A Short Video Recommendation Method Based on Graph Neural Networks

Abstract

This invention discloses a short video recommendation method based on graph neural networks, comprising the following steps: acquiring and preprocessing multi-source recommendation data from a short video platform; dividing short video interest clusters and determining the continuous stimulation intensity of each short video interest cluster for the user; constructing an interest fatigue propagation graph; configuring edge state parameters for the target graph edges in the interest fatigue propagation graph and generating graph propagation data; inputting the graph propagation data into a dual-channel propagation model to aggregate graph information according to different relationship types; determining the net recommendation benefit of each candidate short video in the candidate short video set; sorting the candidate short video set, generating a short video recommendation list, and pushing it to the user terminal. This invention utilizes graph neural networks and multi-source feedback data to characterize interest fatigue, achieving dynamic and accurate short video recommendations, and possessing the advantages of high adaptability and low homogeneous pushes.

Description

Technical Field

[0001] This invention relates to the field of content recommendation technology, and in particular to a short video recommendation method based on graph neural networks. Background Technology

[0002] Short video platforms typically make personalized recommendations based on users' historical behavior data, short video content data, exposure sequence data, and real-time feedback data. They also use graph neural networks to build relationships between users, short videos, content tags, and interactive behaviors to obtain user interest representations, which are then used for ranking candidate short videos.

[0003] Existing recommendation methods primarily focus on positive interest signals such as viewing time, completion rate, likes, and comments, but they are insufficient in capturing the interest fatigue caused by continuous exposure to similar short videos. When a platform continuously pushes short videos with similar themes, semantic tags, background music, visual styles, or emotional types, users may exhibit changes in feedback such as increased scrolling speed and decreased interaction intensity. Existing methods struggle to distinguish between genuine interest enhancement and short-term fatigue decay, easily leading to the continuous push of homogeneous content and reducing the match between recommendation results and the user's current viewing state.

[0004] Therefore, how to provide a short video recommendation method based on graph neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a short video recommendation method based on graph neural networks. This invention utilizes graph neural networks and multi-source feedback data to characterize interest fatigue, thereby achieving dynamic and accurate short video recommendations with the advantages of high adaptability and low homogeneity in push notifications.

[0006] A short video recommendation method based on graph neural networks according to an embodiment of the present invention includes the following steps: Acquire multi-source recommendation data from short video platforms and preprocess it to generate a user short video interaction dataset; Based on the user short video interaction dataset, short video interest clusters are divided, and the continuous stimulation intensity of users for each short video interest cluster is determined. Based on user short video interaction datasets, short video interest clusters, and continuous stimulus intensity, graph nodes and various graph edges are constructed, and interest fatigue propagation graphs are generated according to node relationships. Based on the user short video interaction dataset and continuous stimulus intensity, edge state parameters are configured for the target graph edges in the interest fatigue propagation graph, and graph propagation data is generated by combining node features and relationship types. Input the graph propagation data into the dual-channel propagation model, aggregate the graph information according to different relationship types, and generate user interest enhancement representation and user fatigue decay representation. Obtain a set of candidate short videos, and determine the net recommendation benefit of each candidate short video in the set based on user interest enhancement representation, user fatigue decay representation, and continuous stimulus intensity. The candidate short video set is sorted according to the net recommendation revenue, and a short video recommendation list is generated based on the sorting results and pushed to the user's terminal.

[0007] Optionally, the multi-source recommendation data includes user historical behavior data, short video content data, exposure sequence data, and real-time feedback data. The preprocessing includes data cleaning, user identification unification, short video identification unification, timestamp unification, exposure order rearrangement, behavior record deduplication, abnormal feedback removal, and multi-source data alignment.

[0008] Optionally, the segmentation of short video interest clusters and the determination of continuous stimulus intensity specifically include: Short videos are clustered and divided based on the short video content data in the user short video interaction dataset to generate short video interest clusters; Configure interest cluster identifiers for short video interest clusters and establish the content ownership relationship between short video identifiers and interest cluster identifiers; Based on the content attribution relationship, the short video identifiers in the exposure sequence data are mapped to the corresponding short video interest clusters to generate interest cluster exposure sequences; Based on the exposure sequence of interest clusters, the continuous exposure features corresponding to each short video interest cluster are statistically analyzed, and the continuous exposure features are weighted and normalized to generate the continuous push status of interest clusters. The changes in viewing feedback for each short video interest cluster are normalized based on real-time feedback data to generate user feedback change status. The continuous push status of interest clusters and the user feedback change status are converted to a unified value range and then weighted and summed to obtain the continuous stimulation intensity of users for each short video interest cluster.

