matrix decomposition processing on a second sample matrix corresponding to the second sample data to obtain a second relationship matrix corresponding to the first relationship matrix, the second relationship matrix being used to represent interest levels between the second account and the cluster categories; and determining, from a third relationship matrix determined from the first and second relationship matrices, to recommend candidate content to the second account. This method ensures data security.

Description

Technical Field

[0001] The present application relates to the field of Internet, and in particular to a content recommendation method, apparatus, device and medium. Background Art

[0002] Personalized recommendations are prevalent in various Internet services. Internet platforms need to recommend corresponding content to users based on their preferences in order to obtain higher click-through rates for the corresponding content.

[0003] The personalized recommendation system is a widely used information filtering system. Internet platforms collect a large amount of personal data from users, analyze it, and then build a personalized recommendation system based on the analysis results. This personalized recommendation system can make targeted recommendations based on the explicit and implicit preferences of different users.

[0004] However, during the design process of the above-mentioned personalized recommendation system, its training and analysis requires a large amount of data support, and some of the user's personal data also includes sensitive data. This sensitive data, as public training data, poses security and leakage issues for the user's personal privacy data, and the security of the personalized recommendation system is relatively low. Summary of the Invention

[0005] The embodiments of the present application provide a content recommendation method, apparatus, device, and medium that can improve data security during content recommendation. The technical solution is as follows:

[0006] In one aspect, a content recommendation method is provided, the method comprising:

[0007] Obtaining first sample data corresponding to the first account set, where the first sample data is used to represent the first account's interest in the candidate content to be recommended;

[0008] Performing clustering processing on the first sample data to obtain a first relationship matrix, where the first relationship matrix is ​​used to represent the clustering of the candidate content;

[0009] Obtaining second sample data corresponding to a second account set, where the second account set includes second accounts corresponding to the target platform, and the second sample data includes known interest levels between the second accounts and the candidate content;

[0010] Performing matrix decomposition processing on a second sample matrix corresponding to the second sample data to obtain a second relationship matrix corresponding to the first relationship matrix, where the second relationship matrix is ​​used to represent the interest level between the second account and the cluster category;

[0011] Determine to recommend the candidate content to the second account from a third relationship matrix determined from the first relationship matrix and the second relationship matrix.

[0012] In another aspect, a content recommendation device is provided, the device comprising:

[0013] An acquisition module, configured to acquire first sample data corresponding to a first account set, wherein the first sample data is used to represent the first account's interest in the candidate content to be recommended;

[0014] a first determining module, configured to perform clustering processing on the first sample data to obtain a first relationship matrix, wherein the first relationship matrix is ​​used to represent the clustering of the candidate content;

[0015] The acquisition module is further configured to acquire second sample data corresponding to a second account set, the second account set including second accounts corresponding to the target platform, the second sample data including known interest levels between the second accounts and the candidate content;

[0016] a second determining module, configured to perform matrix decomposition processing on a second sample matrix corresponding to the second sample data to obtain a second relationship matrix corresponding to the first relationship matrix, wherein the second relationship matrix is ​​used to represent the interest level between the second account and the cluster category;

[0017] A recommendation module is configured to determine, from a third relationship matrix determined by the first relationship matrix and the second relationship matrix, to recommend the candidate content to the second account.

[0018] On the other hand, a computer device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set, or an instruction set, and the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement any content recommendation method described in the embodiments of the present application.

[0019] On the other hand, a computer-readable storage medium is provided, in which at least one program code is stored. The program code is loaded and executed by a processor to implement any content recommendation method described in the embodiments of the present application.

[0020] In another aspect, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the content recommendation method described in any of the above embodiments.

[0021] The technical solution provided by this application includes at least the following beneficial effects:

[0022] In order to obtain the unknown interest level between the second account and the candidate content, the first relationship matrix is ​​determined by the first sample data corresponding to the first account set, and the second relationship matrix is ​​determined by the second sample data corresponding to the second account set, wherein the training processes of the first relationship matrix and the second relationship matrix are independent of each other. The first relationship matrix represents the clustering of the candidate content. Since the clustering of the candidate content is an objective attribute of the candidate content, the training of the first relationship matrix does not need to be completed through the full data, so the training process of the first relationship matrix can also be completed independently of the target platform, that is, the target platform can obtain the trained first relationship matrix from other platforms, and then obtain the second relationship matrix based on the first relationship matrix and the second sample data of the second account set corresponding to the target platform. The second relationship matrix represents the interest level between the second account and the cluster category. Then, based on the first relationship matrix and the second relationship matrix, the target platform can determine the interest level between the second account and the candidate content, so as to recommend candidate content that meets the corresponding interest level of the second account to the second account. Since the training process of the first relationship matrix can be carried out independently of the target platform, a distributed server can be used to obtain the first relationship matrix and the second relationship matrix. The first sample data as shared data can be trained in the shared server in the distributed server to obtain the first relationship matrix, while the second user data as non-shared data does not need to be left the domain and can be trained in the server corresponding to the target platform. Therefore, the possibility of leakage of user data in the content recommendation process is reduced, and the security of data in the content recommendation process is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0024] Figure 1 is a schematic diagram of a latent semantic model provided by an exemplary embodiment of the present application;

[0025] Figure 2 This is a schematic diagram of an implementation environment provided by an exemplary embodiment of the present application;

[0026] Figure 3 is a flow chart of a content recommendation method provided by an exemplary embodiment of the present application;

[0027] Figure 4is a flow chart of a method for determining a first relationship matrix provided by an exemplary embodiment of the present application;

[0028] Figure 5 is a flow chart of a method for determining a second relationship matrix provided by an exemplary embodiment of the present application;

[0029] Figure 6 is a flow chart of a content recommendation method provided by another exemplary embodiment of the present application;

[0030] Figure 7 is a mean square error histogram provided by an exemplary embodiment of the present application;

[0031] Figure 8 is a mean square error reduction graph provided by an exemplary embodiment of the present application;

[0032] Figure 9 is a mean square error reduction graph provided by another exemplary embodiment of the present application;

[0033] Figure 10 is a schematic diagram of a content recommendation system provided by an exemplary embodiment of the present application;

[0034] Figure 11 is a schematic diagram of a data processing module provided by an exemplary embodiment of the present application;

[0035] Figure 12 is a structural block diagram of a content recommendation device provided by an exemplary embodiment of the present application;

[0036] Figure 13 is a structural block diagram of a content recommendation device provided by another exemplary embodiment of the present application;

[0037] Figure 14 It is a structural diagram of a server provided by another exemplary embodiment of the present application. DETAILED DESCRIPTION

[0038] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0039] First, a brief introduction to the terms involved in the embodiments of this application is given:

[0040] Artificial Intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also studies the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0041] Artificial intelligence (AI) technology is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0042] Machine learning (ML) is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of AI. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learning through demonstration.

[0043] Recommendation systems: An information filtering system that learns users' interests and hobbies based on their historical usage behavior and predicts their preferences for items or content. This system changes the way platforms communicate with users, enhancing communication efficiency and reducing communication costs. The effectiveness of traditional recommendation systems is proportional to the amount of user data. Existing recommendation systems collect and analyze large amounts of personal and sensitive user data, raising security and data leakage issues.

[0044] In the recommendation system in the embodiment of the present application, among the parameters corresponding to the system, the training of some parameters is realized through shared data, and the training of some parameters is carried out through the platform's own unique data. Since the data is not out of the domain, the part trained with unique data protects the security of user data and avoids the problem of data leakage.

