Media recommendation method and device, electronic equipment, storage medium and program product

Abstract

The application provides a media recommendation method and device, electronic equipment, a storage medium and a program product, which are applied to the technical field of computers. The method comprises the following steps: acquiring attribute information of a plurality of media interacted by a family account in a historical period and a current period; performing feature coding on the attribute information of each media to obtain a coding vector of each media; determining an interaction probability of the family account for each media based on the coding vector of each media and a mixed behavior sequence of the family account; extracting a target behavior sub-sequence of a target family member from the mixed behavior sequence according to the interaction probability of the family account for each media and a real-time behavior sequence of the target family member; and performing media recommendation on the target family member based on the target behavior sub-sequence of the target family member and the real-time behavior sequence of the target family member. The method improves the accuracy of media recommendation.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a media asset recommendation method, apparatus, electronic device, storage medium, and program product. Background Technology

[0002] With the rapid development of mobile internet and video technologies, and the increasing popularity of online TV and short video apps, people's daily lives have been greatly enriched. For recommendation systems of video devices, accurately capturing user interests is a core challenge. Existing media asset recommendation methods primarily rely on users' viewing behavior sequences. However, for large-screen TVs, each TV box ID corresponds to a family, and a family often includes multiple family members with varying viewing interests. Therefore, the user behavior sequence for the same TV box contains a mix of interest records from different family members. Current media asset recommendation methods simply assume that the behavior sequence corresponding to the same TV box ID represents the same user's interests, without distinguishing the behaviors of different family members under the same TV box ID. This makes it impossible to provide targeted recommendations to different members corresponding to a single TV box ID, resulting in low recommendation accuracy. Summary of the Invention

[0003] This application provides a media asset recommendation method, apparatus, electronic device, storage medium, and program product to address the problem of low recommendation accuracy in existing media asset recommendation methods.

[0004] To solve the above-mentioned technical problems, this application is implemented as follows:

[0005] In a first aspect, embodiments of this application provide a media asset recommendation method, which includes:

[0006] Obtain attribute information of multiple media assets interacted with by the family account in historical and current time periods;

[0007] The attribute information of each media asset is feature-encoded to obtain the encoding vector of each media asset;

[0008] Based on the encoding vector of each media asset and the mixed behavior sequence of the family account, the interaction probability of the family account with each media asset is determined, wherein the mixed behavior sequence represents the media asset sequence of the family account in historical time periods and current time periods;

[0009] Based on the interaction probability of the family account with each media asset and the real-time behavior sequence of the target family member, the target behavior subsequence of the target family member is extracted from the mixed behavior sequence. The target family member is the family member who is currently using the family account to interact with the media asset. The target behavior subsequence represents the media asset sequence interacted by the target family member in a historical period. The real-time behavior sequence represents the media asset sequence interacted by the target family member in the current period.

[0010] Media asset recommendations are made for the target family members based on the target family member's target behavior subsequence and the target family member's real-time behavior sequence.

[0011] Optionally, the step of extracting the target behavior subsequence of the target family member from the mixed behavior sequence based on the interaction probability of the family account with each media asset and the real-time behavior sequence of the target family member includes:

[0012] A complete graph is constructed based on the multiple media assets and the interaction probability of the family account with each media asset. The complete graph includes multiple first nodes and multiple first edges. Any two first nodes are connected by a first edge. The first node is used to represent the media asset. The weight of the first node is the interaction probability of the media asset corresponding to the first node. The weight of the first edge is the js divergence calculated based on the interaction probabilities of the media assets corresponding to the two connected first nodes.

[0013] The first edge with a weight greater than a preset value in the complete graph is broken to obtain k subgraphs, where each subgraph is used to represent the first row of a member as a subsequence;

[0014] Based on the first behavior subsequence of each of the k members and the real-time behavior sequence of the target family member, the target behavior subsequence of the target family member is obtained.

[0015] Optionally, obtaining the target behavior subsequence of the target family member based on the first behavior subsequence of each of the k members and the real-time behavior sequence of the target family member includes:

[0016] The first behavior subsequence of each of the k members and the real-time behavior sequence of the target family member are input into the gating structure;

[0017] Based on the real-time behavior sequence of the target family member, the weight of each first behavior subsequence in the k first behavior subsequences is calculated using the self-attention module of the gating structure;

[0018] The k first-behavior subsequences are fused based on the weight of each first-behavior subsequence to obtain the target behavior subsequence of the target family member.

[0019] Optionally, after breaking the first edge with a weight greater than a preset value in the complete graph to obtain k subgraphs, the method further includes:

[0020] A directed graph is created based on the mixed behavior sequence. The directed graph includes multiple second nodes and multiple directed edges. Each directed edge connects two of the multiple second nodes. The second nodes are used to represent the media assets. The interaction time of the media asset corresponding to the starting point of the directed edge is earlier than the interaction time of the media asset corresponding to the ending point of the directed edge. The weight of the directed edge is determined based on a target time interval and the interaction duration of the media asset corresponding to the starting point of the directed edge. The target time interval is the difference between the interaction time of the media asset corresponding to the starting point of the directed edge and the interaction time of the media asset corresponding to the ending point of the directed edge.