[0009] Optionally, the generation of the interest fatigue propagation map specifically includes: Based on user identifiers, short video identifiers, exposure order, short video interest clusters, and real-time feedback data in the user short video interaction dataset, construct graph nodes in the interest fatigue propagation graph; Establish behavioral feedback edges between user nodes and short video nodes based on users' historical behavior data, and record users' interaction relationships with short videos; Establish content ownership edges between short video nodes and interest cluster nodes based on short video interest clusters; Based on the user short video interaction dataset and short video interest clusters, establish interest enhancement edges between user nodes and interest cluster nodes; Based on continuous stimulus intensity and real-time feedback data, a fatigue decay edge is established between user nodes and interest cluster nodes. Based on the exposure sequence data and the intensity of continuous stimuli, establish continuous stimulus edges between exposure sequence nodes and interest cluster nodes; Based on real-time feedback data, feedback write-back edges are established between feedback status nodes and user nodes. Based on graph nodes, behavioral feedback edges, content ownership edges, interest enhancement edges, fatigue decay edges, continuous stimulus edges, and feedback write-back edges, an interest fatigue propagation graph is generated according to node relationships.

[0010] Optionally, the generation of the graph propagation data specifically includes: Configure interest enhancement weights based on the user's historical behavior data corresponding to the interest enhancement edges; Configure fatigue decay weights based on the continuous stimulus intensity and real-time feedback data corresponding to the fatigue decay edge. Configure exposure interval weights based on the exposure time, exposure order, and exposure interval of similar short videos in the exposure sequence data; Configure the feedback change weight based on real-time feedback data, and configure the continuous push density weight based on continuous exposure characteristics; Interest enhancement weight and fatigue decay weight are written into the interest enhancement edge and fatigue decay edge, which are the core propagation edges, respectively. Exposure interval weight and continuous push density weight are written into the continuous stimulus edge. Feedback change weight is written into the feedback write-back edge. Graph propagation data is generated by combining behavioral feedback edge, content ownership edge, node features and relationship type.

[0011] Optionally, the generation of the user interest enhancement representation and the user fatigue decay representation specifically includes: The graph propagation data is input into the dual-channel propagation model. The node features of each graph node in the interest fatigue propagation graph are read through the node feature input layer, and the relationship type and edge state parameters corresponding to the edge of the target graph are read. The dual-channel propagation model includes an interest enhancement propagation branch and a fatigue decay propagation branch. Both the interest enhancement propagation branch and the fatigue decay propagation branch adopt a relation graph convolutional network structure. By dividing the target graph edges in the graph propagation data into branches through a relation type partitioning layer, we obtain an interest-enhancing relation set and a fatigue-decaying relation set. By using the relation-specific weight transformation layer in the interest enhancement propagation branch and the fatigue decay propagation branch, relation-specific transformations are performed on the node features within the corresponding propagation range according to the relation type and edge state parameters to obtain interest enhancement propagation information and fatigue decay propagation information. By using a relational neighbor aggregation layer, interest enhancement propagation information and fatigue decay propagation information are aggregated by relational sub-relationship to generate positive interest aggregation features and fatigue decay aggregation features. The self-loop feature retention layer preserves the node features of the target user node, and the positive interest aggregation feature and fatigue decay aggregation feature corresponding to the target user node are fused with the node features of the target user node respectively. The node representation output layer outputs the user interest enhancement representation and the user fatigue decay representation.

[0012] Optionally, the determination of the recommended net income specifically includes: Obtain a set of candidate short videos and read the short video identifier and content affiliation of each candidate short video. Map each candidate short video to the corresponding short video interest cluster and generate the candidate affiliation relationship between the candidate short video and the short video interest cluster. Based on the candidate affiliation relationship, the target user’s interest enhancement component for the short video interest cluster to which each candidate short video belongs is extracted from the user interest enhancement representation, and the real interest enhancement state value is calculated based on the interest enhancement component. Based on the candidate attribution relationship, the fatigue decay component of the target user to the short video interest cluster to which each candidate short video belongs is extracted from the user fatigue decay representation, and the interest fatigue state value is obtained by combining the continuous stimulus intensity and real-time feedback data. Subtracting the interest fatigue value from the genuine interest enhancement value yields the net recommendation benefit for each candidate short video.

[0013] Optionally, the short video recommendation list includes the recommendation position, short video identifier, short video theme, semantic tag, author identifier, and net recommendation revenue for each candidate short video.

[0014] The beneficial effects of this invention are: This invention acquires user historical behavior data, short video content data, exposure sequence data, and real-time feedback data, and then segments short videos into interest clusters. It can group short videos with similar themes, semantic tags, background music, visual styles, or emotional types into corresponding interest clusters. Based on the exposure sequence statistics of these interest clusters, it statistically analyzes continuous exposure characteristics and determines the intensity of continuous stimuli. This allows the short video recommendation process to go beyond relying solely on positive behavioral signals such as clicks, completions, likes, and favorites. It can also quantify the degree of stimulation caused by the continuous push of similar short videos to the user's viewing state, thereby improving the ability to perceive changes in the user's true interests.

