A content recommendation method, communication device and computer-readable storage medium

Abstract

The present invention relates to the field of recommendation technology, and provides a content recommendation method, a communication device, and a computer-readable storage medium to solve the problem of poor recommendation accuracy. The method includes: obtaining a user behavior recognition result when a terminal starts an application and / or logs into an application through an account, the user behavior recognition result is a recognition result obtained by pre-recognizing user behavior data through a target recognition model, the user behavior recognition result includes multi-user behavior or single-user behavior, the target recognition model is trained by a training sample behavior feature set, and content recommendation is performed based on the user behavior recognition result. The user behavior data of the terminal and / or account can be identified by the target recognition model to determine whether it is a single-user behavior or a multi-user behavior, and the target recognition model is trained by S sample behavior features of multi-user behavior and T sample behavior features of single-user behavior, and the user behavior recognition result is used to recommend content, which can improve the accuracy of recommendation.

Description

Technical Field

[0001] The present invention relates to the field of recommendation technology, and in particular to a content recommendation method, a communication device and a computer-readable storage medium. Background Art

[0002] With the development of Internet technology, a large number of applications can provide users with personalized recommendation services and content that matches their behavioral preferences, etc., in order to increase user stickiness and activity.

[0003] At present, in the process of content recommendation, a common method is to obtain user preferences by analyzing the user's behavior on the terminal or the user's behavior through a certain account of the application, and provide recommended content to the user. However, in this way, the behavior on the terminal or the behavior through a certain account of the application is counted as the behavior of the same user. For example, multiple users in a family share a terminal, and the resulting behavior is multi-user behavior, which is counted as the behavior of the same user. For another example, the same account can be logged in on different terminals, that is, different users can share an account, and the resulting behavior is multi-user behavior, which is counted as the behavior of the same user. In this way, content recommendation through the above common methods is likely to lead to poor accuracy of content recommendation. Summary of the invention

[0004] Embodiments of the present invention provide a content recommendation method, a communication device, and a computer-readable storage medium to solve the problem of poor accuracy in content recommendation.

[0005] In order to solve the above-mentioned technical problems, the present invention is achieved as follows:

[0006] In a first aspect, an embodiment of the present invention provides a content recommendation method, comprising:

[0007] When the terminal starts the application and / or logs in to the application through the account, a user behavior recognition result is obtained, wherein the user behavior recognition result is a recognition result obtained by pre-recognizing the user behavior data through the target recognition model, wherein the user behavior recognition result includes multi-user behavior or single-user behavior, wherein the user behavior data is associated with the terminal and / or the account, wherein the target recognition model is obtained by training a training sample behavior feature set, wherein the training sample behavior feature set includes S sample behavior features whose behaviors are marked as multi-user behaviors and T sample behavior features whose behaviors are marked as single-user behaviors, wherein both S and T are positive integers;

[0008] Content recommendation is performed based on the user behavior recognition result.

[0009] In a second aspect, an embodiment of the present invention further provides a communication device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps in the content recommendation method as described above when executing the computer program.

[0010] In a third aspect, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps in the content recommendation method described above are implemented.

[0011] In the content recommendation method of the present embodiment, the user behavior data of the terminal and / or account can be identified through the target recognition model to obtain the user behavior recognition result, that is, whether it is a single-user behavior or a multi-user behavior, and the target recognition model is trained by S sample behavior features of multi-user behavior and T sample behavior features of single-user behavior, and the obtained user behavior recognition result can be used for content recommendation, that is, in the recommendation process, the user behavior recognition result of the terminal and / or account is taken into account, and the user behavior recognition result is obtained by performing user behavior recognition on the user behavior data of the terminal and / or account using the target recognition model trained by S sample behavior features of multi-user behavior and T sample behavior features of single-user behavior. In this way, the accuracy of content recommendation can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0013] Figure 1 This is one of the flow charts of the content recommendation method provided by an embodiment of the present invention;

[0014] Figure 2 This is the second flowchart of the content recommendation method provided by the embodiment of the present invention;

[0015] Figure 3 is a module diagram of a content recommendation system for implementing a content recommendation method provided by an embodiment of the present invention;

[0016] Figure 4 It is a structural diagram of a communication device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] See also Figure 1 , Figure 1 is a flow chart of a content recommendation method provided by an embodiment of the present invention. Figure 1 As shown, the following steps are included:

[0019] Step 101: When the terminal starts the application and / or logs into the application through an account, obtain the user behavior recognition result.

[0020] The user behavior recognition result is the recognition result obtained by pre-recognizing the user behavior data through the target recognition model. The user behavior recognition result includes multi-user behavior or single-user behavior. The user behavior data is associated with the terminal and / or account. The target recognition model is trained by the training sample behavior feature set. The training sample behavior feature set includes S sample behavior features with the behavior marked as multi-user behavior and T sample behavior features with the behavior marked as single-user behavior. S and T are both positive integers.

[0021] It should be noted that the user who generates behavior under the same terminal or the same account may be a (single) user, that is, the terminal or account is used by the same user, and the user who generates behavior under the same terminal or the same account may not be the same user, and there may be multiple users, that is, multiple users share a terminal or share an account, and perform actions on related content in the application, such as click actions, collection actions, etc. This embodiment does not limit the generation of behavior data. It can be understood that the user behavior data of this embodiment is behavior data related to the application. The user behavior data associated with the above terminal and / or the above account may include at least one behavior content, that is, content that has performed related actions (for example, click actions, etc.), such as items (which may include commodities, etc.), videos, books, news, etc.

[0022] In this embodiment, user behavior data associated with the terminal and / or account can be obtained in advance, and user behavior recognition can be performed on the user behavior data based on the target recognition model to obtain a user behavior recognition result, which can identify whether the user behavior data is single-user behavior or multi-user behavior. It can also be understood as identifying the user behavior category of the terminal and / or account, that is, identifying whether the terminal and / or account is a terminal and / or account with multi-user behavior or a terminal and / or account with single-user behavior.

[0023] It should be noted that the target recognition model is a trained model, which is a model pre-trained through a training sample behavior feature set. The training sample behavior feature set includes sample behavior features of two different behavior categories, namely, S sample behavior features of multi-user behavior and T sample behavior features of single-user behavior. In this way, the user behavior recognition performance of the target recognition model can be improved, and the accuracy of the obtained user behavior recognition results can be improved by performing behavior recognition on user behavior data. When the terminal starts the application and / or logs in to the above application through an account, the user behavior recognition result obtained by the target recognition model is first obtained.

[0024] Step 102: Recommend content based on the user behavior recognition result.

[0025] Since the user using the above-mentioned terminal and account may be a single user or multiple users, in this embodiment, the user behavior identification result of the user behavior data associated with the terminal and / or account can be obtained, and the user behavior identification result can be used to perform content recommendation, that is, the user behavior identification result is taken into consideration in the content recommendation process, and it can be understood that whether it is a single user behavior or multiple user behaviors is taken into consideration. Different content recommendations can be implemented for different user behavior identification results, which can improve the accuracy of content recommendation.