[0021] Randomly select k second nodes as starting points from the directed graph, and perform random walks based on the weights of the directed edges in the directed graph to obtain the second row of each of the k members as a subsequence.

[0022] The process of obtaining the target behavior subsequence of the target family member based on the first behavior subsequence of each of the k members and the real-time behavior sequence of the target family member includes:

[0023] Based on the first behavior subsequence of each of the k members, the second behavior subsequence of each of the k members, and the real-time behavior sequence of the target family member, the target behavior subsequence of the target family member is obtained.

[0024] Optionally, obtaining the target behavior subsequence of the target family member based on the first behavior subsequence of each of the k members, the second behavior subsequence of each of the k members, and the real-time behavior sequence of the target family member includes:

[0025] The k first-line subsequences, the k second-line subsequences, and the real-time behavior sequence of the target family member are input into the gating structure;

[0026] Based on the real-time behavior sequence of the target family member, the weight of each first behavior subsequence in the k first behavior subsequences and the weight of each second behavior subsequence in the k second behavior subsequences are calculated using the self-attention module of the gating structure.

[0027] Based on the weight of each first-beginning subsequence in the k first-beginning subsequences and the weight of each second-beginning subsequence in the k second-beginning subsequences, the k first-beginning subsequences and the k second-beginning subsequences are fused to obtain the target behavior subsequence corresponding to the target family member.

[0028] Optionally, the weight of the directed edge is the reciprocal of the target value, where the target value is the difference between the target time interval and the interaction duration.

[0029] Secondly, embodiments of this application also provide a media asset recommendation device, which includes:

[0030] The first acquisition module is used to acquire attribute information of multiple media assets that the family account interacted with in historical and current time periods;

[0031] The second acquisition module is used to perform feature encoding on the attribute information of each media asset to obtain the encoding vector of each media asset;

[0032] The first determining module is used to determine the interaction probability of the family account with each media asset based on the encoding vector of each media asset and the mixed behavior sequence of the family account;

[0033] The first extraction module is used to extract the target behavior subsequence of the target family member from the mixed behavior sequence based on the interaction probability of the family account with each media asset and the real-time behavior sequence of the target family member. The target family member is the family member who is currently using the family account to interact with the media asset. The target behavior subsequence represents the media asset sequence interacted by the target family member in a historical period. The real-time behavior sequence represents the media asset sequence interacted by the target family member in the current period.

[0034] The first recommendation module is used to recommend media resources to the target family members based on the target behavior subsequences and the real-time behavior sequences of the target family members.

[0035] Thirdly, embodiments of this application also provide an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the media asset recommendation method described above.

[0036] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the media asset recommendation method described above.

[0037] Fifthly, a computer program product is provided, including computer instructions that, when executed by a processor, implement the steps of the media asset recommendation method described above.

[0038] The media asset recommendation method of this application embodiment obtains attribute information of multiple media assets interacted by a family account in historical and current time periods; performs feature encoding on the attribute information of each media asset to obtain an encoding vector for each media asset; determines the interaction probability of the family account with each media asset based on the encoding vector of each media asset and the mixed behavior sequence of the family account; extracts the target behavior subsequence of the target family member from the mixed behavior sequence according to the interaction probability of the family account with each media asset and the real-time behavior sequence of the target family member; and recommends media assets to the target family member based on the target behavior subsequence and the real-time behavior sequence of the target family member. In this method, by extracting the target behavior subsequence of the target family member based on the mixed behavior sequence of the family account and the interaction probability of each media asset, combined with the real-time behavior sequence of the target family member, the problem of difficulty in distinguishing the interests of different family members under a shared family account is solved, enabling personalized recommendations for each family member. Furthermore, by recommending media assets based on the target behavior subsequence and the real-time behavior subsequence, the long-term interests and short-term preferences of family members are considered, thereby further improving the accuracy of media asset recommendations. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart of the media asset recommendation method provided in the embodiments of this application;

[0041] Figure 2a This is one of the execution diagrams of the media asset recommendation method provided in the embodiments of this application;

[0042] Figure 2b This is the second schematic diagram illustrating the execution of the media asset recommendation method provided in the embodiments of this application;

[0043] Figure 3 This is a structural diagram of a media resource recommendation device provided in an embodiment of this application;

[0044] Figure 4 This is a structural diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0045] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0046] This application provides a media asset recommendation method. See also... Figure 1 , Figure 1 This is a flowchart of the media asset recommendation method provided in the embodiments of this application, such as... Figure 1 As shown, it includes the following steps:

[0047] Step 101: Obtain attribute information of multiple media assets that the family account interacted with in historical and current time periods;

[0048] In this step, a family account can be understood as a TV box ID, and a family account may be associated with the media asset interaction behaviors of multiple family members. Media assets can be understood as media resources such as movies, TV series, music songs, short videos, and podcasts. Media asset interaction behaviors can be understood as family users' behaviors such as playing, collecting, and commenting on media assets. The attribute information of media assets includes style information and category information: the style information of media assets can be understood as the artistic or expressive form of the media asset content, such as music genre (rock / classical), film style (science fiction / comedy), etc.; the category information of media assets can be understood as the broader category to which the media asset belongs, such as film and television, music, games, educational content, etc.