[0015] This invention constructs an interest fatigue propagation graph, linking user nodes, short video nodes, interest cluster nodes, exposure sequence nodes, and feedback state nodes. It expresses the interactions between different recommendation factors through behavioral feedback edges, content attribution edges, interest enhancement edges, fatigue decay edges, continuous stimulation edges, and feedback write-back edges. By configuring edge state parameters for the target graph edges, interest enhancement signals, continuous exposure stimulation signals, and real-time feedback change signals can propagate within the graph structure according to different relationship types. This avoids the traditional recommendation model's simple accumulation of all positive historical behaviors and reduces the risk of continuously pushing homogeneous content due to excessive amplification of historical interests.

[0016] This invention generates user interest enhancement representations and user fatigue decay representations through a dual-channel propagation model, and determines the net recommendation benefit of each candidate short video based on the actual interest enhancement state and interest fatigue state. This allows for the differentiation between a user's stable interest in a certain type of short video and the fatigue decay state caused by short-term repeated stimulation. During the candidate short video ranking stage, benefits are deducted for short videos with excessively strong continuous stimulation and deteriorating real-time feedback, making the short video recommendation list more consistent with the user's current viewing state and improving the diversity, adaptability, and user experience of the recommendation results. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a short video recommendation method based on graph neural networks proposed in this invention; Figure 2 This is a flowchart illustrating the graph propagation data generation process of a short video recommendation method based on graph neural networks proposed in this invention. Figure 3 This is a flowchart illustrating the graph information aggregation process of a short video recommendation method based on graph neural networks proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-3 A short video recommendation method based on graph neural networks includes the following steps: Acquire multi-source recommendation data from short video platforms and preprocess it to generate a user short video interaction dataset; Based on the user short video interaction dataset, short video interest clusters are divided, and the continuous stimulation intensity of users for each short video interest cluster is determined. Based on user short video interaction datasets, short video interest clusters, and continuous stimulus intensity, graph nodes and various graph edges are constructed, and interest fatigue propagation graphs are generated according to node relationships. Based on the user short video interaction dataset and continuous stimulus intensity, edge state parameters are configured for the target graph edges in the interest fatigue propagation graph, and graph propagation data is generated by combining node features and relationship types. Input the graph propagation data into the dual-channel propagation model, aggregate the graph information according to different relationship types, and generate user interest enhancement representation and user fatigue decay representation. Obtain a set of candidate short videos, and determine the net recommendation benefit of each candidate short video in the set based on user interest enhancement representation, user fatigue decay representation, and continuous stimulus intensity. The candidate short video set is sorted according to the net recommendation revenue, and a short video recommendation list is generated based on the sorting results and pushed to the user's terminal.

[0020] In this embodiment, the multi-source recommendation data includes user historical behavior data, short video content data, exposure sequence data, and real-time feedback data. User historical behavior data includes viewing duration, completion rate, likes, comments, shares, favorites, and swipe time. Short video content data includes short video theme, semantic tags, author identifiers, background music, visual style, and mood type. Exposure sequence data includes short video exposure time, exposure order, number of consecutive exposures of similar short videos, and exposure interval of similar short videos. Real-time feedback data includes changes in viewing duration, swipe speed, and interaction intensity. Preprocessing includes data cleaning, user identifier unification, short video identifier unification, timestamp unification, exposure order rearrangement, behavior record deduplication, abnormal feedback removal, and multi-source data alignment.

[0021] In this embodiment, the division of short video interest clusters and the determination of continuous stimulus intensity specifically include: Short videos are clustered based on their content data from the user short video interaction dataset to generate short video interest clusters. Clustering is performed based on the similarity of short video themes, semantic tag overlap, background music similarity, visual style similarity, and emotion type consistency. Specifically, theme similarity is determined by whether the two short videos share the same theme level and the distance between those levels; semantic tag overlap is determined by the proportion of identical semantic tags to all unique semantic tags in the two short videos; background music similarity is determined by the matching results of background music tags when the two short videos have the same tags, and by the degree of audio fingerprint matching when the tags are different; visual style similarity is determined by the similarity of color features, brightness features, and shot transition frequency in the two short videos; and emotion type consistency is determined by the proportion of identical emotion types to all unique emotion types in the two short videos. Configure interest cluster identifiers for short video interest clusters and establish content ownership relationships between short video identifiers and interest cluster identifiers; the content ownership relationship is used to record the short video identifier, short video theme, semantic tags, background music, visual style and mood type contained in each short video interest cluster; Based on the content attribution relationship, the short video identifiers in the exposure sequence data are mapped to the corresponding short video interest clusters to generate interest cluster exposure sequences; Based on the exposure sequence of interest clusters, the continuous exposure features corresponding to each short video interest cluster are statistically analyzed, and the continuous exposure features are weighted and normalized to generate the continuous push status of interest clusters. The weighted normalization includes reverse normalization of the exposure interval of similar short videos and forward normalization of the remaining continuous exposure features, and then weighted summation based on the normalized continuous exposure features. The continuous exposure features include the number of consecutive exposures of similar short videos, the exposure interval of similar short videos, the continuous push density of similar background music, the continuous push density of similar visual styles, and the continuous push density of similar emotion types. Among them, the number of consecutive exposures of similar short videos is determined by the number of short videos that continuously belong to the same short video interest cluster in the exposure sequence; similar short videos... The frequency of exposure intervals is determined based on the time difference between the exposure times of two adjacent short videos that satisfy at least one similarity relationship in terms of short video theme, background music, visual style, or emotional type; the continuous push density of similar background music is determined based on the proportion of the number of short videos with similar background music exposed in a continuous exposure segment to the total number of short videos exposed in the continuous exposure segment; the continuous push density of similar visual style is determined based on the proportion of the number of short videos with similar visual style exposed in a continuous exposure segment to the total number of short videos exposed in the continuous exposure segment; the continuous push density of similar emotional type is determined based on the proportion of the number of short videos with similar emotional type exposed in a continuous exposure segment to the total number of short videos exposed in the continuous exposure segment. The changes in viewing feedback for each short video interest cluster are normalized based on real-time feedback data to generate user feedback change status. The continuous push status of interest clusters and the user feedback change status are converted to a unified value range and then weighted and summed to obtain the continuous stimulation intensity of users for each short video interest cluster.