[0026] In the content recommendation method of the present embodiment, the user behavior data of the terminal and / or account can be identified through the target recognition model to obtain the user behavior recognition result, that is, whether it is a single-user behavior or a multi-user behavior, and the target recognition model is trained by S sample behavior features of multi-user behavior and T sample behavior features of single-user behavior, and the obtained user behavior recognition result can be used for content recommendation, that is, in the recommendation process, the user behavior recognition result of the terminal and / or account is taken into account, and the user behavior recognition result is obtained by performing user behavior recognition on the user behavior data of the terminal and / or account using the target recognition model trained by S sample behavior features of multi-user behavior and T sample behavior features of single-user behavior. In this way, the accuracy of content recommendation can be improved.

[0027] In one embodiment, content recommendation based on user behavior recognition results includes:

[0028] When the user behavior recognition result is a multi-user behavior, the first content is output according to the behavior portraits of K users obtained in advance, wherein the behavior portraits of the K users are constructed in advance according to K types of behavior data, and the K types of behavior data are obtained by clustering the user behavior data in advance according to the user behavior recognition result, and K is an integer greater than 1;

[0029] When receiving a first input of a target content in the first content, determining a target user from the K users according to the target content and the K types of behavior data;

[0030] Obtain the behavior profile of the target user from the behavior profiles of K users;

[0031] Recommend the second content based on the behavioral profile of the target user.

[0032] In the case where the user behavior recognition result is a multi-user behavior, it means that there are multiple users using the above-mentioned terminal and / or account. Due to the differences between users, the behaviors of different users are different. Therefore, in the content recommendation process under the multi-user behavior, the first content can be output based on the behavior portraits of K users, that is, presented to the user who currently uses the terminal and / or currently uses the account to log in to the application. It should be noted that the user behavior data associated with the above-mentioned terminal and / or the above-mentioned account can be clustered in advance to determine the K types of behavior data corresponding to the terminal and / or account, that is, the user behavior data can be divided into K categories, each of which corresponds to a user, and the K types of behavior data correspond to different users. In this way, it can correspond to K users, indicating that the above-mentioned user behavior data is the behavior data generated by K users under the terminal and / or the account, and K users use the terminal and / or account to perform related behaviors. One type of behavior data is the behavior data generated by the same user in the application, and the K types of behavior data correspond to the K users one by one. In addition, the K types of behavior data can be used to construct the behavior portraits of K users.

[0033] As an example, in the process of outputting the first content according to the behavior portraits of K users, the corresponding K candidate content sets can be selected according to the behavior portraits of the K users. One candidate content set corresponds to one of the K users. The K candidate content sets correspond to the K users one by one. Any candidate content set includes at least one content. As an example, the content may include but is not limited to items (including commodities, etc.), videos, books, news, etc. Then, the K candidate content sets are sampled to obtain the first content, and the first content is output. It should be noted that the number of content in the first content can be one or more, which is not limited in this embodiment.

[0034] The first input can be understood as the first action on the target content. For example, the first input is the first click input, which can be understood as the first click action. That is, after the first content is output, the current user can view the output first content and select the content. When the first input on the target content in the first content is received, it indicates the selection of the target content. Subsequently, the target user can be determined from K users based on the target content and K types of behavior data, and the target user currently using the terminal and / or the account can be determined.

[0035] Since the target user is a user among K users, the behavior profiles of the K users have been pre-constructed based on K types of behavior data. After the target user is determined, the behavior profile of the target user can be obtained from the behavior profiles of the K users, and the behavior preference of the target user can be obtained.

[0036] By using the behavioral profile of the target user to recommend the second content, that is, recommending content based on the behavioral preferences of the target user, the accuracy of content recommendation can be improved.

[0037] In this embodiment, when the user behavior recognition result is a multi-user behavior, content recommendation is performed. In the content recommendation process, the first content can be outputted first according to the behavior profiles of K users, and then the target user can be determined from the K users according to the target content corresponding to the first input in the first content and the K types of behavior data obtained by clustering the user behavior data, and the second content can be recommended based on the behavior profile of the target user. That is, in this embodiment, even if the user behavior data is the behavior data generated by K users, the first content can be outputted first according to the behavior profiles of K users, and then the first input of the target content in the first content can be received, and the target user who currently uses the terminal and / or the account can be determined according to the target content and the K types of behavior data, and the second content can be recommended by using the behavior profile of the target user who currently uses the terminal and / or the account, so as to improve the accuracy of content recommendation.

[0038] In one embodiment, content recommendation based on user behavior recognition results includes:

[0039] When the user behavior recognition result is a single user behavior, the third content is recommended based on the behavior portraits of K users obtained in advance. The behavior portraits of the K users are constructed in advance based on K types of behavior data. The K types of behavior data are obtained by clustering the user behavior data in advance based on the user behavior recognition result, and K is 1.

[0040] In the case where the user behavior identification result is a single-user behavior, it means that the user using the above-mentioned terminal and / or account is a single user. The user behavior data associated with the terminal and / or account is clustered in advance to obtain a type of behavior data, and a user behavior profile is constructed using this type of behavior data. In the content recommendation process, the behavior profile of this user is considered, and the third content is output to achieve content recommendation in the case of single-user behavior, so as to improve the accuracy of content recommendation in the case of single-user behavior. It can be understood that the content recommendation method of the embodiment of the present invention can use different methods to achieve content recommendation for multi-user behavior and single-user behavior, that is, to formulate corresponding recommendation processes for content recommendation to improve the accuracy of content recommendation.

[0041] In one embodiment, before recommending content based on the user behavior recognition result, the method further includes:

[0042] Obtain user behavior data;

[0043] Generate user behavior features based on user behavior data;

[0044] The user behavior features are input into the target recognition model to perform user behavior recognition and obtain the user behavior recognition results.

[0045] That is, user behavior recognition can be performed through the target recognition model to obtain a user behavior recognition result. In this embodiment, user behavior features are first generated based on user behavior data, and then user behavior recognition is performed on the user behavior features through the target recognition model to obtain a user behavior recognition result to improve the accuracy of user behavior recognition. As an example, the target recognition model can be, but is not limited to, a target factorization machine (FM), etc. As an example, the user behavior data can be behavior data within a preset historical time, for example, the preset historical time can be the last L days, and L can be a positive integer, etc.

[0046] In one example, the user behavior features may include a first-order user behavior feature vector and a second-order user behavior feature matrix. In the process of generating the user behavior features based on the user behavior data, a full behavior content set may be obtained first, and a first-order user behavior feature vector of the user behavior data may be generated according to the arrangement order of the behavior content in the full behavior content set, wherein the dimension of any first-order user behavior feature vector is the same as the number of behavior contents in the full behavior content set, and the value of the reference position in any first-order user behavior feature vector is a first preset value, and the values ​​of other positions except the reference position are second preset values, the reference position matches the sequential position of the reference behavior content in the full behavior content set, the reference behavior content is the behavior content in the user behavior data, and the full behavior content set may be a content set obtained by union processing of historical behavior contents in multiple historical behavior content sequences, and a historical behavior content sequence includes the historical behavior content of one user. Then, according to the number of behavior contents in the full behavior content set, a second-order user behavior feature matrix of the user behavior data is generated, the number of rows and columns of any second-order user behavior feature matrix are the same as the number of behavior contents in the full behavior content set, and when the i-th behavior content and the j-th behavior content in the full behavior content set are both in the spliced ​​behavior content sequence, the value of the i-th row and j-th column in the second-order user behavior feature matrix is ​​set to a first preset value; when at least one of the i-th behavior content and the j-th behavior content in the full behavior content set is not in the user behavior data, the value of the i-th row and j-th column in the second-order user behavior feature matrix is ​​set to a second preset value; wherein, 1≤i≤n, 1≤j≤n, and n is the number of behavior contents in the full behavior content set.