[0045] Latent Factor Model: It is a latent semantic analysis technology used to find the underlying themes or categories between things. In the recommendation system, the latent semantic model can automatically cluster the content based on the user's behavior, that is, divide the content into different categories. The above categories can be understood as the user's interests. Figure 1 As shown, the R matrix 101 is a matrix used to indicate user-content interest, and the matrix parameter R ij represents the interest of user i in content j. The P matrix 102 is a matrix used to indicate user-category interest. ik It represents the interest of user i in category k. The Q matrix 103 is a weight matrix for indicating category-content. kj It represents the weight of content j in category k, where 0<i≤3, 0<j≤4, 0<k≤3, and i, j, and k are all integers.

[0046] The solutions provided in the embodiments of this application involve technologies such as artificial intelligence machine learning. The content recommendation method provided in this application can be applied to the following scenarios:

[0047] Schematically, this method is described by taking the application of it to a film and television / film review platform as an example. Schematically, the film and television / film review platform obtains public or shared user data from other films and television / film reviews, and uses this part of user data to train the parameters used to characterize "film and television works-categories" in the recommendation system, and obtains parameters that can be used to represent the correspondence between film and television works and categories. Then, private user data within its own platform is obtained. This part of user data will not be shared with other film and television / film review platforms. The platform uses this part of private user data to train the parameters used to characterize "users-categories" in the recommendation system, and obtains parameters that can be used to represent the correspondence between users and categories. The platform can configure the recommendation system through the above two parameters, and then use the recommendation system to predict the user's interests and recommend relevant video content based on the prediction results.

[0048] This method can also be applied to recommendation systems in article reading platforms, social platforms, shopping platforms, search engines, etc. It is worth noting that the above application scenarios are only illustrative examples, and the content recommendation method provided in this application can be applied to other application scenarios including recommendation systems.

[0049] In combination with the above-mentioned noun introduction and application scenarios, the implementation environment of the embodiments of the present application is explained.

[0050] Please refer to Figure 2 The implementation environment includes a terminal 201, a server 202, and a communication network 203, wherein the server 202 includes a target platform server 204 and a shared server 205.

[0051] Terminal 201 can be an electronic device such as a mobile phone, tablet computer, e-book reader, multimedia player, wearable device, laptop computer, desktop computer, or biometric identification integrated device. Illustratively, terminal 201 has a target application installed, which includes a service module for recommending candidate content to the user. This service module recommends content when the user uses the target application. Illustratively, the target application can be a traditional application software, a cloud application software, a mini-program or application module within a host application, or a web platform, without limitation.

[0052] The server 202 is used to provide functional services to the terminal 201, wherein the functional services include content recommendation services. Schematically, the server 202 is a distributed server, including a target platform server 204 and a shared server 205, wherein the target platform server 204 is a server corresponding to the target application, and the shared server 205 is a server for providing shared services to multiple target platform servers 204. Schematically, a communication connection is established between the target platform server 204 and the shared server 205. In one example, the target platform server 204 obtains shared parameters from the shared server 205, and the shared parameters represent the clustering of the candidate content to be recommended. The training of the shared parameters is completed by the shared server 205. The target platform server 204 obtains private parameters based on the shared parameters and the platform's own user data training. The private parameters represent the interest between the second account corresponding to the target platform and the cluster category. The target platform server 204 stores the shared parameters and the private parameters. When the target application in the terminal 201 needs to recommend content to the user, the terminal 201 sends a content recommendation request to the target platform server 204. The target platform server 204 reads the stored shared parameters and private parameters according to the content recommendation request, and determines the recommendation value of the candidate content to be recommended relative to the target account corresponding to the terminal 201 based on the shared parameters and private parameters, and determines the candidate content to be returned to the terminal 201 based on the above recommendation value.

[0053] Optionally, target platform server 204 is a physical server or a cloud server. Target platform server 204 can be a single server, a server cluster consisting of multiple servers, or a cloud computing service center. Shared server 205 is a physical server or a cloud server. Shared server 205 can be a single server, a server cluster consisting of multiple servers, or a cloud computing service center.

[0054] The server 202 can establish a communication connection with the terminal 201 via a communication network 203. The network can be a wireless network or a wired network.

[0055] Please refer to Figure 3 , which shows a flow chart of a content recommendation method provided by an exemplary embodiment of the present application. In the embodiment of the present application, the method is described as being executed by a server in an implementation environment. The process of obtaining the first relationship matrix and the second relationship matrix is ​​implemented through a two-stage process. The method includes:

[0056] Step 301: Obtain first sample data corresponding to a first account set.

[0057] The first account in the first account set may be an account from a target platform and other platforms associated with the target platform. Schematically, a communication connection is established between the target platform server and the shared server, and a communication connection is also established between the shared server and the servers corresponding to the above-mentioned other platforms. The shared server is used to provide sharing services for multiple platforms, including data sharing services. Schematically, the shared server stores user data from multiple platforms with a connection relationship. The above-mentioned user data includes personal data actively provided by the user, and may also include personal data disclosed by the user. The above-mentioned user data is stored in the shared server and provided to the platform server that has established a communication connection with the shared server. The platform server can obtain the above-mentioned user data from the shared server.

[0058] In an embodiment of the present application, data processing is performed on the relevant data in the user data to obtain first sample data corresponding to a first account set. The first sample data is shared data and is used to represent the first account's interest in the candidate content to be recommended. Illustratively, the user data includes preference data of the user corresponding to the first account for the candidate content. Illustratively, the preference data includes explicit preference data and implicit preference data. In one example, using the content recommendation method applied to a movie review website, if the first account has reviewed movie A, then the first account has explicit preference data for movie A, where movie A is one of the candidate content to be recommended. If the user corresponding to the first account has not reviewed movie A, but has browsed it, for example, by browsing / saving movie A's introduction or browsing / saving movie A's review, then the user is likely interested in movie A, and therefore the first account has implicit preference data for movie A.

[0059] Step 302: Perform clustering processing on the first sample data to obtain a first relationship matrix.

[0060] In an embodiment of the present application, a two-stage training method is used to obtain the first relationship matrix and the second relationship matrix, that is, the first relationship matrix is ​​first obtained through the first sample data, and then the second relationship matrix is ​​obtained through the second sample data. Schematically, the target platform server can obtain the above-mentioned first sample data from the shared server, and then perform preset processing on the first sample data to obtain the first relationship matrix, or directly obtain the first relationship matrix obtained by the shared server performing preset processing on the first sample data from the shared server, which is not limited here. In an embodiment of the present application, a distributed server is used to train the first relationship matrix and the second relationship matrix. The target platform server obtains the first relationship matrix from the shared server as an example. The shared server obtains the first relationship matrix based on the first sample data. The target platform server sends a data acquisition request to the shared server. The shared server authenticates the data acquisition request. The shared server determines that the target platform server has the authority to obtain the above-mentioned first relationship matrix, and then sends the first relationship matrix to the target platform server.

[0061] Schematically, a first sample matrix is ​​established based on the first sample data. If the first sample data indicates that the first account has explicit preference data for the candidate content, the explicit preference data is used as the matrix parameter of the first sample matrix; if the first sample data indicates that the first account has implicit preference data for the candidate content, the implicit preference data is used as the matrix parameter of the first sample matrix. Schematically, the implicit preference data is represented by first preset data. If the first sample data does not contain the first account's preference data for the candidate content, the second preset data is used as the matrix parameter of the first sample matrix. Schematically, the second preset data can be "0".

[0062] In the embodiment of the present application, the content recommendation method is applied to a movie review website as an example for explanation. The candidate content includes Movie 1, Movie 2, Movie 3, and Movie 4. The first account set includes First Account A, First Account B, and First Account C. If the first account has rated the candidate content, the value corresponding to the rating is used as the corresponding matrix parameter, where the rating ranges from 1 to 5. If the first account has not rated the candidate content but has browsed it, the corresponding matrix parameter is determined to be 1. If there is neither a rating nor a browsing operation, the corresponding matrix parameter is determined to be 0. The first sample matrix is ​​shown in Table 1.