[0022] In this embodiment, the generation of the interest fatigue propagation map specifically includes: Based on user identifiers, short video identifiers, exposure order, short video interest clusters, and real-time feedback data in the user short video interaction dataset, a graph node is constructed in the interest fatigue propagation graph; the graph node includes user nodes, short video nodes, interest cluster nodes, exposure sequence nodes, and feedback status nodes; Based on historical user behavior data, behavioral feedback edges are established between user nodes and short video nodes, and user interaction relationships with short videos are recorded. Behavioral feedback edges are used to represent user viewing time, completion rate, likes, comments, shares, favorites, and swipe time for short videos. Content attribution edges are established between short video nodes and interest cluster nodes based on short video interest clusters. Content attribution edges are used to represent the attribution relationship between short video nodes and interest cluster nodes. When the short video corresponding to a short video node is classified into a short video interest cluster, a content attribution edge is established between the short video node and the corresponding interest cluster node, and the similarity of short video theme, semantic tag overlap, background music similarity, visual style similarity, and emotional type consistency are used as edge attributes of the content attribution edge. Based on the user short video interaction dataset and short video interest clusters, interest enhancement edges are established between user nodes and interest cluster nodes. Interest enhancement edges are used to represent the positive interest accumulation of user nodes on interest cluster nodes. When a user generates a completion result, like result, comment result, forward result, or favorite result for a short video within the same short video interest cluster, an interest enhancement edge is established between the user node and the corresponding interest cluster node, and the viewing duration, completion result, like result, comment result, forward result, and favorite result are used as edge attributes of the interest enhancement edge. Based on continuous stimulus intensity and real-time feedback data, a fatigue decay edge is established between user nodes and interest cluster nodes. The fatigue decay edge is used to characterize the decline in user node interest in interest cluster nodes. When the value corresponding to the continuous stimulus intensity of the user for the same short video interest cluster is greater than zero, and the real-time feedback data shows a decrease in viewing time, an increase in scrolling speed, or a decrease in interaction intensity, a fatigue decay edge is established between the user node and the corresponding interest cluster node, and the continuous stimulus intensity, changes in viewing time, changes in scrolling speed, and changes in interaction intensity are used as edge attributes of the fatigue decay edge. Based on exposure sequence data and continuous stimulus intensity, continuous stimulus edges are established between exposure sequence nodes and interest cluster nodes. Continuous stimulus edges are used to characterize the continuous stimulus relationship between exposure sequence nodes and interest cluster nodes. When short videos belonging to the same short video interest cluster appear consecutively in the exposure sequence data, or when adjacent exposed short videos satisfy at least one similarity relationship among similar themes, similar background music, similar visual styles, or similar emotional types, continuous stimulus edges are established between exposure sequence nodes and corresponding interest cluster nodes. The short video exposure time, exposure order, number of consecutive exposures of similar short videos, exposure interval of similar short videos, and continuous stimulus intensity are used as edge attributes of the continuous stimulus edges. Feedback write-back edges are established between feedback status nodes and user nodes based on real-time feedback data. An interest fatigue propagation graph is generated based on the relationships between graph nodes, behavioral feedback edges, content attribution edges, interest enhancement edges, fatigue decay edges, continuous stimulation edges, and feedback write-back edges. Feedback write-back edges are used to represent the state write-back relationship from feedback status nodes to user nodes. When the real-time feedback data corresponding to the feedback status node belongs to the target user, a feedback write-back edge is established between the feedback status node and the user node corresponding to the target user. Changes in viewing time, scrolling speed, and interaction intensity are used as edge attributes of the feedback write-back edge. The node-based association includes: associating behavioral feedback edges with user nodes and short video nodes; associating content attribution edges with short video nodes and interest cluster nodes; associating interest enhancement edges and fatigue decay edges with user nodes and interest cluster nodes; associating continuous stimulation edges with exposure sequence nodes and interest cluster nodes; and associating feedback write-back edges with feedback status nodes and user nodes.