[0047] In one embodiment, after inputting the user behavior features into the target recognition model to perform user behavior recognition and obtaining the user behavior recognition result, the method further includes:

[0048] When the user behavior recognition result is a single-user behavior, K is set to 1, or, when the user behavior recognition result is a multi-user behavior, K is set to an integer greater than 1, where K is the number of clusters;

[0049] According to the value of K, the user behavior data is clustered to obtain K types of behavior data.

[0050] Different user behavior recognition results use different values ​​of the number of clusters K, resulting in different clustering results. In this embodiment, if the user behavior recognition result is a single-user behavior, the value of K is set to 1. It can be understood that when the value of K is 1, the user behavior data is clustered to obtain K types of behavior data, that is, 1 type of behavior data is obtained, and the noise data in the user behavior data (not in the above 1 type of behavior data) is filtered. If the user behavior recognition result is a multi-user behavior, the value of K is set to an integer greater than 1, and the user behavior data is clustered according to the value of K to obtain multiple types of behavior data. It can be understood that in the clustering process, different values ​​are set for K according to different user behavior recognition results, and clustering is performed according to the set value of K, which can improve the accuracy of clustering.

[0051] In one embodiment, the user behavior data includes M behavior contents, where M is a positive integer;

[0052] According to the value of K, the user behavior data is clustered to obtain K types of behavior data, including:

[0053] Clustering is performed according to the value of K and the first feature vectors of the M behavior contents to obtain K types of behavior data, wherein the first feature vectors of the M behavior contents are obtained in advance through feature learning of the target recognition model, and the user behavior features are generated according to the user behavior data.

[0054] Each of the M behavioral contents has a corresponding first feature vector, and the first feature vectors of the M behavioral contents can be understood as M first feature vectors. In the process of pre-training the initial recognition model through the second-order feature matrix in multiple sample behavioral features to obtain the target recognition model, the initial recognition model can be used to learn the corresponding first feature vector for each behavioral content in the full behavioral content set (for example, when the target recognition model is the target FM, the first feature vector can be understood as a latent vector), that is, feature learning is performed, and the second-order feature matrix includes the combined feature vector of each behavioral content in the full behavioral content set. Each element in the combined feature vector of any behavioral content in the full behavioral content set is the combined feature value between the behavioral content and the behavioral content in the full behavioral content set. If the two behavioral contents are in the spliced ​​behavioral content sequence corresponding to the same sample behavioral feature, the combined feature value between the two behavioral contents is the first preset value, otherwise it is the second preset value. When the training is completed, the feature learning is completed and the target recognition model is obtained. At this time, the first feature vector of each behavior content in the full behavior content set can be obtained. The full behavior content set includes the above-mentioned M behavior contents, that is, the first feature vectors of the M behavior contents can be obtained in advance through feature learning of the target recognition model.

[0055] In this embodiment, feature learning is performed in advance through the target recognition model to obtain the first feature vectors of M behavior contents. During the clustering process, the first feature vectors of the M behavior contents and the value of K can be used to cluster the user behavior data to obtain K types of behavior data, thereby improving the clustering accuracy.

[0056] In one embodiment, the target recognition model is obtained by the following training method:

[0057] Acquire multiple historical behavior content sequences, wherein one historical behavior content sequence includes historical behavior content of a user;

[0058] Based on multiple historical behavior content sequences, a target behavior sample set is constructed;

[0059] Generate a training sample behavior feature set based on the target behavior sample set and the historical behavior contents in the plurality of historical behavior content sequences;

[0060] The initial recognition model is trained based on the training sample behavior feature set to obtain a target recognition model.

[0061] Among them, any historical behavior content sequence may include multiple historical behavior contents. As an example, a historical behavior content sequence may include the historical behavior content of a preset time period within a preset historical time under a terminal and / or an account in the above-mentioned application. Multiple historical behavior content sequences may include historical behavior content of a preset time period within a preset historical time under multiple terminals and / or multiple accounts. The preset time period may be a time period between a first preset moment and a second preset moment. For example, the first preset moment may be eight o'clock in the morning, the second preset moment may be seven o'clock in the evening, etc. This embodiment does not make specific limitations.

[0062] In this embodiment, a target behavior sample set is constructed through multiple historical behavior content sequences, and then a training sample behavior feature set is generated by utilizing the target behavior sample set and the historical behavior contents in the multiple historical behavior content sequences. The training sample behavior feature set is used to train the initial recognition model to obtain the target recognition model, thereby realizing the training of the recognition model to improve the performance of the trained target recognition model.

[0063] In one embodiment, a target behavior sample set is constructed based on multiple historical behavior content sequences, including:

[0064] For each of the plurality of historical behavior content sequences, segment the historical behavior content sequence to obtain segmented behavior content of the historical behavior content sequence;

[0065] The segmented behavior contents of multiple historical behavior content sequences are spliced ​​to obtain a target behavior sample set, which includes multiple spliced ​​behavior content sequences.

[0066] Each historical behavior content sequence is segmented, and the number of segments of the segmented behavior content of each historical behavior content sequence is at least one, and any segmented behavior content includes at least one behavior content, and then the segmented behavior contents of multiple historical behavior content sequences can be spliced ​​to obtain a target behavior sample set. It should be noted that the segmented behavior contents of different historical behavior content sequences can be spliced, and different segmented behavior contents in the same historical behavior content sequence can be spliced.

[0067] In this embodiment, the target behavior sample set is obtained by splicing segmented behavior contents of multiple historical behavior content sequences. In this way, the obtained spliced ​​behavior content sequence can be made more diverse. The obtained target behavior sample set is used to generate a training sample behavior feature set, and the initial recognition model is trained accordingly to obtain a target recognition model, thereby improving the performance of the trained target recognition model.

[0068] In one embodiment, the target behavior sample set includes a first behavior sample set and a second behavior sample set, the first behavior sample set includes S spliced ​​behavior content sequences, and the second behavior sample set includes T spliced ​​behavior content sequences;

[0069] Among them, any sequence in the S spliced ​​behavior content sequences is spliced ​​together from at least two segmented behavior contents of different users, and the first similarity of any sequence in the S spliced ​​behavior content sequences is less than a first preset threshold value, the first similarity of a spliced ​​behavior content sequence is the similarity between second feature vectors of historical behavior content sequences of different users corresponding to the spliced ​​behavior content sequence, the second feature vector of the historical behavior content sequence is obtained by the embedding vector of each behavior content in the historical behavior content sequence, and the embedding vector of a behavior content is obtained by vector mapping through the identity identifier of the behavior content;

[0070] The T spliced ​​behavior content sequences include q spliced ​​content sequences and h spliced ​​behavior content sequences, any sequence in the q spliced ​​behavior content sequences is spliced ​​together from at least two segmented behavior content of the same user, any sequence in the h spliced ​​behavior content sequences is spliced ​​together from at least two segmented behavior content of different users, the second similarity of any sequence in the h spliced ​​content sequences is greater than a second preset threshold, and the second similarity of a spliced ​​behavior content sequence is the similarity between second feature vectors of historical behavior content sequences of different users corresponding to the spliced ​​behavior content sequences.