[0063] Table 1

[0064] Movie 1 Movie 2 Movie 3 Movie 4 First Account A 4 0 0 1 First Account B 1 1 2 3 First Account C 0 5 2 0

[0065] Schematically, there may be direct or indirect relationships between candidate contents, so the candidate contents can be classified. Among them, because different users have different ways of dividing different contents, there are different classification methods. For example, Movie 1 is a movie containing robot elements, and Movie 4 is a movie containing conscious life elements. Therefore, Movie 1 and Movie 4 can both be classified as science fiction movies based on the subject matter of the movies. Therefore, it can be inferred that User A corresponding to the first account A is interested in science fiction movies; Movie 1, Movie 2, Movie 3, and Movie 4 are all starred by Actor A. Therefore, Movie 1, Movie 2, Movie 3, and Movie 4 can all be classified as movies related to Actor A based on the actor object. Therefore, it can be inferred that User B corresponding to the first account B is interested in movies starring A. Therefore, in the embodiment of the present application, it is necessary to cluster the candidate contents using the first sample data to obtain a first relationship matrix for representing the clustering of the candidate contents. That is, the first relationship matrix is ​​used to represent the clustering of the candidate contents.

[0066] Schematically, candidate contents are clustered using a deep learning-based clustering algorithm, wherein the above-mentioned clustering algorithm includes but is not limited to SOM (Self-organizing Maps) algorithm, BP (Error Back Propagation, multi-layer feedforward network) algorithm, K-means (K-means Clustering Algorithm, k-means clustering) algorithm and other algorithms, and the clustering results are converted to obtain a first relationship matrix.

[0067] Schematically, the first relationship matrix can also be obtained through the latent semantic model, that is, the first sample matrix is ​​subjected to matrix decomposition to obtain the first relationship matrix for representing the clustering of the candidate content. In the embodiment of the present application, the first relationship matrix obtained through matrix decomposition is used as an example for explanation.

[0068] Schematically, the above-mentioned preset processing method for the first sample data includes: obtaining a first initial matrix of candidate content and cluster categories; performing parameter fitting training on the matrix parameters in the first initial matrix in combination with the first sample data to obtain first matrix parameters for representing the clustering of the candidate content; and determining a first relationship matrix containing the first matrix parameters. The matrix parameters in the first initial matrix are an estimated value used to represent the weight of the candidate content in the cluster category before training. The above-mentioned first initial matrix can be one of a zero matrix, a unit matrix, a mean matrix, or other matrices with preset matrix parameters. The first sample matrix corresponding to the first sample data is used to perform parameter fitting training on the matrix parameters in the first initial matrix to obtain the first matrix parameters. The training process is to continuously update the matrix parameters in the first initial matrix until the relationship between the first matrix parameters in the corresponding first relationship matrix approaches the relationship between the candidate content and cluster categories implied by the first sample data.

[0069] Step 303: Obtain second sample data corresponding to the second account set.

[0070] The target platform server stores second sample data corresponding to a second set of accounts. This second sample data is non-shared data. The second set of accounts includes the second account corresponding to the target platform. The second sample data includes known interest between the second account and the candidate content. That is, the user data corresponding to the second account includes data indicating the second account's interest in some of the candidate content. This data can be obtained through the second account's rating of the candidate content, through the second account's browsing of the candidate content, or through the second account's attention to tags related to the candidate content, without limitation. However, the above data does not include all candidate content, that is, there may be another portion of candidate content whose interest in the second account is unknown. Illustratively, the first sample data and the second sample data can be two data sets with a non-empty intersection or two data sets without an intersection, without limitation. The data set in the second sample data that does not belong to the first sample data will not be transmitted to the shared server or other platform servers. That is, this portion of data is private to the target platform server and is guaranteed secure by the target platform server.

[0071] Step 304 : Perform matrix decomposition processing on the second sample matrix corresponding to the second sample data to obtain a second relationship matrix corresponding to the first relationship matrix.

[0072] Illustratively, the target platform server obtains the second sample data from a database and then performs pre-defined processing on the second sample data to obtain a second relationship matrix, which is used to represent the interest between the second account and the cluster category. A second sample matrix is ​​established based on the second sample data. The method for establishing the second sample matrix is ​​the same as the method for establishing the first sample matrix and is not further described here. The matrix parameters in the second sample matrix include known and unknown interest levels of the second account in the candidate content.

[0073] The target platform server obtains a second relationship matrix based on the second sample data and the first relationship matrix. Schematically, in an embodiment of the present application, the second relationship matrix is ​​obtained through a latent semantic model, that is, the second sample matrix is ​​subjected to matrix decomposition to obtain a second relationship matrix for representing the interest level between the second account and the cluster category.

[0074] In an embodiment of the present application, the preset processing method for the second sample data includes: obtaining a second initial matrix of the second account and the cluster category; performing parameter fitting training on the matrix parameters in the second initial matrix in combination with the second sample matrix and the first relationship matrix to obtain second matrix parameters for representing the interest level between the second account and the cluster category; and determining a second relationship matrix containing the second matrix parameters. The matrix parameters in the second initial matrix are an estimated value for representing the interest level of the second account in the cluster category before training. The second initial matrix can be a zero matrix, a unit matrix, a mean matrix, or other matrices with preset matrix parameters. The second sample matrix corresponding to the second sample data is used to perform parameter fitting training on the matrix parameters in the second initial matrix to obtain the second matrix parameters. The training process is to continuously update the matrix parameters in the second initial matrix until the relationship between the second matrix parameters in the corresponding second relationship matrix approaches the relationship between the second account and the cluster category implied by the second sample data.

[0075] The acquisition process of the second relationship matrix is ​​completed in the target platform server, and the second sample data, which is non-shared data, will not leave the domain, thereby ensuring the data security of the second sample data and protecting the privacy information of the second account.

[0076] Step 305 : Determine candidate content to recommend to the second account from a third relationship matrix determined by the first relationship matrix and the second relationship matrix.

[0077] In an embodiment of the present application, the process of determining the first relationship matrix and the second relationship matrix can be determined in real time when the terminal initiates a content recommendation request to the target platform server; or the target platform server obtains sample data at a preset frequency, processes the sample data, and stores the determined first relationship matrix and the second relationship matrix in a database. In one example, the target platform server obtains the first relationship matrix from the shared server as an example. The shared server obtains the first sample data at a preset frequency and obtains the first relationship matrix based on the first sample data, and the target platform server obtains the first relationship matrix from the shared server at the same preset frequency and stores it in the database. Schematically, the target platform server can determine the second relationship matrix based on the first relationship matrix and the second sample data at the preset frequency, or at other frequencies based on the latest first relationship matrix stored in the database, which is not limited here. Schematically, the preset frequency can be set to once every 12 hours, once a day, once a week, or other frequencies, which is not limited here. When the terminal sends a content recommendation request to the target platform server, the target platform server directly obtains the first relationship matrix and the second relationship matrix from the database.

[0078] Determining the first relationship matrix means determining the weight of the candidate content in the cluster category, and determining the second relationship matrix means determining the interest of the second account in the cluster category. Therefore, based on the first relationship matrix and the second relationship matrix, the recommendation value of the candidate content to be recommended by the second account can be determined.

[0079] When the target platform server receives a content recommendation request for a second account from a terminal, it obtains the determined first and second relationship matrices and recommends suitable candidate content to the second account based on the first and second relationship matrices. The content recommendation request is used to request the target platform server to send candidate content that meets the interests of the second account to the terminal corresponding to the second account. The target platform server determines whether to recommend the candidate content to the first account based on the recommendation value corresponding to the candidate content. In an exemplary embodiment, the inner product of the first and second relationship matrices is calculated to obtain a third relationship matrix. The matrix parameters in the third relationship matrix are used to determine the recommendation value of the candidate content relative to the second account. In response to the recommendation value being greater than a preset recommendation value, the candidate content is recommended to the second account.