[0049] Step 102: Encode the attribute information of each media asset to obtain the encoding vector of each media asset;

[0050] In this step, see Figure 2a Based on the probabilistic coding module, feature encoding of the attribute information of each media asset can be understood as the process of converting the attribute information of the media asset into digital codes that a computer can process and understand. Because computers can typically only process digital data, it is necessary to convert various types of attribute information into digital form. Common encoding methods include one-hot encoding and category encoding. After feature encoding, all the attribute information of each media asset is converted into digital form and combined into a vector. This vector is the encoding vector of the media asset; it contains the key feature information of the media asset and can represent the feature representation of the media asset in the computer.

[0051] Step 103: Based on the encoding vector of each media asset and the mixed behavior sequence of the family account, determine the interaction probability of the family account with each media asset, wherein the mixed behavior sequence represents the media asset sequence of the family account in historical time periods and current time periods;

[0052] In this step, the mixed behavior sequence represents the sequence of media assets interacting with the family account over historical and current periods. For example, the mixed behavior sequence = [Media Asset A (time t1), Media Asset B (time t2), ..., Media Asset B (time t3)]. Machine learning models (such as collaborative filtering or sequence prediction models) can be used to calculate the probability of each media asset being interacted with by the "family as a whole" based on the encoding vector of each media asset and the mixed behavior sequence of the family account. For example, if multiple people in the family have watched movie A, then movie A has a higher probability of being interacted with by the family as a whole.

[0053] Step 104: Based on the interaction probability of the family account for each media asset and the real-time behavior sequence of the target family member, extract the target behavior subsequence of the target family member from the mixed behavior sequence, wherein the target family member is the family member currently using the family account to perform media asset interaction behavior, the target behavior subsequence represents the media asset sequence interacted by the target family member in a historical period, and the real-time behavior sequence of the target family member represents the media asset sequence interacted by the target family member in the current period;

[0054] In this step, the target family member is the member currently using the family account for media interaction, which could be a father, child, or someone else. The target family member's target behavior subsequence represents the media asset records that the target family member has interacted with independently within a historical time period. The target family member's real-time behavior sequence represents the media asset interaction records that the target family member is currently performing or has just completed (e.g., watching a new episode or just playing a song).

[0055] It should be noted that the current time period mentioned in this application refers to a short time interval, during which only one family member is assumed to be interacting with media assets. Therefore, the real-time behavior sequence obtained through the family account during this time period is assumed to be the real-time behavior sequence of the target family member.

[0056] Then, based on the interaction probability of the family account for each media asset and the real-time behavior sequence of the target family members, the target behavior subsequence of the target family members is extracted from the mixed behavior sequence. This can be understood as finding the content that the target family user has liked in the past based on the overall preferences of the family and the current behavior of the target family user. For example, suppose a family frequently watches action movies and comedies, and the current user is watching a comedy. Then, it can be inferred that the user may have watched comedy movies in the past, so the comedy movie can be added to the target behavior subsequence.

[0057] Step 105: Based on the target behavior subsequence of the target family member and the real-time behavior sequence of the target family member, recommend media resources for the target family member. The real-time behavior sequence represents the media resource sequence interacted by the target family member in the current time period.

[0058] In this step, the target family member's target behavior subsequence reflects their past interests and preferences, while their real-time behavior sequence reflects their current interests and preferences. Media asset recommendations for the target family member can be made by combining their past and current interests and preferences.

[0059] In one implementation, attribute information of multiple media assets interacted by a family account in historical and current time periods is obtained; feature encoding is performed on the attribute information of each media asset to obtain an encoding vector for each media asset; based on the encoding vector of each media asset and the mixed behavior sequence of the family account, the interaction probability of the family account with each media asset is determined; according to the interaction probability of the family account with each media asset and the real-time behavior sequence of the target family member, a target behavior sub-sequence of the target family member is extracted from the mixed behavior sequence, wherein the target family member is the family member currently using the family account to interact with media assets, the mixed behavior sequence represents the media asset sequence interacted by the family account in historical and current time periods, and the target behavior sub-sequence represents the media asset sequence interacted by the target family member in historical time periods; based on the target behavior sub-sequence of the target family member and the real-time behavior sequence of the target family member, media assets are recommended to the target family member, wherein the real-time behavior sequence represents the media asset sequence interacted by the target family member in the current time period.

[0060] In this implementation, by combining the mixed behavioral sequence based on the family account and the interaction probability of each media asset with the real-time behavioral sequence of the target family member, the target behavioral subsequence of the target family member is extracted. This solves the problem of difficulty in distinguishing the interests of different family members under the family shared account, and can provide personalized recommendations for each family member. On this basis, media asset recommendations are made based on the target behavioral subsequence and the real-time behavioral subsequence, taking into account the long-term interests and short-term preferences of family members, thereby further improving the accuracy of media asset recommendations.