[0071] It can be understood that different historical behavior content sequences are historical behaviors of different users, and the spliced ​​behavior content sequence in the first behavior sample set is spliced ​​together from at least two segmented behavior contents of different users, that is, the spliced ​​behavior content sequence in the first behavior sample set is obtained by splicing together the segmented behavior contents of different historical behavior content sequences, that is, the behavior content in the same spliced ​​behavior content sequence in the first behavior sample set includes segmented behavior contents of multiple different users, which is multi-user behavior sample data, and can also be understood as negative sample data. The mark corresponding to the first behavior sample set is multi-user behavior, for example, the mark can be 0. The same spliced ​​behavior content sequence of the q spliced ​​behavior content sequences in the second behavior sample set is formed by the splicing of at least two segmented behavior contents of the same user, which can be understood as the splicing of different segmented behavior contents in the same historical behavior content sequence. The behavior content in the same spliced ​​behavior content sequence in the second behavior sample set includes different segmented behavior contents of the same user, that is, single-user behavior sample data. In addition, although the same spliced ​​behavior content sequence of the h spliced ​​behavior content sequences in the second behavior sample set is formed by the splicing of at least two segmented behavior contents of different users, the similarity between the second feature vectors of the corresponding historical behavior content sequences of different users is greater than the second preset threshold, that is, the same spliced ​​behavior content sequence of the h spliced ​​behavior content sequences is the splicing of segmented behavior contents between similar users, and their behaviors are similar. It can also be understood as single-user behavior sample data, or as positive sample data. The corresponding mark of the second behavior sample set is single-user behavior. For example, the mark can be 1.

[0072] In addition, in order to reduce the similarity of the segmented behavior content of different users in the spliced ​​behavior content sequence in the first behavior sample set and ensure the difference of the segmented behavior content in the spliced ​​behavior content sequence in the first behavior sample set, in this embodiment, it is necessary to ensure that the similarity between the historical behavior content of different users associated with the spliced ​​behavior content sequence in the first behavior sample set is low, that is, the first similarity of any spliced ​​behavior content sequence in the first behavior sample set is less than the first preset threshold. In addition, in order to improve the similarity between the segmented behavior content of different users in the spliced ​​behavior content sequence in h spliced ​​behavior contents, in this embodiment, it is necessary to ensure that the similarity between the second feature vectors of the historical behavior content sequences of different users associated with the spliced ​​behavior content sequence in h spliced ​​behavior contents is high, that is, the second similarity of any spliced ​​behavior content sequence in h spliced ​​behavior contents is greater than the second preset threshold, and the second preset threshold is greater than or equal to the first preset threshold. Through the above-mentioned target behavior sample set including the first behavior sample set and the second behavior sample set, a training sample behavior feature set is generated to train the initial recognition model to obtain a target recognition model, which can improve the performance of the model.

[0073] In one example, the second feature vector of a segmented behavior content may be the embedding vector of each behavior content in the segmented behavior content, and the second feature vector of a historical behavior content sequence may be the average of the embedding vectors of each behavior content in the historical behavior content sequence. The embedding vector of a behavior content is obtained by vector mapping through the identity identifier of the behavior content. For example, the identity identifier of the behavior content can be input into the item2vec model for mapping to obtain the second embedding vector of the behavior content.

[0074] In one embodiment, any sample behavior feature in the training sample behavior feature set includes a first-order feature vector and a second-order feature matrix;

[0075] According to the target behavior sample set and the historical behavior content in multiple historical behavior content sequences, a training sample behavior feature set is generated, including:

[0076] Performing a union process on the historical behavior contents in multiple historical behavior content sequences to obtain a full behavior content set;

[0077] Generate a first-order feature vector for each concatenated behavior content sequence in the target behavior sample set according to the arrangement order of the behavior content in the full behavior content set, wherein the dimension of any first-order feature vector is the same as the number of behavior content in the full behavior content set, and the value of the target position in any first-order feature vector is a first preset value, and the values ​​of other positions except the target position are second preset values, the target position matches the sequential position of the target behavior content in the full behavior content set, and the target behavior content is the behavior content of the concatenated behavior content sequence corresponding to the first-order feature vector;

[0078] According to the number of behavioral contents in the full behavioral content set, a second-order feature matrix of each concatenated behavioral content sequence in the target behavioral sample set is generated, and the number of rows and columns of any second-order feature matrix is ​​the same as the number of behavioral contents in the full behavioral content set;

[0079] Among them, S sample behavior features correspond to S spliced ​​behavior content sequences, and T sample behavior features correspond to T spliced ​​behavior content sequences.

[0080] For example, the first preset value is 1, the second preset value is 0, and the full set of behavior content is {item1, item2, ..., item n-1 ,item n}, a concatenated behavior content sequence is {item1, item2}, then the value of the corresponding position of item1-item2 in the first-order feature vector of the concatenated behavior content sequence is 1, and item3-item nThe value of the corresponding position is 0, that is, the first-order feature vector of the splicing behavior content sequence is (1, 1, 0, ..., 0), and the dimension is n. In addition, the size of the second-order feature matrix of any splicing behavior content sequence is n rows and n columns. The value of the i-th row and j-th column in the second-order feature matrix of a splicing behavior content sequence is item i and item j The combined eigenvalue between is 1 or 0.

[0081] In this embodiment, in the process of generating a training sample feature set, historical behavior contents in multiple historical behavior content sequences are firstly unioned to obtain a full behavior content set, and then a first-order feature vector of each spliced ​​behavior content sequence in the target behavior sample set is generated according to the arrangement order of the behavior contents in the full behavior content set, and a second-order feature matrix of each spliced ​​behavior content sequence in the target behavior sample set is generated according to the number of behavior contents in the full behavior content set. Each sample behavior feature in the obtained training sample behavior feature set includes a first-order feature vector and a second-order feature matrix, and the training sample behavior feature set is used for model training to improve the performance of the model.

[0082] In one embodiment, according to the number of behavior contents in the full behavior content set, a second-order feature matrix of each concatenated behavior content sequence in the target behavior sample set is generated, including:

[0083] For each spliced ​​behavior content sequence in the target behavior sample set, when the i-th behavior content and the j-th behavior content in the full behavior content set are both in the spliced ​​behavior content sequence, the value of the i-th row and j-th column in the second-order feature matrix of the spliced ​​behavior content sequence is set to a first preset value; when at least one of the i-th behavior content and the j-th behavior content in the full behavior content set is not in the spliced ​​behavior content sequence, the value of the i-th row and j-th column in the second-order feature matrix of the spliced ​​behavior content sequence is set to a second preset value;

[0084] Among them, 1≤i≤n, 1≤j≤n, and n is the number of behavioral contents in the full behavioral content set.

[0085] For example, the complete set of behavior content above is {item1, item2, ..., item n-1 ,item n} and the splicing behavior content sequence is {item1, item2}, the value of the 1st row and 1st column in the second-order feature matrix of the splicing behavior content sequence obtained represents the combined feature value between item1 and item1, and since item1 is in the splicing behavior content sequence, the combined feature value is 1, the value of the 1st row and 2nd column represents the combined feature value between item1 and item2, and since both item1 and item2 are in the splicing behavior content sequence, the combined feature value is 1, the value of the 2nd row and 1st column represents the combined feature value between item2 and item1, and since both item1 and item2 are in the splicing behavior content sequence, the combined feature value is 1, the value of the 2nd row and 2nd column represents the combined feature value between item2 and item2, and since item2 is in the splicing behavior content sequence, the combined feature value is 1, and the rest of the values ​​in the second-order feature matrix of the splicing behavior content sequence are all 0.