[0080] In summary, the content recommendation method provided in the embodiment of the present application is to obtain the unknown interest situation between the second account and the candidate content, determine the first relationship matrix through the first sample data corresponding to the first account set, and determine the second relationship matrix through the second sample data corresponding to the second account set, wherein the training processes of the first relationship matrix and the second relationship matrix are independent of each other, and the first relationship matrix represents the clustering situation of the candidate content. Since the clustering situation of the candidate content is an objective attribute of the candidate content, the training of the first relationship matrix does not need to be completed through the full amount of data, so the training process of the first relationship matrix can also be completed independently of the target platform, that is, the target platform can obtain the trained first relationship matrix from other platforms, and then obtain the second relationship matrix based on the first relationship matrix and the second sample data of the second account set corresponding to the target platform. The second relationship matrix represents the interest situation between the second account and the cluster category, and then the target platform can determine the interest situation between the second account and the candidate content based on the first relationship matrix and the second relationship matrix, so as to recommend candidate content that meets the corresponding interest level of the second account to the second account. Since the training process of the first relationship matrix can be carried out independently of the target platform, a distributed server can be used to obtain the first relationship matrix and the second relationship matrix. The first sample data as shared data can be trained in the shared server in the distributed server to obtain the first relationship matrix, while the second user data as non-shared data does not need to be left the domain and can be trained in the server corresponding to the target platform. Therefore, the possibility of leakage of user data in the content recommendation process is reduced, and the security of data in the content recommendation process is improved.

[0081] Please refer to Figure 4 , which shows a first relationship matrix determination method provided in an embodiment of the present application. The method can be executed by a target platform server or a shared server, and is not limited here. The method includes:

[0082] Step 401: Obtain first sample data corresponding to a first account set.

[0083] The first sample data is used to represent the first account's interest in the candidate content to be recommended. The first sample data is processed to obtain a first sample matrix.

[0084] Step 402: Obtain a first initial matrix of candidate content and cluster categories.

[0085] The matrix parameters in the first initial matrix can be one of a zero matrix, an identity matrix, a mean matrix, or other matrices with preset matrix parameters, which are not limited here. Schematically, the first initial matrix corresponds to the Q matrix in the latent semantic model, and a first auxiliary matrix is ​​obtained at the same time. The first auxiliary matrix is ​​the P matrix in the latent semantic model. The matrix parameters in the first auxiliary matrix can be one of a zero matrix, an identity matrix, a mean matrix, or other matrices with preset matrix parameters, which are not limited here.

[0086] Step 403: Obtain a first loss function.

[0087] The first loss function is used to determine the error between the clustering of the candidate content expressed by the first initial matrix during the training process and the clustering of the candidate content corresponding to the first sample data. Schematically, in the embodiment of the present application, the first loss function adopts the mean square error loss function. In order to prevent overfitting during the model training process, a penalty term is added after the mean square error loss function to obtain the first loss function, as shown in Formula 1, where Z c is the number of first accounts in the first sample data, r ui is the matrix parameter in the first sample matrix, q i is the matrix parameter vector in the first initial matrix, P u is the matrix parameter vector in the first auxiliary matrix, ρ is the penalty coefficient, and m is the number of cluster categories.

[0088] Formula 1:

[0089] The first sample data is substituted into the first loss function, and parameter fitting training is performed on the matrix parameters in the first initial matrix to obtain first matrix parameters. During the parameter fitting training of the matrix parameters in the first initial matrix, the matrix parameters in the first auxiliary matrix do not change, that is, only the matrix parameters in the first initial matrix are updated.

[0090] Step 404: Based on the first loss function, determine an update function of the first initial matrix by a gradient descent method.

[0091] According to the first loss function shown in formula 1, the update function corresponding to the first initial matrix is ​​determined. The update function is used to update the matrix parameters in the first initial matrix. Schematically, the update function of the first initial matrix is ​​obtained by the gradient descent method, that is, the update function determines the parameter to be estimated q in the first initial matrix by taking the partial derivative of the first loss function. i*The update function corresponding to the first initial matrix is ​​obtained by advancing in the opposite direction of the gradient. The component parameters of the update function are the same as those of the first loss function. Schematically, a learning rate parameter can also be introduced into the above update function to control the training speed of the first initial matrix.

[0092] Step 405: Substitute the first sample data into the update function of the first initial matrix, update the matrix parameters in the first initial matrix, and obtain first updated parameters.

[0093] The first sample data corresponds to a first sample matrix. The matrix parameters in the first sample matrix are substituted into the above-mentioned update function to update the matrix parameters in the first initial matrix. The matrix parameters in the updated first initial matrix are composed of the first update parameters.

[0094] Step 406 : In response to a first loss value corresponding to the first update parameter being less than a first numerical requirement, determining the first update parameter as a first matrix parameter.

[0095] A first loss value is obtained by applying a first loss function to the matrix parameters in the updated first initial matrix. The first loss value is used to represent the error between the currently updated first initial matrix and the first relationship matrix to be obtained. A larger loss value indicates a larger error between the first initial matrix and the first relationship matrix to be obtained, indicating that the matrix parameters in the first initial matrix still need to be updated until the first loss value obtained by the first loss function converges. In response to the first loss value corresponding to the first update parameter being less than the first numerical requirement, the first update parameter is determined as the first matrix parameter.

[0096] Step 407: Determine a first relationship matrix including first matrix parameters.

[0097] When the first loss value corresponding to the first update parameter is less than the first numerical requirement, the determined first update parameter is the matrix parameter in the first relationship matrix.

[0098] Since the clustering categories between contents are determined by the intrinsic attributes of the contents, and users' evaluations of the contents are independent of each other, the matrix parameters in the first relationship matrix are not affected by the data volume of the first sample data. The first relationship matrix obtained by this method reduces the data volume of the required first sample data and at the same time reduces the computational complexity of obtaining the first relationship matrix.

[0099] In summary, the first relationship matrix determination method provided in the embodiment of the present application updates the matrix parameters in the first initial matrix according to the gradient descent method for the first sample data corresponding to the first account set and the first loss function, so as to obtain a first relationship matrix that can represent the clustering of the candidate content. In the process of determining the first relationship matrix, the amount of data of the first sample data corresponding to the first account set can be less than the amount of data of the second sample data corresponding to the second account set, thereby reducing the amount of data calculation in the content recommendation process. At the same time, because the first sample data corresponds to shared data, the process of determining the first relationship matrix can be executed in either a shared server or a target platform server.

[0100] Please refer to Figure 5 , which shows a second relationship matrix determination method provided in an embodiment of the present application. The method can be executed by a target platform server or a terminal, and is not limited here. The method includes:

[0101] Step 501: Obtain second sample data corresponding to a second account set.

[0102] The second sample data includes known interest levels between the second account and the candidate content. The second sample data is processed to obtain a second sample matrix.

[0103] Step 502: Obtain a second initial matrix of second accounts and cluster categories.

[0104] The matrix parameters in the second initial matrix can be one of a zero matrix, an identity matrix, a mean matrix, or other matrices with preset matrix parameters, which are not limited here. Schematically, the second initial matrix corresponds to the P matrix in the latent semantic model.

[0105] Step 503: Obtain a second loss function.

[0106] The second loss function is used to determine the interest between the second account and the cluster category expressed by the second initial matrix during the training process, and the error between the interest between the second account and the cluster category corresponding to the second sample data. Schematically, in the embodiment of the present application, the first loss function adopts the mean square error loss function. In order to prevent overfitting during the model training process, a penalty term is added after the mean square error loss function to obtain the second loss function, as shown in Formula 2, where Z u is the number of second accounts in the second sample data, r ui is the matrix parameter in the second sample matrix, q i* is the matrix parameter in the first relation matrix, P u is the matrix parameter vector in the second initial matrix, and ρ is the penalty coefficient.