[0023] In this embodiment, the generation of graph propagation data specifically includes: The interest enhancement weight is configured based on the user's historical behavior data corresponding to the interest enhancement edge; the interest enhancement weight increases with the increase of viewing time, completion of the broadcast, likes, comments, reposts, and collections. The fatigue decay weight is configured based on the continuous stimulus intensity and real-time feedback data corresponding to the fatigue decay edge; the fatigue decay weight increases with the increase of continuous stimulus intensity, decrease of viewing time, increase of sliding speed and decrease of interaction intensity; The exposure interval weight is configured based on the exposure time, exposure order, and exposure interval of similar short videos in the exposure sequence data; the exposure interval weight increases as the exposure interval of similar short videos shortens and decreases as the exposure interval of similar short videos lengthens; The feedback change weight is configured based on real-time feedback data, and the continuous push density weight is configured based on continuous exposure characteristics. The feedback change weight increases with the increase of the decrease in viewing time, the increase in scrolling speed, and the decrease in interaction intensity. The continuous push density weight increases with the increase of the continuous push density of similar background music, similar visual style, and similar emotion type. Interest enhancement weight and fatigue decay weight are written into the interest enhancement edge and fatigue decay edge, which are the core propagation edges, respectively. Exposure interval weight and continuous push density weight are written into the continuous stimulus edge. Feedback change weight is written into the feedback write-back edge. Graph propagation data is generated by combining behavioral feedback edge, content ownership edge, node features and relationship type.

[0024] In this embodiment, the generation of user interest enhancement representation and user fatigue decay representation specifically includes: The graph propagation data is input into the dual-channel propagation model. The node features of each graph node in the interest fatigue propagation graph are read through the node feature input layer, and the relationship type and edge state parameters corresponding to the edge of the target graph are read. The dual-channel propagation model includes an interest enhancement propagation branch and a fatigue decay propagation branch. Both the interest enhancement propagation branch and the fatigue decay propagation branch adopt a relation graph convolutional network structure. By dividing the target graph edges in the graph propagation data into branches using a relation type partitioning layer, we obtain an interest enhancement relation set and a fatigue decay relation set. The interest enhancement relation set includes interest enhancement edges, behavioral feedback edges, and content attribution edges, which are used to limit the positive interest propagation range between user nodes, short video nodes, and interest cluster nodes. The fatigue decay relation set includes fatigue decay edges, continuous stimulus edges, and feedback write-back edges, which are used to limit the continuous exposure decay propagation range between user nodes, interest cluster nodes, exposure sequence nodes, and feedback status nodes. By using the relation-specific weight transformation layer in the interest enhancement propagation branch and the fatigue decay propagation branch, relation-specific transformations are performed on the node features within the corresponding propagation range according to the relation type and edge state parameters to obtain interest enhancement propagation information and fatigue decay propagation information. The relation-specific transformation includes: identifying the relation type of the edge in the current graph to be propagated, and calling the transformation parameters corresponding to the relation type; normalizing the edge state parameters of the current graph to be propagated to obtain the propagation adjustment coefficient, which is used to characterize the degree of intensity adjustment of the current graph edge's propagation of neighbor node features to the current node to be updated; inputting the neighbor node features into the transformation parameters corresponding to the relation type, mapping them to relation propagation features of the same dimension as the current node to be updated; using the propagation adjustment coefficient to perform weighted adjustment on the relation propagation features to obtain edge constraint propagation features; propagating the edge constraint propagation features along the current graph edge to be propagated to the current node to be updated; edge constraint propagation features belonging to the interest-enhancing relation set form interest-enhancing propagation information, and edge constraint propagation features belonging to the fatigue-decay relation set form fatigue-decay propagation information; By using a relational neighbor aggregation layer, interest enhancement propagation information and fatigue decay propagation information are aggregated by relational sub-relationship to generate positive interest aggregation features and fatigue decay aggregation features. Relational aggregation includes: using the current node to be updated as the aggregation center, dividing the edge constraint propagation features passed to the current node to be updated into relation feature groups corresponding to each relation type according to the relation type; adding the edge constraint propagation features in each relation feature group item by item to obtain the relation-internal aggregation features of the corresponding relation type; according to the branch affiliation corresponding to the relation type, inputting the relation-internal aggregation features belonging to the interest enhancement propagation branch into the interest enhancement aggregation channel, and inputting the relation-internal aggregation features belonging to the fatigue decay propagation branch into the fatigue decay aggregation channel; in the interest enhancement aggregation channel, adding the relation-internal aggregation features corresponding to different relation types item by item to obtain the positive interest aggregation features; in the fatigue decay aggregation channel, adding the relation-internal aggregation features corresponding to different relation types item by item to obtain the fatigue decay aggregation features. The self-loop feature retention layer preserves the target user node's own node features, and fuses the positive interest aggregation features and fatigue decay aggregation features corresponding to the target user node with its own node features respectively. The node representation output layer outputs user interest enhancement representations and user fatigue decay representations. Specifically: the self-loop feature retention layer uses the target user node as the retention object, reads the node features of the target user node at the input of the current propagation layer, and generates self-loop retained features; it performs a same-dimensional mapping on the self-loop retained features, the positive interest aggregation features corresponding to the target user node, and the fatigue decay aggregation features corresponding to the target user node; and it then merges the same-dimensionally mapped self-loop retained features with… The positive interest aggregation features corresponding to the target user nodes after same-dimensional mapping are added item by item, and the addition result is subjected to non-linear activation and normalization to obtain the interest-enhanced user node representation. The self-loop-preserving features after same-dimensional mapping are added item by item to the fatigue decay aggregation features corresponding to the target user nodes after same-dimensional mapping, and the addition result is subjected to non-linear activation and normalization to obtain the fatigue decay user node representation. The node representation output layer is used to output map the interest-enhanced user node representation to generate the user interest-enhanced representation, and the node representation output layer is used to output map the fatigue decay user node representation to generate the user fatigue decay representation.