[0086] In this embodiment, by judging whether the ith behavior content and the jth behavior content in the full behavior content set are both in the spliced ​​behavior content sequence, the value of the ith row and jth column in the second-order feature matrix of the spliced ​​behavior content sequence is set according to the result, and different values ​​are set for different results. In this way, through similar judgments, corresponding values ​​are set for corresponding positions in the second-order feature matrix for every two behavior contents in the full behavior content set, thereby obtaining the second-order feature matrix of the spliced ​​behavior content sequence, and taking it together with the first-order feature matrix as a sample behavior feature. In this way, a training sample behavior feature set is constructed for model training to improve the performance of the model.

[0087] In one embodiment, determining a target user from K users according to the target content and K types of behavior data includes:

[0088] Determine target class behavior data from the K class behavior data according to the first feature vector of the target content and the cluster center feature vectors of each class behavior data in the K class behavior data;

[0089] Determine the user corresponding to the target behavior data as the target user;

[0090] The cluster center feature vector of the reference class behavior data is obtained according to the first feature vector of each behavior content in the reference class behavior data, and the reference class behavior data is any type of behavior data in the K types of behavior data.

[0091] The first feature vector of the target content is obtained in advance through feature learning of the target recognition model. For example, the above-mentioned full behavioral content set includes the target content, and the first feature vector of each behavioral content in the full behavioral content set is obtained in advance, and the first feature vector of the target content can be obtained from the first feature vector of the full behavioral content set. In one example, the cluster center feature vector of the reference class behavioral data can be the average of the first feature vectors of each behavioral content in the reference class behavioral data.

[0092] In this embodiment, the target class behavior data is determined from the K class behavior data through the first feature vector of the target content and the cluster center feature vectors of each class of behavior data in the K class behavior data, and then the user corresponding to the target class behavior data is determined as the target user. In this way, the accuracy of the determined target user can be improved, and the second content can be recommended using the behavior portrait of the target user, thereby improving the accuracy of content recommendation.

[0093] In one example, in the process of determining the target class behavior data from the K class behavior data, based on the first feature vector of the target content and the cluster center feature vectors of each class behavior data in the K class behavior data, the similarity between the first feature vector of the target content and the K cluster center feature vectors can be calculated to obtain K similarities, and the class behavior data corresponding to the maximum similarity among the K similarities is determined as the target class behavior data, and then the user corresponding to the target class behavior data is determined as the target user. It can be understood that the most similar user is determined as the target user, that is, the user currently using the above terminal and / or the above account.

[0094] The process of the above content recommendation method is described in detail below with reference to a specific embodiment.

[0095] like Figure 2 As shown, the process of the content recommendation method of this embodiment is as follows:

[0096] First, collect behavioral data such as multiple users clicking or playing digital content in the application to obtain multiple historical behavior content sequences, and construct a training sample behavior feature set. Use the training sample behavior feature set to train the model to obtain the target FM model, and obtain the latent vector of each behavior content in the full behavior content set;

[0097] Then, the user behavior data of a certain terminal or a certain account is obtained, and the user behavior data can be identified through the target FM model to obtain the user behavior identification result. The user behavior data is clustered using the user behavior identification result to obtain K types of behavior data, and the K types of behavior data are used to construct the behavior portraits of K users;

[0098] Secondly, when the terminal starts the application or the account logs into the application, the user behavior identification result is obtained, and content recommendation is performed based on the user behavior identification result.

[0099] The process of content recommendation based on user behavior recognition results is as follows:

[0100] If the user behavior recognition result is a multi-user behavior, the system can select a corresponding candidate content set based on the behavior portraits of K users, sample the candidate content set to obtain the first content, and output the first content. In this case, K is greater than 1.

[0101] The user clicks on the target content in the first content according to his / her own preference;

[0102] The system determines the target user from K users based on the target content clicked by the user;

[0103] The system obtains the behavior profile of the target user from the behavior profiles of K users, and pushes the second content based on the behavior profile of the target user, thereby realizing multiple user identification and personalized content recommendation under the same terminal or the same account.

[0104] It should be noted that the above content recommendation method can be implemented through a content recommendation system, such as Figure 3 As shown, the content recommendation system may include a user behavior collection module, a multi-user analysis and identification module, a specific user exploration module, a specific user switching module, and a user personalized content recommendation module.

[0105] User behavior collection module: responsible for collecting and storing user behavior data on digital content applications through embedding methods;

[0106] Multi-user analysis and identification module: responsible for analyzing and mining the user behavior under the same account or the same terminal from massive user behavior data, building a multi-user behavior identification model (target identification model), and characterizing user portrait data based on user behavior data, that is, obtaining user behavior portraits;

[0107] Specific user exploration module: determines user preferences based on the user's interactive behavior in digital content applications, and further confirms the specific user identity of the current user under the same terminal or the same account, which can be understood as determining the target user;

[0108] Specific user switching module: The person in charge notifies the personalized recommendation module to switch the current behavior profile to the behavior profile of a specific user;

[0109] User personalized content recommendation module: pushes relevant digital content to the current user based on the behavior portrait confirmed by the switch and the user's real-time behavior.

[0110] The process of model training to obtain the target recognition model is as follows:

[0111] (1) Single-user client behavior sampling: From the total user behavior, extract the user behavior of a single-user client in the last L days. Since we do not know which users are single-user clients, we select the user behavior data that occurs during working hours on weekdays (between 8:00 and 19:00, which has the highest probability of being generated by the same user) and mark it as the behavior data generated by a single user. For each user, the content (item) sequence of his / her behavior in the last L days is generated according to the occurrence time in the following format (i.e., the historical behavior content sequence):

[0112] Sep X : {item1, item2,…, item m ,};

[0113] Among them, Sep X For user X X )’s historical behavior content sequence, m is user X The total number of items in the last L days. In an example, a minimum threshold Size can be pre-set. min , sampling needs to satisfy m>Size min To ensure that the user has a certain level of activity;

[0114] (2) Segmentation of sampled user behavior sequences: The historical behavior content sequence of each user in (1) above is divided into k segments (the division method may be equal or divided according to the date of occurrence of the behavior, which is not limited in this embodiment). Taking equal division as an example, segmented behavior content in the following format is generated:

[0115] Sep X1 : {item1, item2,…, item m / k};

[0116] Sep X2 :{item m / k+1 , item m / k+2 ,……,item m*2 / k};

[0117] …

[0118] Sep Xi :{item m(i-1) / k+1 , item m(i-1) / k+2 ,……,item m*i / k};

[0119] …

[0120] Sep Xk :{item m(k-1) / k+1, item m(k-1) / k+2 ,……,item m};

[0121] Among them, k is the number of segments, Sep Xi For the i-th segment behavior content corresponding to the historical behavior content sequence of user X, a minimum threshold C is set for m / k min , set the size of k according to the m value of each user, and m / k>C min , that is, to ensure that the number of segmented behavior contents of each user is greater than C min ;

[0122] (3) Constructing training samples: Generate positive and negative samples for determining multi-user clients through the segmented behavior content of the user's historical behavior content sequence in (2). Assume that a client is a single user and is positive, while a client is a multi-user and is negative. The method for constructing positive and negative samples is as follows:

[0123] Negative sample construction (i.e. the first line is sample set construction): for two different users user X and user Y (User Y) randomly extracts several segments (from 1 to k) from their own historical behavior content sequences and splices them to generate a spliced ​​behavior content sequence. For example, extract Sep Xi and Sep Yi (is the i-th segment behavior content corresponding to the historical behavior content sequence of user Y) and the new connected behavior content sequence is generated as: Sep XiYi :{Sep Xi , Sep Yi}, due to Sep XiYi is user X and user Y The i-th segment behavior content is spliced ​​together and comes from multiple users, so Sep XiYi As a negative sample, its label is 0. Taking this as an example, repeating the above random extraction and merging operations multiple times can generate several negative samples; it should be noted that considering the user X and user Y It may be similar users with the same interests. It is not accurate to splice the behavior sequences of similar users as negative samples. Therefore, when constructing negative samples, it is necessary to ensure that the user X and user YThe similarity cannot be too high. The historical behavior content sequence in (1) can be used to train an item2vec model (which is an evolution of the word2vec model (word vector model) and can treat items as words) to obtain the embedding vector of the entire behavior content. Then, for each user, the embedding vectors of each behavior content in the user's historical behavior sequence are summed and averaged to represent the user as the embedding vector of the user. X and user Y The similarity is obtained by the inner product of the user's embedding vector, and a first preset threshold simi_score is set False , for the user who selects to construct negative samples X and user Y , which needs to satisfy the inner product of the embedded vector <simi_score False ;

[0124] Positive sample construction (i.e. the second behavior is sample set construction): for a user user X , randomly extract the segmented behavior content Sep X1 ~Sep Xk Several segments (from 1 to k) in the splicing are combined to generate a new splicing behavior content sequence, such as extracting Sep Xi and Sep Xj , then the generated new splicing behavior content sequence is: Sep XiXj :{Sep Xi , Sep Xj}, due to Sep XiXj is user X The behavior sequences of different segments are spliced ​​together, all of which are user X Therefore, Sep XiXj As a positive sample, it is marked as 1; in addition, you can refer to the negative sample construction method and select two users with high similarity. X and user Y , randomly extract several segments (from 1 to k) of each segmented behavior content and splice them as a supplement to the positive sample, and set a second preset threshold simi_score True , for the user who selects the positive sample X and user Y , the inner product of the embedded vector must be greater than simi_score True Taking this as an example, repeating the above random extraction and merging operations multiple times can generate several positive samples;

[0125] In this way, a target behavior sample set including the first behavior sample set and the second behavior sample set is generated, and then a training sample behavior feature set is generated based on the full behavior content set of the target behavior sample set.

[0126] (4) Identification model (classification model) training: Use the training sample behavior feature set generated in (3) as the input of the identification model to train a multi-user client discrimination model, that is, the target identification model. The core idea is to use the behavior features of the full amount of behavior content as the features of the model for learning. When two contents with large differences appear in the same behavior sequence, the probability of being determined as multi-user will increase, while when two contents with small differences appear in the same behavior sequence, the probability of being determined as single-user will increase. Second-order combination of all features and full learning of the influence of combined features on the classification result. The model expression is as follows:

[0127]

[0128] In this embodiment, x in the first-order feature vector of the model i is the feature score of the content item that appears in the target behavior sample set i . Assume that the full amount of behavior content set is {item1, item2,..., item n-1 , item n}. If a spliced behavior content sequence is {item1, item2,..., item m} (m < n), then the values at the corresponding positions of x1 to x m in the first-order feature vector input to the model are equal to 1, and the values at the corresponding positions of x m+1 to x n are equal to 0; the value x i x j at the i-th row and j-th column of the second-order feature matrix of the model is the combined feature value (combined feature score) of the content item i and item j in the target behavior sample set. For the spliced behavior content sequence {item1, item2,..., item m} (m < n), the combined feature scores of any two items between item1 and item m are 1, and the scores of other combined features are 0; in addition, if it is a single-user sample, the model y value is 1, and if it is a multi-user sample, the model y value is 0. Through training with a large number of samples, a multi-user client discrimination model based on FM can be obtained.

[0129] (5) After training, discriminate all users. Use the user behavior data of all clients in the most recent L days (the difference from (1) is that the user behavior data during non-working hours is included at this time), generate the user behavior data of each client, and use it as the model input. Through model prediction, it can be determined whether the client is a single user or a multi-user;

[0130] (6) Based on the user behavior recognition result, the user behavior data of the client can be clustered to obtain K types of behavior data, and behavior profiles of K users can be constructed. When the client subsequently launches the application, the first content can be output based on the constructed behavior profiles of the K users, and then the target user can be determined from the K users based on the user's click behavior on the target content in the first user, and the second content can be recommended using the behavior profile of the target user;

[0131] It should be noted that in addition to realizing the recognition of multi-user clients, the above recognition model also obtains the latent vectors of all items. Since the items are used as feature inputs, the latent vector of each item can be directly obtained. Assuming that the total number of items is n, one advantage of the FM model is that after factoring the n*n dimensional symmetric second-order feature matrix composed of the second-order cross feature weights, an n*v dimensional low-dimensional matrix can be obtained, where the i-th row of the low-dimensional matrix is ​​the latent vector of the i-th item. Two content items appearing in the behavior sequence of the same client x and item y , the larger the inner product of its latent vector, the greater the probability that the client is a single user, which means that the two items are more similar. Therefore, for each client, the items in its behavior sequence can be k-means clustered according to the latent vector. The number of clusters K represents the number of users of the client. For clients judged as single users, the number of users is 1, that is, the number of clusters K = 1, and a few items away from the cluster center are filtered as noise, so that the single user preference characterization is more accurate; for clients judged as multi-users, K>1 is specified for clustering, and the items in the client behavior sequence are divided into K categories to realize the preference characterization of multiple users of the client and realize the personalized recommendation strategy for multi-user clients.

[0132] The present application proposes a multi-user identification and content recommendation method, which identifies different users by analyzing the user viewing or on-demand behavior of a client or account, selects candidate sets of recommended content according to different users, determines to switch to a specific user based on the user terminal interaction behavior, and pushes digital content that meets the preferences of the specific user.

[0133] See also Figure 4 , a content recommendation device 400 of an embodiment is provided, comprising:

[0134] The first acquisition module 401 is used to acquire a user behavior recognition result when the terminal starts the application and / or logs in to the application through an account. The user behavior recognition result is a recognition result obtained by pre-identifying the user behavior data through a target recognition model. The user behavior recognition result includes multi-user behavior or single-user behavior. The user behavior data is associated with the terminal and / or the account. The target recognition model is trained by a training sample behavior feature set. The training sample behavior feature set includes S sample behavior features with a behavior mark of multi-user behavior and T sample behavior features with a behavior mark of single-user behavior. Both S and T are positive integers.

[0135] The recommendation module 402 is used to recommend content based on the user behavior recognition result.

[0136] In one embodiment, the recommendation module 402 includes:

[0137] an output module, configured to output the first content according to the behavior portraits of K users obtained in advance when the user behavior recognition result is a multi-user behavior, wherein the behavior portraits of the K users are constructed in advance according to K types of behavior data, and the K types of behavior data are obtained by clustering the user behavior data in advance according to the user behavior recognition result, and K is an integer greater than 1;

[0138] a determination module, configured to determine a target user from K users according to the target content and K types of behavior data when receiving a first input of the target content in the first content;

[0139] The second acquisition module is used to acquire the behavior profile of the target user from the behavior profiles of K users;

[0140] The first recommendation submodule is used to recommend the second content based on the behavior profile of the target user.