[0107] Formula 2:

[0108] The second sample data and the first relationship matrix are substituted into the second loss function, and parameter fitting training is performed on the matrix parameters in the second initial matrix to obtain second matrix parameters. During the parameter fitting training process for the matrix parameters in the second initial matrix, the matrix parameters in the first relationship matrix do not change, that is, only the matrix parameters in the second initial matrix are updated.

[0109] Step 504 : Based on the second loss function and the first relationship matrix, determine an update function of the second initial matrix by a gradient descent method.

[0110] According to the second loss function shown in formula 2, the update function corresponding to the second initial matrix is ​​determined. The update function is used to update the matrix parameters in the second initial matrix. Schematically, the update function of the second initial matrix is ​​obtained by the gradient descent method, that is, the update function determines the parameter to be estimated P in the second initial matrix by taking the partial derivative of the second loss function. u The update function corresponding to the second initial matrix is ​​obtained by advancing in the opposite direction of the gradient. The component parameters of the update function are the same as those of the second loss function. Schematically, a learning rate parameter can also be introduced into the above update function to control the training speed of the first initial matrix.

[0111] Step 505: Input the second sample data into the update function of the second initial matrix, update the matrix parameters in the second initial matrix, and obtain second updated parameters.

[0112] The second sample data corresponds to a second sample matrix. The matrix parameters in the second sample matrix are substituted into the above-mentioned update function to update the matrix parameters in the second initial matrix. The matrix parameters in the updated second initial matrix are composed of the second update parameters.

[0113] Step 506: In response to the second loss value corresponding to the second update parameter being less than the second numerical requirement, the second update parameter is determined as a second matrix parameter.

[0114] The matrix parameters in the updated second initial matrix are applied to the second loss function to obtain a second loss value. The second loss value is used to represent the error between the currently updated second initial matrix and the second relationship matrix to be obtained. The larger the loss value, the greater the error between the second initial matrix and the second relationship matrix to be obtained, indicating that the matrix parameters in the second initial matrix still need to be updated until the second loss value obtained by the second loss function converges. In response to the second loss value corresponding to the second update parameter being less than the second numerical requirement, the second update parameter is determined as the second matrix parameter.

[0115] Step 507: Determine a second relationship matrix including second matrix parameters.

[0116] When the second loss value corresponding to the second update parameter is less than the second numerical requirement, the determined second update parameter is the matrix parameter in the second relationship matrix.

[0117] The training process of the first relationship matrix can be completed in a shared server or in a target platform server, and the training process of the second relationship matrix can be completed in a target platform server or in a terminal. Schematically, if the training process of the second relationship matrix is ​​completed in the target platform server, the second sample data is the known interest level between the second account and the candidate content in the second account set corresponding to the target platform; if the training process of the second relationship matrix is ​​completed in the terminal, the second sample data is the known interest level between the second account corresponding to the terminal and the candidate content, that is, the terminal obtains the trained first relationship matrix from the target platform server, and then determines the second sample data based on the relevant data of the second account logged in by the target application running in the terminal, and obtains the second relationship matrix based on the second sample data and the first relationship matrix training, and the recommendation value of the candidate content relative to the second account corresponding to the terminal is calculated jointly by the first relationship matrix and the second relationship matrix.

[0118] In summary, the second relationship matrix method provided in the embodiment of the present application updates the matrix parameters in the second initial matrix according to the gradient descent method based on the second sample data corresponding to the second account set and the second loss function, so as to obtain a second relationship matrix that can represent the interest level of the second account and the cluster category. In the process of determining the second relationship matrix, since the second sample data corresponding to the second account set will not be transmitted to other servers or terminals, the corresponding second relationship matrix is ​​determined only in the current server or terminal, thereby protecting the security of user data and improving the security of account data in the content recommendation process.

[0119] Please refer to Figure 6 , which shows a flow chart of a content recommendation method provided by an embodiment of the present application. Since user preferences for content are independent of each other, and users' ratings of content are subjective, and the positioning of recommendation scenarios and the differences in recommended content will lead to differences in the overall rating distribution, in this embodiment of the present application, a global influence parameter is introduced to eliminate the above influence. The method includes:

[0120] Step 601: Obtain global impact parameters.

[0121] Schematically, the global impact parameter is determined by the mean of the first sample data, as shown in Formula 3, where μ *is the global average of the first sample matrix corresponding to the first sample data, Z c is the number of first accounts in the first account set, r ui is the matrix parameter in the first sample matrix.

[0122] Formula 3:

[0123] Step 602: Obtain a first relationship matrix.

[0124] In the embodiment of the present application, the process of determining the first relationship matrix is ​​the same as steps 401 to 407 and will not be described in detail here.

[0125] Step 603: Obtain a second relationship matrix.

[0126] In the embodiment of the present application, the process of determining the second relationship matrix is ​​the same as steps 501 to 507 and will not be described in detail here.

[0127] Step 604: Modify the third relationship matrix using the global influence parameter to determine the recommendation value of the candidate content relative to the second account.

[0128] In an embodiment of the present application, the inner product of the first relationship matrix and the second relationship matrix is ​​calculated to obtain a third relationship matrix, and the reference recommendation value of the candidate content relative to the second account is determined based on the third matrix parameters in the third relationship matrix; the reference recommendation value and the global influence parameter are summed to obtain the recommendation value of the candidate content relative to the second account.

[0129] Indicatively, the formula for calculating the recommended value is shown in Formula 4, where: is the recommendation value of candidate content i relative to the second account u, μ * is the global influencing parameter, P u* is the matrix parameter vector in the second relation matrix, q i* is the matrix parameter vector in the first relation matrix.

[0130] Formula 4:

[0131] Step 605 : In response to the recommendation value reaching the preset recommendation value, recommending candidate content to the second account.

[0132] After the recommendation value of the candidate content relative to the second account is predicted by the above formula 4, whether to push the candidate content to the second account is determined based on the obtained recommendation value. If the recommendation value corresponding to the candidate content reaches the preset recommendation value, it means that the user of the second account is more likely to be interested in the candidate content, that is, the candidate content is recommended to the second account. If the recommendation value corresponding to the candidate content does not reach the preset recommendation value, it means that the user of the second account is less likely to be interested in the candidate content, that is, the candidate content will not be pushed to the second account.

[0133] Schematically, in the embodiment of the present application, the content recommendation method is applied to the MovieLens dataset and the Netflix dataset for testing, wherein a preset percentage (e.g., 20%) of the data in the dataset is obtained as the first sample data by random sampling or user-based sampling, and the first relationship matrix and the global influence parameter are trained. The remaining data in the dataset is used as the second sample data to train the second relationship matrix. Figure 7 , which shows the mean square error histogram 700 corresponding to five algorithms: USST-SVD-Rnd (a recommendation algorithm implemented by random sampling of a data set in the content recommendation method provided in the embodiment of the present application), USST-SVD-UB (a recommendation algorithm implemented by user sampling of a data set in the content recommendation method provided in the embodiment of the present application), SVD (latent semantic model) 703, NMBP (multi-layer feedforward network) 704, and PRIBCF (collaborative filtering) 705, where the ordinate is the mean square error. Please refer to Figure 8 , which shows a mean square error reduction graph 800 of USST-SVD-Rnd801, USST-SVD-UB802 and SVD803 when the MovieLens dataset is used as sample data, wherein the horizontal axis represents the preset percentage of the first sample data in the total sample data, and the vertical axis represents the mean square error. Please refer to Figure 9 , which shows a mean square error reduction graph 900 of USST-SVD-Rnd901, USST-SVD-UB902 and SVD903 when the Netflix dataset is used as sample data, where the horizontal axis represents the preset percentage of the first sample data in the total sample data, and the vertical axis represents the mean square error.