[0025] In this embodiment, the determination of the recommended net income specifically includes: Obtain a candidate short video set and read the short video identifier and content affiliation of each candidate short video. Map each candidate short video to the corresponding short video interest cluster and generate the candidate affiliation relationship between the candidate short video and the short video interest cluster. The candidate short video set includes the short video identifier, short video theme, semantic tag, author identifier and content affiliation relationship of the short video to be recommended. Based on the candidate affiliation relationship, the target user’s interest enhancement component for the short video interest cluster to which each candidate short video belongs is extracted from the user interest enhancement representation, and the real interest enhancement state value is calculated based on the interest enhancement component. The calculation of the true interest enhancement state value includes: taking the short video interest cluster to which the candidate short video belongs as the interest state extraction object, and determining the interest cluster identifier corresponding to the interest state extraction object; matching the corresponding interest cluster representation position in the user interest enhancement representation according to the interest cluster identifier; extracting the target user's representation value of the interest state extraction object from the interest cluster representation position to obtain the interest enhancement component; calculating the interest enhancement components corresponding to each candidate short video in the current candidate short video set to obtain the minimum and maximum values ​​of the interest enhancement components; when the maximum and minimum values ​​of the interest enhancement components are the same, determining the initial interest enhancement state value to be the same value as the interest enhancement component; when the maximum and minimum values ​​of the interest enhancement components are different, performing interval normalization on the interest enhancement components according to the minimum and maximum values ​​of the interest enhancement components to obtain the initial interest enhancement state value; determining the candidate affiliation strength between the candidate short video and the interest state extraction object according to the candidate affiliation relationship; and multiplying the initial interest enhancement state value and the candidate affiliation strength item by item to obtain the true interest enhancement state value. The calculation of candidate attribution strength includes: extracting content attribution features between candidate short videos and interest state extraction objects based on candidate attribution relationships; normalizing the short video theme similarity, semantic tag overlap, background music similarity, visual style similarity, and emotion type consistency in the content attribution features; summing the normalized content attribution feature values ​​one by one and dividing by the number of content attribution feature values ​​to obtain the candidate attribution strength. Based on the candidate attribution relationship, the fatigue decay component of the target user to the short video interest cluster to which each candidate short video belongs is extracted from the user fatigue decay representation, and the interest fatigue state value is obtained by combining the continuous stimulus intensity and real-time feedback data. The process of obtaining the interest fatigue state value includes: taking the short video interest cluster to which the candidate short video belongs as the fatigue state extraction object, and determining the interest cluster identifier corresponding to the fatigue state extraction object; matching the corresponding interest cluster representation position in the user fatigue decay representation according to the interest cluster identifier; extracting the representation value of the target user for the fatigue state extraction object from the interest cluster representation position to obtain the fatigue decay component; statistically analyzing the fatigue decay components corresponding to each candidate short video in the current candidate short video set to obtain the minimum and maximum values ​​of the fatigue decay components; when the maximum and minimum values ​​of the fatigue decay components are the same, the fatigue state value is determined to be the same value as the fatigue decay component; when the maximum and minimum values ​​of the fatigue decay components are not the same, the fatigue state value is determined according to the minimum and maximum values ​​of the fatigue decay components. The fatigue attenuation component is normalized within an interval to obtain a fatigue state value. The continuous stimulus intensity corresponding to the fatigue state extraction object is determined based on candidate affiliation relationships, and the continuous stimulus intensity of the short video interest cluster to which each candidate short video belongs in the current candidate short video set is normalized within an interval to obtain a continuous stimulus state value. The degree of decrease in viewing time, increase in scrolling speed, and decrease in interaction intensity of the target user for the fatigue state extraction object are determined based on real-time feedback data and converted into fatigue direction feedback values. The fatigue direction feedback values ​​are normalized within an interval, and the normalized fatigue direction feedback values ​​are summed one by one and divided by the number of fatigue direction feedback values ​​to obtain the feedback fatigue state value. The fatigue state value, continuous stimulus state value, and feedback fatigue state value are summed one by one and divided by the number of state values ​​to obtain the interest fatigue state value. The conversion of fatigue direction feedback values ​​includes: taking the absolute value of the portion of the viewing duration change value that is less than zero to obtain the viewing duration decrease value, and setting the portion of the viewing duration change value that is greater than or equal to zero to zero; retaining the portion of the gliding speed change value that is greater than zero to obtain the gliding speed increase value, and setting the portion of the gliding speed change value that is less than or equal to zero to zero; taking the absolute value of the portion of the interaction intensity change value that is less than zero to obtain the interaction intensity decrease value, and setting the portion of the interaction intensity change value that is greater than or equal to zero to zero; and determining the viewing duration decrease value, gliding speed increase value, and interaction intensity decrease value as fatigue direction feedback values ​​respectively. Subtracting the interest fatigue value from the genuine interest enhancement value yields the net recommendation benefit for each candidate short video.