[0061] Optionally, the step of extracting the target behavior subsequence of the target family member from the mixed behavior sequence based on the interaction probability of the family account with each media asset and the real-time behavior sequence of the target family member includes:

[0062] A complete graph is constructed based on the multiple media assets and the interaction probability of the family account with each media asset. The complete graph includes multiple first nodes and multiple first edges. Any two first nodes are connected by a first edge. The first node is used to represent the media asset. The weight of the first node is the interaction probability of the media asset corresponding to the first node. The weight of the first edge is the js divergence calculated based on the interaction probabilities of the media assets corresponding to the two connected first nodes.

[0063] The first edge with a weight greater than a preset value in the complete graph is broken to obtain k subgraphs, where each subgraph is used to represent the first row of a member as a subsequence;

[0064] Based on the first behavior subsequence of each of the k members and the real-time behavior sequence of the target family member, the target behavior subsequence of the target family member is obtained.

[0065] In one implementation, see further. Figure 2a A complete graph is constructed using a probability consistency module. A complete graph is an undirected graph where all nodes are paired, and each node and edge has a weight. In this application, the first node of the complete graph represents a media asset, and its weight is the family account interaction probability of that media asset; the first edge connects any two first nodes, and its weight is the JS divergence between the interaction probabilities of the media assets corresponding to these two first nodes. JS divergence is a distance metric in information theory used to compare the similarity of two interaction probability distributions. The smaller the value, the closer the distributions are; the larger the value, the greater the difference. Its calculation method is the same as that of JS divergence in related technologies, and will not be repeated here.

[0066] The first edge with a weight greater than a preset value in the complete graph is disconnected, resulting in k subgraphs. For example, suppose the complete graph has four first nodes A, B, C, and D. The weights of A, B, C, and D are 0.8, 0.5, 0.7, and 0.6. The calculated weights of edges AB, BC, AC, BD, CD, and AD are 0.2, 0.1, 0.3, 0.15, 0.1, and 0.05 respectively. Setting the preset value to 0.25, disconnecting edges with weights greater than 0.25 will disconnect AC, while the other edges will be preserved.

[0067] Based on the above processing principles, the complete graph is divided into multiple subgraphs. The JS divergence between the internal nodes of each subgraph is small. Media resources within the same subgraph have similar family preference patterns and can be identified as behavioral characteristics belonging to the same family member. That is, a subgraph can be identified as representing the first behavioral subsequence of a member.

[0068] Then, the similarity between the first row of each subgraph and the real-time behavior sequence of the target family member can be compared to determine the weight of the first row of each subgraph. The higher the similarity with the real-time behavior sequence of the target family member, the greater the weight of the first row of the subgraph. Then, based on the weight of each first row of the subgraph, multiple first row subgraphs are fused to obtain the target behavior subgraph of the target family member.

[0069] In this implementation, by constructing a complete graph and then decomposing it, the consistency of family members' interests and preferences is fully utilized, effectively distinguishing the behaviors of different family members and avoiding the mixing of interests among multiple family members. The mixed behavior sequence corresponding to the family account is decoupled to obtain the first behavior subsequence of individual family members. Then, the target family member's target behavior subsequence is obtained by combining the target family member's real-time behavior sequence and multiple first behavior subsequences. This allows for targeted media asset recommendations to be made to the target family member based on the target behavior subsequence, rather than the mixed behavior sequence corresponding to the family account, thus improving the accuracy of media asset recommendations.

[0070] Optionally, obtaining the target behavior subsequence of the target family member based on the first behavior subsequence of each of the k members and the real-time behavior sequence of the target family member includes:

[0071] The first behavior subsequence of each of the k members and the real-time behavior sequence of the target family member are input into the gating structure;

[0072] Based on the real-time behavior sequence of the target family member, the weight of each first behavior subsequence in the k first behavior subsequences is calculated using the self-attention module of the gating structure;

[0073] The k first-behavior subsequences are fused based on the weight of each first-behavior subsequence to obtain the target behavior subsequence of the target family member.

[0074] In one implementation, see Figure 2b The k first-behavior subsequences and the real-time behavior sequences of the target family members are input into the gating structure, and the attention mechanism of the gating structure (i.e., ...) is used to... Figure 2b In the interest extraction layer (of the model), the model focuses on the relevance of each first behavior subsequence to the real-time behavior sequence of the target family members, thereby assigning a weight to each first behavior subsequence. For example, if the target user is currently watching a science fiction movie, the first behavior subsequence belonging to family members with a strong science fiction preference will be given a higher weight.

[0075] Then, based on the weight of each first-row subsequence, the k first-row subsequences are fused, and after normalization, the target behavior subsequence of the target family member is obtained.

[0076] In this implementation, the self-attention mechanism can identify which first behavior subsequences are most relevant to the target family members. Considering that the interests of the members may be persistent, the weight of the first behavior sequence can be accurately determined by the attention mechanism of the gating structure and the real-time behavior sequence of the target family members. Based on this, the target behavior subsequence that conforms to the historical interests of the target members can be obtained by fusing multiple first behavior sequences based on the weight of the first behavior sequence.