[0134] from Figure 7 、 Figure 8 and Figure 9It can be seen that USST-SVD-Rnd can surpass the NMBP algorithm and PRIBCF algorithm trained on the full data, while USST-SVD-UB can approach the SVD model trained on the full data. In other words, the content recommendation method provided by the embodiment of the application can reduce the computational complexity of the content recommendation process while ensuring the accuracy of candidate content recommendations.

[0135] To sum up, the content recommendation method provided in the embodiment of the present application eliminates the differences caused by the distribution of sample data through global influence parameters, determines the recommendation value of the candidate content relative to the second account based on the first relationship matrix, the second relationship matrix and the global influence parameters, and recommends the candidate content that reaches the preset recommendation value to the second account, thereby improving the accuracy of content recommendation.

[0136] Please refer to Figure 10 , which shows a content recommendation system provided by an embodiment of the present application. Schematically, the content recommendation system is set in a target platform server as an example for description. The system includes:

[0137] The data collection module 1010 is used to obtain first sample data 1011 corresponding to the first account set, where the first sample data is used to represent the first account's interest in the candidate content to be recommended; send the first sample data 1012 to the data processing module 1020; obtain second sample data 1013 corresponding to the second account set, where the second account set includes the second account corresponding to the target platform, and the second sample data includes the known interest between the second account and the candidate content; and send the second sample data 1014 to the data processing module 1020.

[0138] Illustratively, a platform cluster including the target platform obtains user content behavior preference data through voluntary or paid user contributions, pre-processes the behavior preference data, and obtains first sample data. Alternatively, the target platform collects user data publicly available on internet platforms and processes the data to obtain the first sample data.

[0139] Illustratively, the target platform server obtains the behavior preference data corresponding to the internal account of the platform from the database, and preprocesses the behavior preference data to obtain the second sample data.

[0140] The data processing module 1020 is used to process the first sample data to obtain a first sample matrix 1021; send the first sample matrix 1022 to the matrix training module 1030; process the second sample data to obtain a second sample matrix 1023; and send the second sample matrix 1024 to the matrix training module 1030.

[0141] Schematically, the data processing module 1020, such as Figure 11As shown, taking the acquisition of the first sample matrix as an example, the data collection module 1110 inputs the first sample data into the data processing module 1120. The data processing module 1120 first determines whether the first account has a rating operation on the candidate content 1121. If so, the specific rating is filled in the first sample matrix 1122; if not, it determines whether the first account has a browsing operation on the candidate content 1123. If so, 1 is filled in the first sample matrix 1124; if not, 0 is filled in the first sample matrix 1125; and the first sample matrix 1126 is generated.

[0142] The matrix training module 1030 is used to cluster the first sample data to obtain a first relationship matrix 1031, which is used to represent the clustering of the candidate content; perform matrix decomposition processing on the second sample matrix corresponding to the second sample data to obtain a second relationship matrix 1032 corresponding to the first relationship matrix, which is used to represent the interest level between the second account and the cluster category; and send the first relationship matrix and the second relationship matrix 1033 to the prediction module 1040.

[0143] Optionally, the matrix training module 1030 is also used to obtain a first initial matrix of candidate content and cluster categories; perform parameter fitting training on the matrix parameters in the first initial matrix in combination with the first sample data to obtain first matrix parameters for representing the clustering situation of the candidate content; and determine a first relationship matrix containing the first matrix parameters.

[0144] Optionally, the matrix training module 1030 is also used to obtain a first loss function, which is used to determine the clustering of candidate content expressed by the first initial matrix during the training process and the error between the clustering of candidate content corresponding to the first sample data; the first sample data is substituted into the first loss function, and parameter fitting training is performed on the matrix parameters in the first initial matrix to obtain the first matrix parameters.

[0145] Optionally, the matrix training module 1030 is also used to determine the update function of the first initial matrix by the gradient descent method based on the first loss function; substitute the first sample data into the update function of the first initial matrix, update the matrix parameters in the first initial matrix, and obtain the first update parameters; in response to the first loss value corresponding to the first update parameter being less than the first numerical requirement, determine the first update parameter as the first matrix parameter.

[0146] Optionally, the matrix training module 1030 is also used to obtain a second initial matrix of the second account and the cluster category; perform parameter fitting training on the matrix parameters in the second initial matrix in combination with the second sample matrix and the first relationship matrix to obtain second matrix parameters for representing the interest level between the second account and the cluster category; and determine a second relationship matrix containing the second matrix parameters.

[0147] Optionally, the matrix training module 1030 is also used to obtain a second loss function, which is used to determine the interest level between the second account and the cluster category expressed by the second initial matrix during the training process, and the error between the interest level between the second account and the cluster category corresponding to the second sample data; the second sample data and the first relationship matrix are substituted into the second loss function, and parameter fitting training is performed on the matrix parameters in the second initial matrix to obtain the second matrix parameters.

[0148] Optionally, the matrix training module 1030 is also used to determine the update function of the second initial matrix by the gradient descent method based on the second loss function and the first relationship matrix; input the second sample data into the update function of the second initial matrix, update the matrix parameters in the second initial matrix, and obtain second update parameters; in response to the second loss value corresponding to the second update parameter being less than the second numerical requirement, determine the second update parameter as the second matrix parameter.

[0149] The prediction module 1040 is configured to determine, from a third relationship matrix determined by the first relationship matrix and the second relationship matrix, candidate content 1041 to be recommended to the second account.

[0150] Optionally, the prediction module 1040 is also used to obtain a global influence parameter, which is determined by the mean of the first sample data; correct the third relationship matrix through the global influence parameter to determine the recommendation value of the candidate content relative to the second account; and recommend the candidate content to the second account in response to the recommendation value reaching a preset recommendation value.

[0151] Optionally, the prediction module 1040 is also used to calculate the inner product of the first relationship matrix and the second relationship matrix to obtain a third relationship matrix; determine the reference recommendation value of the candidate content relative to the second account based on the third matrix parameters in the third relationship matrix; sum the reference recommendation value and the global influence parameter to obtain the recommendation value of the candidate content relative to the second account.

[0152] Optionally, the prediction module 1040 is further configured to receive a content recommendation request for the second account, where the content recommendation request is configured to request that candidate content that meets the interests of the second account be sent to a terminal corresponding to the second account.

[0153] In summary, the content recommendation system provided in the embodiment of the present application determines a first relationship matrix using first sample data corresponding to a first account set, and determines a second relationship matrix using second sample data corresponding to a second account set, in order to obtain the unknown interest level between the second account and the candidate content. The first relationship matrix represents the clustering of the candidate content, and the second relationship matrix represents the interest level between the second account and the cluster category. Based on the above-mentioned first and second relationship matrices, the interest level between the second account and the candidate content can be determined, thereby recommending candidate content that meets the corresponding interest level of the second account to the second account. Since the clustering of the candidate content is an objective attribute of the candidate content, the first relationship matrix can be determined using a smaller amount of first sample data, reducing the amount of data calculation in the content recommendation process.

[0154] Please refer to Figure 12 , is a structural block diagram of a content recommendation device provided by an exemplary embodiment of the present application, the device comprising:

[0155] An acquisition module 1210 is configured to acquire first sample data corresponding to a first account set, where the first sample data is used to indicate the first account's interest in the candidate content to be recommended;

[0156] A first determining module 1220 is configured to perform clustering processing on the first sample data to obtain a first relationship matrix, where the first relationship matrix is ​​used to represent the clustering of the candidate content;

[0157] The acquisition module 1210 is further configured to acquire second sample data corresponding to a second account set, wherein the second account set includes second accounts corresponding to the target platform, and the second sample data includes known interest levels between the second accounts and the candidate content;

[0158] A second determining module 1230 is configured to perform matrix decomposition processing on a second sample matrix corresponding to the second sample data to obtain a second relationship matrix corresponding to the first relationship matrix, where the second relationship matrix is ​​used to represent the interest level between the second account and the cluster category;

[0159] The recommendation module 1240 is configured to determine, from a third relationship matrix determined by the first relationship matrix and the second relationship matrix, to recommend the candidate content to the second account.