[0026] In this embodiment, the short video recommendation list includes the recommendation position, short video identifier, short video theme, semantic tag, author identifier, and net recommendation revenue for each candidate short video, wherein the recommendation position is determined according to the sorting number in the candidate short video ranking results.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to the evening content recommendation scenario of a short video platform. During peak user activity periods before and after holidays, the platform tends to continuously push short videos with similar themes, background music, and visual styles to the same user. Although some users may have previously watched, liked, or saved such content, after continuous viewing, they tend to scroll faster, watch for shorter periods, and have less willingness to interact. This causes the existing recommendation method to continue amplifying past interests, and the recommendation list gradually becomes homogenized.

[0028] In this scenario, the platform acquires user historical behavior data, short video content data, exposure sequence data, and real-time feedback data. It then segments short videos into interest clusters, categorizing them based on similar themes, semantic tags, background music, visual styles, and emotional types. The system generates an exposure sequence for each interest cluster based on the user's recent exposure sequence, determining whether similar content appears continuously, and assessing the intensity of continuous stimulation by considering changes in viewing time, scrolling speed, and interaction intensity. Subsequently, the system constructs an interest fatigue propagation graph, establishing connections between the user, short video, interest cluster, exposure sequence, and feedback status. Interest enhancement edges represent the user's true preferences, while fatigue decay edges, continuous stimulation edges, and feedback write-back edges represent the fatigue changes caused by repeated stimulation. A dual-channel propagation model is then used to obtain the user's interest enhancement representation and user fatigue decay representation, respectively.

[0029] Platform backend logs show that during online applications by the same user group within the same recommendation period and region, the recommendation system can identify short video interest clusters where users still have basic interest but their feedback weakens in a short period of time. It also reduces the net benefit of recommending similar content when ranking candidate short videos. The issue of consecutive appearances of similar content in the recommendation list is alleviated, and the matching degree between current user feedback and recommendation results is more stable. This indicates that the invention can effectively distinguish between genuine interest enhancement and short-term interest fatigue, improving the adaptability of short video recommendations.

[0030] Table 1. Performance Comparison of the Invention and Traditional Short Video Recommendation Methods

[0031] As can be clearly seen from Table 1, the method of the present invention is superior to the traditional method in many indicators.

[0032] Under the average viewing duration metric, the traditional recommendation method achieves 41.8 seconds, while the method of this invention achieves 43.9 seconds, representing an improvement of 5.02%. This result demonstrates that the present invention can reduce user fatigue caused by continuous push of similar short videos. This is because the present invention identifies repetitive stimulus states by short video interest clusters and continuous stimulus intensity, and suppresses candidate short videos with strong fatigue in the net recommendation benefit, allowing users to receive content that is more in line with their current viewing state.

[0033] Under the completion rate metric, the traditional recommendation method achieves 62.4%, while the method of this invention achieves 64.7%, an improvement of 2.30 percentage points (pp). This improvement stems from the fact that this invention simultaneously generates user interest enhancement representations and user fatigue decay representations, which can distinguish between users' true interests and short-term fatigue states, avoiding the simple amplification of historical positive behavior, and making the ranking of candidate short videos more in line with users' current viewing intentions.