[0077] Optionally, after breaking the first edge with a weight greater than a preset value in the complete graph to obtain k subgraphs, the method further includes:

[0078] A directed graph is created based on the mixed behavior sequence. The directed graph includes multiple second nodes and multiple directed edges. Each directed edge connects two of the multiple second nodes. The second nodes are used to represent the media assets. The interaction time of the media asset corresponding to the starting point of the directed edge is earlier than the interaction time of the media asset corresponding to the ending point of the directed edge. The weight of the directed edge is determined based on a target time interval and the interaction duration of the media asset corresponding to the starting point of the directed edge. The target time interval is the difference between the interaction time of the media asset corresponding to the starting point of the directed edge and the interaction time of the media asset corresponding to the ending point of the directed edge.

[0079] Randomly select k second nodes as starting points from the directed graph, and perform random walks based on the weights of the directed edges in the directed graph to obtain the second row of each of the k members as a subsequence;

[0080] The process of obtaining the target behavior subsequence of the target family member based on the first behavior subsequence of each of the k members and the real-time behavior sequence of the target family member includes:

[0081] Based on the first behavior subsequence of each of the k members, the second behavior subsequence of each of the k members, and the real-time behavior sequence of the target family member, the target behavior subsequence of the target family member is obtained.

[0082] In one implementation, based on the k subgraphs obtained by complete graph segmentation in the above embodiments, k family members with different interests are estimated (it should be noted that, in this application, family members with the same interests can be considered as one family member by default). To further accurately consider the interests and preferences of these k family members, based on the first behavioral sequence of these k family members, a second behavioral sequence of these k family members can be obtained using directed graph techniques. Specifically:

[0083] See also Figure 2a Using a time-series walk decomposition module, a directed graph is created based on the mixed behavioral sequences of family accounts. In the directed graph, a second node represents a media asset. Each directed edge connects two second nodes, and the media asset corresponding to the starting point is interacted with earlier than the one pointing to the ending point. For example, if a user watches movie A (starting point) and then movie B (ending point), there exists an edge from A to B. The weight of a directed edge is the reciprocal of the time interval between the interactions of the two media assets minus the playback duration of the previous media asset. K second nodes are randomly selected as starting points (corresponding to k members). Starting from each starting point, a random walk is performed according to the weight of the second edge, ultimately forming a subsequence of second behaviors for each member. Generally, the interactions generated by the same family member are continuous; that is, the larger the weight of the second edge, the greater the probability of it being selected for a random walk.

[0084] In this implementation, a time-ordered directed graph is constructed, and behavioral features are captured using edge weights. Then, individualized behavioral subsequences are generated using random walks. This method effectively separates the behavioral patterns of mixed users without requiring labeled data.

[0085] Optionally, obtaining the target behavior subsequence of the target family member based on the first behavior subsequence of each of the k members, the second behavior subsequence of each of the k members, and the real-time behavior sequence of the target family member includes:

[0086] The k first-line subsequences, the k second-line subsequences, and the real-time behavior sequence of the target family member are input into the gating structure;

[0087] Based on the real-time behavior sequence of the target family member, the weight of each first behavior subsequence in the k first behavior subsequences and the weight of each second behavior subsequence in the k second behavior subsequences are calculated using the self-attention module of the gating structure.

[0088] Based on the weight of each first-beginning subsequence in the k first-beginning subsequences and the weight of each second-beginning subsequence in the k second-beginning subsequences, the k first-beginning subsequences and the k second-beginning subsequences are fused to obtain the target behavior subsequence corresponding to the target family member.

[0089] In one implementation, see further. Figure 2b The first behavior subsequence of k family members obtained based on complete graph technology, the second behavior subsequence of k family members obtained based on directed graph technology, and the real-time behavior sequence of the target family member are all input into the gating structure. The first and second behavior subsequences can be used to more accurately evaluate the interests and preferences of the k family members.

[0090] Then, through the attention module of the gating structure (i.e. the interest extraction layer in the figure), the model will focus on the correlation between each first behavior subsequence and each second behavior subsequence and the real-time behavior sequence of the target family member, thereby assigning a weight to each first behavior subsequence and each second behavior subsequence.

[0091] Then, based on the weights of each first-row subsequence and each second-row subsequence, the k first-row subsequences and k second-row subsequences are fused to obtain the target behavior subsequences of the target family members.

[0092] In this implementation, the self-attention mechanism can identify which first and second behavior subsequences are most relevant to the target family member. Considering that the member's interests may be persistent, the weights of the first and second behavior sequences can be accurately determined through the gating structure attention mechanism and the real-time behavior sequence of the target family member. Based on this, multiple first behavior sequences and multiple second behavior sequences can be fused based on the weights of the first and second behavior sequences to obtain the target behavior subsequence that conforms to the target member's historical interests.

[0093] Optionally, the weight of the directed edge is the reciprocal of the target value, where the target value is the difference between the target time interval and the interaction duration.