[0160] In an alternative embodiment, please refer to Figure 13 The first determining module 1220 further includes:

[0161] A first acquiring unit 1221 is configured to acquire a first initial matrix of the candidate content and the cluster categories;

[0162] A first training unit 1222 is configured to perform parameter fitting training on the matrix parameters in the first initial matrix in combination with the first sample data to obtain first matrix parameters for representing the clustering of the candidate content;

[0163] The first determining unit 1223 is configured to determine the first relationship matrix including the first matrix parameters.

[0164] In an optional embodiment, the first acquiring unit 1221 is further configured to acquire a first loss function, wherein the first loss function is used to determine an error between the clustering of the candidate content expressed by the first initial matrix and the clustering of the candidate content corresponding to the first sample data during training;

[0165] The first training unit 1222 is further configured to substitute the first sample data into the first loss function, perform parameter fitting training on the matrix parameters in the first initial matrix, and obtain the first matrix parameters.

[0166] In an optional embodiment, the first determining unit 1223 is further configured to determine an update function of the first initial matrix by a gradient descent method based on the first loss function;

[0167] The first training unit 1222 is further configured to substitute the first sample data into an update function of the first initial matrix, update matrix parameters in the first initial matrix, and obtain first updated parameters;

[0168] The first determining unit 1223 is further configured to determine the first update parameter as the first matrix parameter in response to a first loss value corresponding to the first update parameter being less than a first numerical requirement.

[0169] In an optional embodiment, the second determining module further includes:

[0170] A second acquiring unit 1231 is configured to acquire a second initial matrix of the second account and the cluster category;

[0171] A second training unit 1232 is configured to perform parameter fitting training on the matrix parameters in the second initial matrix in combination with the second sample matrix and the first relationship matrix to obtain second matrix parameters for representing the interest level between the second account and the cluster category;

[0172] The second determining unit 1233 is configured to determine the second relationship matrix including the second matrix parameters.

[0173] In an optional embodiment, the second acquisition unit 1231 is further configured to acquire a second loss function, wherein the second loss function is configured to determine an error between the interest level between the second account and the cluster category expressed by the second initial matrix during training and the interest level between the second account and the cluster category corresponding to the second sample data;

[0174] The second training unit 1232 is further configured to substitute the second sample data and the first relationship matrix into the second loss function, perform parameter fitting training on the matrix parameters in the second initial matrix, and obtain the second matrix parameters.

[0175] In an optional embodiment, the second determining unit 1233 is further configured to determine an update function of the second initial matrix by a gradient descent method based on the second loss function and the first relationship matrix;

[0176] The second training unit 1232 is further configured to input the second sample data into an update function of the second initial matrix, update matrix parameters in the second initial matrix, and obtain second updated parameters;

[0177] The second determining unit 1233 is further configured to determine the second update parameter as the second matrix parameter in response to a second loss value corresponding to the second update parameter being less than a second numerical requirement.

[0178] In an optional embodiment, the acquisition module 1210 is further configured to acquire a global impact parameter, where the global impact parameter is determined by a mean value of the first sample data;

[0179] The recommendation module 1240 further includes:

[0180] a determining unit 1241 configured to modify the third relationship matrix using the global influence parameter to determine a recommendation value of the candidate content relative to the second account;

[0181] The recommendation unit 1242 is configured to recommend the candidate content to the second account in response to the recommendation value reaching a preset recommendation value.

[0182] In an optional embodiment, the recommendation module 1240 further includes:

[0183] a calculating unit 1243, configured to calculate the inner product of the first relationship matrix and the second relationship matrix to obtain a third relationship matrix;

[0184] The determining unit 1241 is further configured to determine a reference recommendation value of the candidate content relative to the second account based on a third matrix parameter in the third relationship matrix;

[0185] The determining unit 1241 is further configured to sum the reference recommendation value and the global influence parameter to obtain a recommendation value of the candidate content relative to the second account.

[0186] In an optional embodiment, the recommendation module 1240 further includes:

[0187] The receiving unit 1244 is configured to receive a content recommendation request for the second account, where the content recommendation request is used to request that the candidate content that meets the interests of the second account be sent to the terminal corresponding to the second account.

[0188] In summary, the content recommendation device provided in the embodiments of the present application determines a first relationship matrix based on first sample data corresponding to a first account set and a second relationship matrix based on second sample data corresponding to a second account set to obtain unknown interest between a second account and candidate content. The training processes of the first and second relationship matrices are independent of each other. The first relationship matrix represents the clustering of candidate content. Since the clustering of candidate content is an objective attribute of the candidate content, the training of the first relationship matrix does not need to be completed using the full data set. Therefore, the training process of the first relationship matrix can also be completed independently of the target platform. That is, the target platform can obtain the trained first relationship matrix from another platform and then obtain the second relationship matrix based on the first relationship matrix and the second sample data of the second account set corresponding to the target platform. The second relationship matrix represents the interest between the second account and the cluster category. The target platform can then determine the interest between the second account and the candidate content based on the first and second relationship matrices, thereby recommending candidate content that meets the corresponding interest level of the second account to the second account. Since the training process of the first relationship matrix can be performed independently of the target platform, the possibility of user data leakage during the content recommendation process is reduced, and the data security during the content recommendation process is improved.

[0189] It should be noted that the content recommendation device provided in the above embodiment is merely an example of the division of the aforementioned functional modules. In actual applications, the aforementioned functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the content recommendation device provided in the above embodiment and the content recommendation method embodiment are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0190] Figure 14 The following is a schematic diagram showing the structure of a server provided by an exemplary embodiment of the present application. Specifically:

[0191] Server 1400 includes a central processing unit (CPU) 1401, a system memory 1404 including a random access memory (RAM) 1402 and a read-only memory (ROM) 1403, and a system bus 1405 connecting system memory 1404 and CPU 1401. Server 1400 also includes a mass storage device 1406 for storing an operating system 1413, application programs 1414, and other program modules 1415.

[0192] The mass storage device 1406 is connected to the central processing unit 1401 through a mass storage controller (not shown) connected to the system bus 1405. The mass storage device 1406 and its associated computer-readable media provide non-volatile storage for the server 1400. In other words, the mass storage device 1406 may include a computer-readable medium (not shown) such as a hard disk or a compact disc read-only memory (CD-ROM) drive.

[0193] Without loss of generality, computer-readable media may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules or other data. Computer storage media include RAM, ROM, Erasable Programmable Read Only Memory (EPROM), Electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other solid-state storage technologies, CD-ROM, Digital Versatile Disc (DVD) or other optical storage, tape cassettes, magnetic tape, disk storage or other magnetic storage devices. Of course, those skilled in the art will appreciate that computer storage media are not limited to the above-mentioned ones. The above-mentioned system memory 1404 and mass storage device 1406 can be collectively referred to as memory.

[0194] According to various embodiments of the present application, the server 1400 may also be connected to a remote computer on a network such as the Internet for operation. That is, the server 1400 may be connected to the network 1412 via the network interface unit 1411 connected to the system bus 1405, or the network interface unit 1411 may be used to connect to other types of networks or remote computer systems (not shown).

[0195] The memory also includes one or more programs, which are stored in the memory and configured to be executed by the CPU.

[0196] Embodiments of the present application also provide a computer device comprising a processor and memory, wherein the memory stores at least one instruction, at least one program, code set, or instruction set, and the at least one instruction, at least one program, code set, or instruction set is loaded and executed by the processor to implement the content recommendation methods provided in the above-described method embodiments. Optionally, the computer device may be a terminal or a server.