[0141] In one embodiment, the recommendation module 402 further includes:

[0142] The second recommendation submodule is used to recommend the third content based on the behavior portraits of K users obtained in advance when the user behavior recognition result is a single user behavior. The behavior portraits of the K users are constructed in advance based on K types of behavior data. The K types of behavior data are obtained by clustering the user behavior data in advance based on the user behavior recognition result, and K is 1.

[0143] In one embodiment, the apparatus 400 further includes:

[0144] The third acquisition module is used to acquire user behavior data;

[0145] A first feature generation module, used to generate user behavior features according to user behavior data;

[0146] The behavior recognition module is used to input user behavior features into the target recognition model to perform user behavior recognition and obtain user behavior recognition results.

[0147] In one embodiment, the apparatus 400 further includes:

[0148] A setting module, used to set K to 1 when the user behavior recognition result is a single-user behavior, or to set K to an integer greater than 1 when the user behavior recognition result is a multi-user behavior, where K is the number of clusters;

[0149] The behavior data clustering module is used to cluster user behavior data according to the value of K to obtain K types of behavior data.

[0150] In one embodiment, the user behavior data includes M behavior contents, where M is a positive integer;

[0151] Among them, the behavior data clustering module is specifically used to cluster according to the value of K and the first feature vectors of M behavior contents to obtain K types of behavior data, wherein the first feature vectors of M behavior contents are pre-obtained through feature learning of the target recognition model, and the user behavior features are generated according to the user behavior data.

[0152] In one embodiment, the target recognition model is obtained by the following training method:

[0153] Acquire multiple historical behavior content sequences, wherein one historical behavior content sequence includes historical behavior content of a user;

[0154] Based on multiple historical behavior content sequences, a target behavior sample set is constructed;

[0155] Generate a training sample behavior feature set based on the target behavior sample set and the historical behavior contents in the plurality of historical behavior content sequences;

[0156] The initial recognition model is trained based on the training sample behavior feature set to obtain a target recognition model.

[0157] In one embodiment, a target behavior sample set is constructed based on multiple historical behavior content sequences, including:

[0158] For each of the plurality of historical behavior content sequences, segment the historical behavior content sequence to obtain segmented behavior content of the historical behavior content sequence;

[0159] The segmented behavior contents of multiple historical behavior content sequences are spliced ​​to obtain a target behavior sample set, which includes multiple spliced ​​behavior content sequences.

[0160] In one embodiment, the target behavior sample set includes a first behavior sample set and a second behavior sample set, the first behavior sample set includes S spliced ​​behavior content sequences, and the second behavior sample set includes T spliced ​​behavior content sequences;

[0161] Among them, any sequence in the S spliced ​​behavior content sequences is spliced ​​together from at least two segmented behavior contents of different users, and the first similarity of any sequence in the S spliced ​​behavior content sequences is less than a first preset threshold value, the first similarity of a spliced ​​behavior content sequence is the similarity between second feature vectors of historical behavior content sequences of different users corresponding to the spliced ​​behavior content sequence, the second feature vector of the historical behavior content sequence is obtained by the embedding vector of each behavior content in the historical behavior content sequence, and the embedding vector of a behavior content is obtained by vector mapping through the identity identifier of the behavior content;

[0162] The T spliced ​​behavior content sequences include q spliced ​​content sequences and h spliced ​​behavior content sequences, any sequence in the q spliced ​​behavior content sequences is spliced ​​together from at least two segmented behavior content of the same user, any sequence in the h spliced ​​behavior content sequences is spliced ​​together from at least two segmented behavior content of different users, the second similarity of any sequence in the h spliced ​​content sequences is greater than a second preset threshold, and the second similarity of a spliced ​​behavior content sequence is the similarity between second feature vectors of historical behavior content sequences of different users corresponding to the spliced ​​behavior content sequences.

[0163] In one embodiment, any sample behavior feature in the training sample behavior feature set includes a first-order feature vector and a second-order feature matrix;

[0164] According to the target behavior sample set and the historical behavior content in multiple historical behavior content sequences, a training sample behavior feature set is generated, including:

[0165] Performing a union process on the historical behavior contents in multiple historical behavior content sequences to obtain a full behavior content set;

[0166] Generate a first-order feature vector for each concatenated behavior content sequence in the target behavior sample set according to the arrangement order of the behavior content in the full behavior content set, wherein the dimension of any first-order feature vector is the same as the number of behavior content in the full behavior content set, and the value of the target position in any first-order feature vector is a first preset value, and the values ​​of other positions except the target position are second preset values, the target position matches the sequential position of the target behavior content in the full behavior content set, and the target behavior content is the behavior content of the concatenated behavior content sequence corresponding to the first-order feature vector;

[0167] According to the number of behavioral contents in the full behavioral content set, a second-order feature matrix of each concatenated behavioral content sequence in the target behavioral sample set is generated, and the number of rows and columns of any second-order feature matrix is ​​the same as the number of behavioral contents in the full behavioral content set;

[0168] Among them, S sample behavior features correspond to S spliced ​​behavior content sequences, and T sample behavior features correspond to T spliced ​​behavior content sequences.

[0169] In one embodiment, according to the number of behavior contents in the full behavior content set, a second-order feature matrix of each concatenated behavior content sequence in the target behavior sample set is generated, including:

[0170] For each spliced ​​behavior content sequence in the target behavior sample set, when the i-th behavior content and the j-th behavior content in the full behavior content set are both in the spliced ​​behavior content sequence, the value of the i-th row and j-th column in the second-order feature matrix of the spliced ​​behavior content sequence is set to a first preset value; when at least one of the i-th behavior content and the j-th behavior content in the full behavior content set is not in the spliced ​​behavior content sequence, the value of the i-th row and j-th column in the second-order feature matrix of the spliced ​​behavior content sequence is set to a second preset value;

[0171] Among them, 1≤i≤n, 1≤j≤n, and n is the number of behavioral contents in the full behavioral content set.

[0172] In one embodiment, the determining module includes:

[0173] A target class behavior data determination module, used to determine the target class behavior data from the K class behavior data according to the first feature vector of the target content and the cluster center feature vectors of each class of behavior data in the K class behavior data;

[0174] A target user determination module is used to determine the user corresponding to the target class behavior data as the target user;

[0175] The cluster center feature vector of the reference class behavior data is obtained according to the first feature vector of each behavior content in the reference class behavior data, and the reference class behavior data is any type of behavior data in the K types of behavior data.

[0176] The technical features of the above-mentioned content recommendation device 400 correspond to the technical features of the above-mentioned content recommendation method. The various processes of the above-mentioned content recommendation method are implemented by the content recommendation device 400, and the same effect can be obtained. To avoid repetition, they are not described again here.

[0177] An embodiment of the present invention further provides a communication device, including a processor and a memory, wherein the memory stores a computer program that can be run on the processor. When the computer program is executed by the processor, the various processes in the above-mentioned content recommendation method embodiment are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0178] An embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the various processes of the above-mentioned content recommendation method embodiment are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0179] Among them, the computer readable storage medium, such as read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), a disk or an optical disk, etc.