[0094] See Figure 3 , Figure 3 This is a structural diagram of a media asset recommendation device provided in an embodiment of this application. Figure 3 As shown, the media asset recommendation device 300 includes:

[0095] The first acquisition module 301 is used to acquire attribute information of multiple media assets that the family account interacted with in historical and current time periods;

[0096] The second acquisition module 302 is used to perform feature encoding on the attribute information of each media asset to obtain the encoding vector of each media asset;

[0097] The first determining module 303 is used to determine the interaction probability of the family account to each media asset based on the encoding vector of each media asset and the mixed behavior sequence of the family account, wherein the mixed behavior sequence represents the media asset sequence of the family account in historical time periods and current time periods;

[0098] The first extraction module 304 extracts the target behavior subsequence of the target family member from the mixed behavior sequence based on the interaction probability of the family account to each media asset and the real-time behavior sequence of the target family member. The target family member is the family member currently using the family account to perform media asset interaction behavior. The target behavior subsequence represents the media asset sequence interacted by the target family member in a historical period. The real-time behavior sequence represents the media asset sequence interacted by the target family member in the current period.

[0099] The first recommendation module 305 is used to recommend media resources to the target family members based on the target behavior subsequences of the target family members and the real-time behavior sequences of the target family members.

[0100] Optionally, the first extraction module includes:

[0101] The first construction unit is configured to construct a complete graph based on the multiple media assets and the interaction probability of the family account with each media asset. The complete graph includes multiple first nodes and multiple first edges. Any two first nodes among the multiple first nodes are connected by a first edge. The first node is used to represent the media asset. The weight of the first node is the interaction probability of the media asset corresponding to the first node. The weight of the first edge is the js divergence calculated based on the interaction probabilities of the media assets corresponding to the two connected first nodes respectively.

[0102] The first processing unit is used to disconnect the first edge with a weight greater than a preset value in the complete graph to obtain k subgraphs, wherein a subgraph is used to represent the first row of a member as a subsequence;

[0103] The second processing unit is used to obtain the target behavior subsequence of the target family member based on the first behavior subsequence of each of the k members and the real-time behavior sequence of the target family member.

[0104] Optionally, the second processing unit includes:

[0105] The first input subunit is used to input the first behavior subsequence of each of the k members and the real-time behavior sequence of the target family member into the gating structure;

[0106] The first calculation subunit is used to calculate the weight of each of the k first behavior subsequences based on the real-time behavior sequence of the target family member using the self-attention module of the gating structure;

[0107] The first fusion subunit is used to fuse the k first behavior subsequences based on the weight of each first behavior subsequence in the k first behavior subsequences to obtain the target behavior subsequence of the target family member.

[0108] Optionally, the device further includes:

[0109] A first creation module is used to create a directed graph based on the mixed behavior sequence. The directed graph includes multiple second nodes and multiple directed edges. Each directed edge connects two of the multiple second nodes. The second nodes are used to represent the media assets. The interaction time of the media asset corresponding to the starting point of the directed edge is earlier than the interaction time of the media asset corresponding to the ending point of the directed edge. The weight of the directed edge is determined based on a target time interval and the interaction duration of the media asset corresponding to the starting point of the directed edge. The target time interval is the difference between the interaction time of the media asset corresponding to the starting point of the directed edge and the interaction time of the media asset corresponding to the ending point of the directed edge.

[0110] The first processing module is used to randomly select k second nodes from the directed graph as starting points, and perform random walks based on the weights of the directed edges in the directed graph to obtain the second row of each of the k members as a subsequence.

[0111] The second processing unit includes:

[0112] The first processing subunit is used to obtain the target behavior subsequence of the target family member based on the first behavior subsequence of each of the k members, the second behavior subsequence of each of the k members, and the real-time behavior sequence of the target family member.

[0113] Optionally, the first processing subunit is specifically used for:

[0114] The k first-line subsequences, the k second-line subsequences, and the real-time behavior sequence of the target family member are input into the gating structure;

[0115] Based on the real-time behavior sequence of the target family member, the weight of each first behavior subsequence in the k first behavior subsequences and the weight of each second behavior subsequence in the k second behavior subsequences are calculated using the self-attention module of the gating structure.

[0116] Based on the weight of each first-beginning subsequence in the k first-beginning subsequences and the weight of each second-beginning subsequence in the k second-beginning subsequences, the k first-beginning subsequences and the k second-beginning subsequences are fused to obtain the target behavior subsequence corresponding to the target family member.

[0117] Optionally, the weight of the directed edge is the reciprocal of the target value, where the target value is the difference between the target time interval and the interaction duration.

[0118] This application also provides an electronic device. Since the principle by which the electronic device solves the problem is similar to the media asset recommendation method in this application, the implementation of this electronic device can be found in the implementation of the method, and repeated details will not be described again. Figure 4 As shown, the electronic device according to an embodiment of this application includes: a processor 400, configured to read a program from a memory 420 and execute the following processes:

[0119] Obtain attribute information of multiple media assets interacted with by the family account in historical and current time periods;

[0120] The attribute information of each media asset is feature-encoded to obtain the encoding vector of each media asset;

[0121] Based on the encoding vector of each media asset and the mixed behavior sequence of the family account, the interaction probability of the family account with each media asset is determined, wherein the mixed behavior sequence represents the media asset sequence of the family account in historical time periods and current time periods;

[0122] Based on the interaction probability of the family account with each media asset and the real-time behavior sequence of the target family member, the target behavior subsequence of the target family member is extracted from the mixed behavior sequence. The target family member is the family member who is currently using the family account to interact with the media asset. The target behavior subsequence represents the media asset sequence interacted by the target family member in a historical period. The real-time behavior sequence represents the media asset sequence interacted by the target family member in the current period.