[0197] An embodiment of the present application also provides a computer-readable storage medium, which stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by a processor to implement the content recommendation method provided by the above-mentioned method embodiments.

[0198] Embodiments of the present application also provide a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the content recommendation method described in any of the above embodiments.

[0199] Optionally, the computer-readable storage medium may include: a read-only memory (ROM), a random access memory (RAM), a solid-state drive (SSD), or an optical disk. Among them, the random access memory may include a resistance random access memory (ReRAM) and a dynamic random access memory (DRAM). The serial numbers of the above embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0200] Those skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or by a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, which may be a read-only memory, a disk, or an optical disk, etc.

[0201] The above description is merely an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A content recommendation method, characterized in that: The method comprises: Obtaining first sample data corresponding to the first account set, where the first sample data is used to represent the first account's interest in the candidate content to be recommended; Performing clustering processing on the first sample data to obtain a first relationship matrix, where the first relationship matrix is ​​used to represent the clustering of the candidate content, and a first matrix parameter in the first relationship matrix represents the weight of the candidate content in the cluster category; Obtaining second sample data corresponding to a second account set, where the second account set includes second accounts corresponding to the target platform, and the second sample data includes known interest levels between the second accounts and the candidate content; Obtain a second sample matrix corresponding to the second sample data; obtain a second initial matrix of the second account number and the cluster category; Obtaining a second loss function, where the second loss function is used to determine the interest level between the second account and the cluster category expressed by the second initial matrix during training, and an error between the interest level between the second account and the cluster category corresponding to the second sample data; Substituting the second sample data and the first relationship matrix into the second loss function, and performing parameter fitting training on the matrix parameters in the second initial matrix while keeping the matrix parameters in the first relationship matrix unchanged, to obtain second matrix parameters for representing the interest level between the second account and the cluster category; Determining a second relationship matrix including the second matrix parameters, where the second relationship matrix is ​​used to represent the interest level between the second account and the cluster category; Determine to recommend the candidate content to the second account from a third relationship matrix determined from the first relationship matrix and the second relationship matrix.

2. The method according to claim 1, characterized in that The clustering process is performed on the first sample data to obtain a first relationship matrix, including: Obtaining a first initial matrix of the candidate content and the cluster categories; performing parameter fitting training on matrix parameters in the first initial matrix in combination with the first sample data to obtain first matrix parameters for representing the clustering of the candidate content; The first relationship matrix including the first matrix parameters is determined.

3. The method according to claim 2, characterized in that The performing parameter fitting training on the matrix parameters in the first initial matrix in combination with the first sample data to obtain first matrix parameters for representing the clustering of the candidate content includes: Obtaining a first loss function, where the first loss function is used to determine an error between a clustering of the candidate content expressed by the first initial matrix during training and a clustering of the candidate content corresponding to the first sample data; Substituting the first sample data into the first loss function, performing parameter fitting training on the matrix parameters in the first initial matrix, and obtaining the first matrix parameters.

4. The method according to claim 3, characterized in that Substituting the first sample data into the first loss function and performing parameter fitting training on the matrix parameters in the first initial matrix to obtain the first matrix parameters includes: Based on the first loss function, determining an update function of the first initial matrix by a gradient descent method; Substituting the first sample data into the update function of the first initial matrix, updating the matrix parameters in the first initial matrix, and obtaining first updated parameters; In response to a first loss value corresponding to the first update parameter being less than a first numerical requirement, the first update parameter is determined as the first matrix parameter.

5. The method according to claim 1, wherein Substituting the second sample data and the first relationship matrix into the second loss function, and performing parameter fitting training on the matrix parameters in the second initial matrix while the matrix parameters in the first relationship matrix remain unchanged, to obtain second matrix parameters for representing the interest level between the second account and the cluster category, includes: Determining an update function of the second initial matrix by a gradient descent method based on the second loss function and the first relationship matrix; Inputting the second sample data into an update function of the second initial matrix, updating matrix parameters in the second initial matrix, and obtaining second updated parameters; In response to a second loss value corresponding to the second update parameter being less than a second numerical requirement, the second update parameter is determined as the second matrix parameter.

6. The method according to any one of claims 1 to 5, characterized in that: The method further comprises: Obtaining a global impact parameter, where the global impact parameter is determined by a mean value of the first sample data; The step of determining, from a third relationship matrix determined from the first relationship matrix and the second relationship matrix, to recommend the candidate content to the second account includes: Modifying the third relationship matrix using the global influence parameter to determine a recommendation value of the candidate content relative to the second account; In response to the recommendation value reaching a preset recommendation value, the candidate content is recommended to the second account.

7. The method according to claim 6, characterized in that The modifying the third relationship matrix by using the global influence parameter to determine the recommendation value of the candidate content relative to the second account includes: Calculating the inner product of the first relationship matrix and the second relationship matrix to obtain the third relationship matrix; determining a reference recommendation value of the candidate content relative to the second account based on a third matrix parameter in the third relationship matrix; The reference recommendation value and the global influence parameter are summed to obtain a recommendation value of the candidate content relative to the second account.

8. The method according to any one of claims 1 to 5, characterized in that: Before determining to recommend the candidate content to the second account in a third relationship matrix determined from the first relationship matrix and the second relationship matrix, the method further includes: A content recommendation request for the second account is received, where the content recommendation request is used to request that the candidate content that meets the interests of the second account be sent to a terminal corresponding to the second account.

9. A content recommendation device, characterized in that: The device comprises: An acquisition module, configured to acquire first sample data corresponding to a first account set, wherein the first sample data is used to represent the first account's interest in the candidate content to be recommended; a first determining module, configured to perform clustering processing on the first sample data to obtain a first relationship matrix, wherein the first relationship matrix is ​​used to represent the clustering of the candidate content, and a first matrix parameter in the first relationship matrix represents the weight of the candidate content in the cluster category; The acquisition module is further configured to acquire second sample data corresponding to a second account set, the second account set including second accounts corresponding to the target platform, the second sample data including known interest levels between the second accounts and the candidate content; A second determination module is configured to obtain a second sample matrix corresponding to the second sample data; obtain a second initial matrix of the second account and the cluster category; obtain a second loss function, wherein the second loss function is configured to determine the error between the interest level between the second account and the cluster category expressed by the second initial matrix during training and the interest level between the second account and the cluster category corresponding to the second sample data; substitute the second sample data and the first relationship matrix into the second loss function, and perform parameter fitting training on the matrix parameters in the second initial matrix without changing the matrix parameters in the first relationship matrix to obtain second matrix parameters for representing the interest level between the second account and the cluster category; determine a second relationship matrix including the second matrix parameters, wherein the second relationship matrix is ​​configured to represent the interest level between the second account and the cluster category; A recommendation module is configured to determine, from a third relationship matrix determined by the first relationship matrix and the second relationship matrix, to recommend the candidate content to the second account.

10. The device according to claim 9, characterized in that The first determining module further includes: A first acquiring unit, configured to acquire a first initial matrix of the candidate content and the cluster categories; a first training unit, configured to perform parameter fitting training on matrix parameters in the first initial matrix in combination with the first sample data, to obtain first matrix parameters for representing clustering of the candidate content; The first determining unit is configured to determine the first relationship matrix including the first matrix parameters.

11. A computer device, characterized in that: The computer device includes a processor and a memory, wherein the memory stores at least one program, and the at least one program is loaded and executed by the processor to implement the content recommendation method according to any one of claims 1 to 8.

12. A computer-readable storage medium, characterized in that The computer-readable storage medium stores at least one program code, and the program code is loaded and executed by a processor to implement the content recommendation method according to any one of claims 1 to 8.

13. A computer program product, characterized in that The method comprises computer instructions, which, when executed by a processor, implements the content recommendation method according to any one of claims 1 to 8.

Images