[0180] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0181] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a communication device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods of various embodiments of the present invention.

[0182] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A content recommendation method, characterized in that: include: When the terminal starts the application and / or logs in to the application through the account, a user behavior recognition result is obtained, wherein the user behavior recognition result is a recognition result obtained by pre-recognizing the user behavior data through the target recognition model, wherein the user behavior recognition result includes multi-user behavior or single-user behavior, wherein the user behavior data is associated with the terminal and / or the account, wherein the target recognition model is obtained by training a training sample behavior feature set, wherein the training sample behavior feature set includes S sample behavior features whose behaviors are marked as multi-user behaviors and T sample behavior features whose behaviors are marked as single-user behaviors, wherein both S and T are positive integers; Recommending content based on the user behavior recognition result; The content recommendation based on the user behavior recognition result includes: when the user behavior recognition result is a multi-user behavior, outputting the first content according to the behavior portraits of K users obtained in advance, wherein the behavior portraits of the K users are constructed in advance according to K types of behavior data, and the K types of behavior data are obtained by clustering the user behavior data in advance according to the user behavior recognition result, and K is an integer greater than 1; when receiving a first input of a target content in the first content, determining a target user from the K users according to the target content and the K types of behavior data; obtaining a behavior portrait of the target user from the behavior portraits of the K users; and recommending a second content based on the behavior portrait of the target user; In the case where the user behavior identification result is a single-user behavior, the third content is recommended based on the behavior portraits of K users obtained in advance, where the behavior portraits of the K users are constructed in advance based on K types of behavior data, and the K types of behavior data are obtained by clustering the user behavior data in advance based on the user behavior identification result, and K is 1.

2. The method according to claim 1, characterized in that Before performing content recommendation based on the user behavior recognition result, the method further includes: Obtaining the user behavior data; generating user behavior features according to the user behavior data; The user behavior feature is input into the target recognition model to perform user behavior recognition to obtain the user behavior recognition result.

3. The method according to claim 2, characterized in that The user behavior data includes M behavior contents, where M is a positive integer; The user behavior data is clustered to obtain the K types of behavior data, including: Clustering is performed according to the value of K and the first feature vectors of the M behavior contents to obtain the K types of behavior data, wherein the first feature vectors of the M behavior contents are obtained in advance by feature learning through the target recognition model.

4. The method according to claim 1, characterized in that: The target recognition model is obtained by the following training method: Acquire multiple historical behavior content sequences, wherein one historical behavior content sequence includes historical behavior content of a user; Based on the multiple historical behavior content sequences, construct a target behavior sample set; Generating the training sample behavior feature set according to the target behavior sample set and the historical behavior contents in the plurality of historical behavior content sequences; The initial recognition model is trained based on the training sample behavior feature set to obtain the target recognition model.

5. The method according to claim 4, characterized in that The step of constructing a target behavior sample set based on the plurality of historical behavior content sequences includes: For each of the plurality of historical behavior content sequences, segmenting the historical behavior content sequence to obtain segmented behavior content of the historical behavior content sequence; The target behavior sample set is obtained by splicing the segmented behavior contents of the plurality of historical behavior content sequences, wherein the target behavior sample set includes a plurality of spliced ​​behavior content sequences.

6. The method according to claim 5, characterized in that The target behavior sample set includes a first behavior sample set and a second behavior sample set, the first behavior sample set includes S spliced ​​behavior content sequences, and the second behavior sample set includes T spliced ​​behavior content sequences; Among them, any sequence in the S spliced ​​behavior content sequences is spliced ​​together from at least two segmented behavior contents of different users, and the first similarity of any sequence in the S spliced ​​behavior content sequences is less than a first preset threshold value, the first similarity of a spliced ​​behavior content sequence is the similarity between second feature vectors of historical behavior content sequences of different users corresponding to the spliced ​​behavior content sequence, the second feature vector of the historical behavior content sequence is obtained by the embedding vector of each behavior content in the historical behavior content sequence, and the embedding vector of a behavior content is obtained by vector mapping through the identity identifier of the behavior content; The T spliced ​​behavior content sequences include q spliced ​​behavior content sequences and h spliced ​​behavior content sequences, any sequence in the q spliced ​​behavior content sequences is spliced ​​together from at least two segmented behavior contents of the same user, any sequence in the h spliced ​​behavior content sequences is spliced ​​together from at least two segmented behavior contents of different users, the second similarity of any sequence in the h spliced ​​behavior content sequences is greater than a second preset threshold, and the second similarity of a spliced ​​behavior content sequence is the similarity between second feature vectors of historical behavior content sequences of different users corresponding to the spliced ​​behavior content sequence.

7. The method according to claim 6, characterized in that Any sample behavior feature in the training sample behavior feature set includes a first-order feature vector and a second-order feature matrix; The step of generating a training sample behavior feature set according to the target behavior sample set and the historical behavior contents in the plurality of historical behavior content sequences includes: Performing a union process on the historical behavior contents in the plurality of historical behavior content sequences to obtain a full behavior content set; Generate a first-order feature vector for each concatenated behavior content sequence in the target behavior sample set according to the arrangement order of the behavior content in the full behavior content set, wherein the dimension of any first-order feature vector is the same as the number of behavior content in the full behavior content set, and the value of the target position in any first-order feature vector is a first preset value, and the values ​​of other positions except the target position are second preset values, the target position matches the sequential position of the target behavior content in the full behavior content set, and the target behavior content is the behavior content of the concatenated behavior content sequence corresponding to the first-order feature vector; According to the number of behavior contents in the full behavior content set, a second-order feature matrix of each concatenated behavior content sequence in the target behavior sample set is generated, wherein the number of rows and the number of columns of any second-order feature matrix are the same as the number of behavior contents in the full behavior content set; The S sample behavior features correspond to the S spliced ​​behavior content sequences, and the T sample behavior features correspond to the T spliced ​​behavior content sequences.

8. The method according to claim 7, characterized in that The step of generating a second-order feature matrix of each spliced ​​behavior content sequence in the target behavior sample set according to the number of behavior contents in the full behavior content set includes: For each spliced ​​behavior content sequence in the target behavior sample set, when the ith behavior content and the jth behavior content in the full behavior content set are both in the spliced ​​behavior content sequence, the value of the ith row and jth column in the second-order feature matrix of the spliced ​​behavior content sequence is set to a first preset value; when at least one of the ith behavior content and the jth behavior content in the full behavior content set is not in the spliced ​​behavior content sequence, the value of the ith row and jth column in the second-order feature matrix of the spliced ​​behavior content sequence is set to a second preset value; Wherein, 1≤i≤n, 1≤j≤n, and n is the number of behavior contents in the full behavior content set.

9. The method according to claim 1, characterized in that: The determining a target user from the K users according to the target content and the K types of behavior data includes: Determining target class behavior data from the K class behavior data according to the first feature vector of the target content and the cluster center feature vectors of each class of behavior data in the K class behavior data; Determine the user corresponding to the target class behavior data as the target user; The cluster center feature vector of the reference behavior data is obtained according to the first feature vector of each behavior content in the reference behavior data, and the reference behavior data is any type of behavior data in the K types of behavior data.

10. A communication device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps in the content recommendation method as described in any one of claims 1 to 9 are implemented.

11. A computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of the content recommendation method as described in any one of claims 1 to 9.

Images