[0123] Media asset recommendations are made for the target family members based on the target family member's target behavior subsequence and the target family member's real-time behavior sequence.

[0124] Among them, Figure 4In this context, the bus architecture can include any number of interconnected buses and bridges, specifically linking various circuits of one or more processors represented by processor 400 and memory represented by memory 420 together. The bus architecture can also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides the interface.

[0125] Optionally, the processor 400 is configured to read the program from the memory 420 and execute the following processes:

[0126] A complete graph is constructed based on the multiple media assets and the interaction probability of the family account with each media asset. The complete graph includes multiple first nodes and multiple first edges. Any two first nodes are connected by a first edge. The first node is used to represent the media asset. The weight of the first node is the interaction probability of the media asset corresponding to the first node. The weight of the first edge is the js divergence calculated based on the interaction probabilities of the media assets corresponding to the two connected first nodes.

[0127] The first edge with a weight greater than a preset value in the complete graph is broken to obtain k subgraphs, where each subgraph is used to represent the first row of a member as a subsequence;

[0128] Based on the first behavior subsequence of each of the k members and the real-time behavior sequence of the target family member, the target behavior subsequence of the target family member is obtained.

[0129] Optionally, the processor 400 is configured to read the program from the memory 420 and execute the following processes:

[0130] The first behavior subsequence of each of the k members and the real-time behavior sequence of the target family member are input into the gating structure;

[0131] Based on the real-time behavior sequence of the target family member, the weight of each first behavior subsequence in the k first behavior subsequences is calculated using the self-attention module of the gating structure;

[0132] The k first-behavior subsequences are fused based on the weight of each first-behavior subsequence to obtain the target behavior subsequence of the target family member.

[0133] Optionally, the processor 400 is configured to read the program from the memory 420 and execute the following processes:

[0134] A directed graph is created based on the mixed behavior sequence. The directed graph includes multiple second nodes and multiple directed edges. Each directed edge connects two of the multiple second nodes. The second nodes are used to represent the media assets. The interaction time of the media asset corresponding to the starting point of the directed edge is earlier than the interaction time of the media asset corresponding to the ending point of the directed edge. The weight of the directed edge is determined based on a target time interval and the interaction duration of the media asset corresponding to the starting point of the directed edge. The target time interval is the difference between the interaction time of the media asset corresponding to the starting point of the directed edge and the interaction time of the media asset corresponding to the ending point of the directed edge.

[0135] Randomly select k second nodes as starting points from the directed graph, and perform random walks based on the weights of the directed edges in the directed graph to obtain the second row of each of the k members as a subsequence;

[0136] The process of obtaining the target behavior subsequence of the target family member based on the first behavior subsequence of each of the k members and the real-time behavior sequence of the target family member includes:

[0137] Based on the first behavior subsequence of each of the k members, the second behavior subsequence of each of the k members, and the real-time behavior sequence of the target family member, the target behavior subsequence of the target family member is obtained.

[0138] Optionally, the processor 400 is configured to read the program from the memory 420 and execute the following processes:

[0139] The k first-line subsequences, the k second-line subsequences, and the real-time behavior sequence of the target family member are input into the gating structure;

[0140] Based on the real-time behavior sequence of the target family member, the weight of each first behavior subsequence in the k first behavior subsequences and the weight of each second behavior subsequence in the k second behavior subsequences are calculated using the self-attention module of the gating structure.

[0141] Based on the weight of each first-beginning subsequence in the k first-beginning subsequences and the weight of each second-beginning subsequence in the k second-beginning subsequences, the k first-beginning subsequences and the k second-beginning subsequences are fused to obtain the target behavior subsequence corresponding to the target family member.

[0142] Optionally, the weight of the directed edge is the reciprocal of the target value, where the target value is the difference between the target time interval and the interaction duration.

[0143] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the various processes of the above-described media resource recommendation method embodiments and achieves the same technical effects. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0144] This application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the above-described... Figure 1 The various processes of the method embodiments shown can achieve the same technical effect, and will not be described again here to avoid repetition.

[0145] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0146] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0147] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A media asset recommendation method, characterized in that, The method includes: Obtain attribute information of multiple media assets interacted with by the family account in historical and current time periods; The attribute information of each media asset is feature-encoded to obtain the encoding vector of each media asset; Based on the encoding vector of each media asset and the mixed behavior sequence of the family account, the interaction probability of the family account with each media asset is determined, wherein the mixed behavior sequence represents the media asset sequence of the family account in historical time periods and current time periods; Based on the interaction probability of the family account with each media asset and the real-time behavior sequence of the target family member, the target behavior subsequence of the target family member is extracted from the mixed behavior sequence. The target family member is the family member who is currently using the family account to interact with the media asset. The target behavior subsequence represents the media asset sequence interacted by the target family member in a historical period. The real-time behavior sequence represents the media asset sequence interacted by the target family member in the current period. Media asset recommendations are made for the target family members based on the target family member's target behavior subsequence and the target family member's real-time behavior sequence.

2. The media asset recommendation method according to claim 1, characterized in that, The step of extracting the target behavior subsequence of the target family member from the mixed behavior sequence based on the interaction probability of the family account with each media asset and the real-time behavior sequence of the target family member includes: A complete graph is constructed based on the multiple media assets and the interaction probability of the family account with each media asset. The complete graph includes multiple first nodes and multiple first edges. Any two first nodes are connected by a first edge. The first node is used to represent the media asset. The weight of the first node is the interaction probability of the media asset corresponding to the first node. The weight of the first edge is the js divergence calculated based on the interaction probabilities of the media assets corresponding to the two connected first nodes. The first edge with a weight greater than a preset value in the complete graph is broken to obtain k subgraphs, where each subgraph is used to represent the first row of a member as a subsequence; Based on the first behavior subsequence of each of the k members and the real-time behavior sequence of the target family member, the target behavior subsequence of the target family member is obtained.

3. The media asset recommendation method according to claim 2, characterized in that, The process of obtaining the target behavior subsequence of the target family member based on the first behavior subsequence of each of the k members and the real-time behavior sequence of the target family member includes: The first behavior subsequence of each of the k members and the real-time behavior sequence of the target family member are input into the gating structure; Based on the real-time behavior sequence of the target family member, the weight of each first behavior subsequence in the k first behavior subsequences is calculated using the self-attention module of the gating structure; The k first-behavior subsequences are fused based on the weight of each first-behavior subsequence to obtain the target behavior subsequence of the target family member.

4. The media asset recommendation method according to claim 2, characterized in that, After breaking the first edge with a weight greater than a preset value in the complete graph to obtain k subgraphs, the method further includes: A directed graph is created based on the mixed behavior sequence. The directed graph includes multiple second nodes and multiple directed edges. Each directed edge connects two of the multiple second nodes. The second nodes are used to represent the media assets. The interaction time of the media asset corresponding to the starting point of the directed edge is earlier than the interaction time of the media asset corresponding to the ending point of the directed edge. The weight of the directed edge is determined based on a target time interval and the interaction duration of the media asset corresponding to the starting point of the directed edge. The target time interval is the difference between the interaction time of the media asset corresponding to the starting point of the directed edge and the interaction time of the media asset corresponding to the ending point of the directed edge. Randomly select k second nodes as starting points from the directed graph, and perform random walks based on the weights of the directed edges in the directed graph to obtain the second row of each of the k members as a subsequence. The process of obtaining the target behavior subsequence of the target family member based on the first behavior subsequence of each of the k members and the real-time behavior sequence of the target family member includes: Based on the first behavior subsequence of each of the k members, the second behavior subsequence of each of the k members, and the real-time behavior sequence of the target family member, the target behavior subsequence of the target family member is obtained.

5. The media asset recommendation method according to claim 4, characterized in that, The process of obtaining the target behavior subsequence of the target family member based on the first behavior subsequence of each of the k members, the second behavior subsequence of each of the k members, and the real-time behavior sequence of the target family member includes: The k first-line subsequences, the k second-line subsequences, and the real-time behavior sequence of the target family member are input into the gating structure; Based on the real-time behavior sequence of the target family member, the weight of each first behavior subsequence in the k first behavior subsequences and the weight of each second behavior subsequence in the k second behavior subsequences are calculated using the self-attention module of the gating structure. Based on the weight of each first-beginning subsequence in the k first-beginning subsequences and the weight of each second-beginning subsequence in the k second-beginning subsequences, the k first-beginning subsequences and the k second-beginning subsequences are fused to obtain the target behavior subsequence corresponding to the target family member.

6. The media asset recommendation method according to claim 4, characterized in that, The weight of the directed edge is the reciprocal of the target value, where the target value is the difference between the target time interval and the interaction duration.

7. A media asset recommendation device, characterized in that, The device includes: The first acquisition module is used to acquire attribute information of multiple media assets that the family account interacted with in historical and current time periods; The second acquisition module is used to perform feature encoding on the attribute information of each media asset to obtain the encoding vector of each media asset; The first determining module is used to determine the interaction probability of the family account to each media asset based on the encoding vector of each media asset and the mixed behavior sequence of the family account, wherein the mixed behavior sequence represents the media asset sequence of the family account in historical time periods and current time periods; The first extraction module is used to extract the target behavior subsequence of the target family member from the mixed behavior sequence based on the interaction probability of the family account with each media asset and the real-time behavior sequence of the target family member. The target family member is the family member who is currently using the family account to interact with the media asset. The target behavior subsequence represents the media asset sequence interacted by the target family member in a historical period. The real-time behavior sequence represents the media asset sequence interacted by the target family member in the current period. The first recommendation module is used to recommend media resources to the target family members based on the target behavior subsequences and the real-time behavior sequences of the target family members.

8. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the media asset recommendation method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the media asset recommendation method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, Includes computer instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 6.

Images