[0034] Under the rapid scroll-away rate metric, the traditional recommendation method yields 28.6%, while the method of this invention yields 26.9%, a reduction of 1.70 percentage points (pp). This reduction indicates that fewer users immediately scroll away after encountering inappropriate content. This is because fatigue decay edges, continuous stimulation edges, and feedback write-back edges can incorporate changes in scroll-away speed, viewing duration, and interaction intensity into the graph propagation process, thus lowering the ranking position of similar content with deteriorating feedback.

[0035] Under the metric of continuous exposure of homogeneous content, the traditional recommendation method achieves 34.5%, while the method of this invention achieves 31.8%, a reduction of 2.70 percentage points (pp). This result indicates that the present invention applies fatigue constraints to short videos with similar themes, semantic tags, background music, visual styles, and emotional types, reducing the continuous appearance of similar content. The recommendation click-through rate increased from 9.6% to 10.1%, an improvement of 0.50 pp, demonstrating that the recommendation list maintains good interest matching ability even after reducing homogeneous pushes.

[0036] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A short video recommendation method based on graph neural networks, characterized in that, Includes the following steps: Acquire multi-source recommendation data from short video platforms and preprocess it to generate a user short video interaction dataset; Based on the user short video interaction dataset, short video interest clusters are divided, and the continuous stimulation intensity of users for each short video interest cluster is determined. Based on user short video interaction datasets, short video interest clusters, and continuous stimulus intensity, graph nodes and various graph edges are constructed, and interest fatigue propagation graphs are generated according to node relationships. Based on the user short video interaction dataset and continuous stimulus intensity, edge state parameters are configured for the target graph edges in the interest fatigue propagation graph, and graph propagation data is generated by combining node features and relationship types. Input the graph propagation data into the dual-channel propagation model, aggregate the graph information according to different relationship types, and generate user interest enhancement representation and user fatigue decay representation. Obtain a set of candidate short videos, and determine the net recommendation benefit of each candidate short video in the set based on user interest enhancement representation, user fatigue decay representation, and continuous stimulus intensity. The candidate short video set is sorted according to the net recommendation revenue, and a short video recommendation list is generated based on the sorting results and pushed to the user's terminal.

2. The short video recommendation method based on graph neural networks according to claim 1, characterized in that, The multi-source recommendation data includes user historical behavior data, short video content data, exposure sequence data, and real-time feedback data. The preprocessing includes data cleaning, user identification unification, short video identification unification, timestamp unification, exposure order rearrangement, behavior record deduplication, abnormal feedback removal, and multi-source data alignment.

3. The short video recommendation method based on graph neural networks according to claim 1, characterized in that, The division of short video interest clusters and the determination of continuous stimulus intensity specifically include: Short videos are clustered and divided based on the short video content data in the user short video interaction dataset to generate short video interest clusters; Configure interest cluster identifiers for short video interest clusters and establish the content ownership relationship between short video identifiers and interest cluster identifiers; Based on the content attribution relationship, the short video identifiers in the exposure sequence data are mapped to the corresponding short video interest clusters to generate interest cluster exposure sequences; Based on the exposure sequence of interest clusters, the continuous exposure features corresponding to each short video interest cluster are statistically analyzed, and the continuous exposure features are weighted and normalized to generate the continuous push status of interest clusters. The changes in viewing feedback for each short video interest cluster are normalized based on real-time feedback data to generate user feedback change status. The continuous push status of interest clusters and the user feedback change status are converted to a unified value range and then weighted and summed to obtain the continuous stimulation intensity of users for each short video interest cluster.

4. The short video recommendation method based on graph neural networks according to claim 1, characterized in that, The generation of the interest fatigue propagation map specifically includes: Based on user identifiers, short video identifiers, exposure order, short video interest clusters, and real-time feedback data in the user short video interaction dataset, construct graph nodes in the interest fatigue propagation graph; Establish behavioral feedback edges between user nodes and short video nodes based on users' historical behavior data, and record users' interaction relationships with short videos; Establish content ownership edges between short video nodes and interest cluster nodes based on short video interest clusters; Based on the user short video interaction dataset and short video interest clusters, establish interest enhancement edges between user nodes and interest cluster nodes; Based on continuous stimulus intensity and real-time feedback data, a fatigue decay edge is established between user nodes and interest cluster nodes; Based on the exposure sequence data and the intensity of continuous stimuli, establish continuous stimulus edges between exposure sequence nodes and interest cluster nodes; Based on real-time feedback data, feedback write-back edges are established between feedback status nodes and user nodes. Based on graph nodes, behavioral feedback edges, content ownership edges, interest enhancement edges, fatigue decay edges, continuous stimulus edges, and feedback write-back edges, an interest fatigue propagation graph is generated according to node relationships.

5. The short video recommendation method based on graph neural networks according to claim 1, characterized in that, The generation of the graph propagation data specifically includes: Configure interest enhancement weights based on the user's historical behavior data corresponding to the interest enhancement edges; Configure fatigue decay weights based on the continuous stimulus intensity and real-time feedback data corresponding to the fatigue decay edge. Configure exposure interval weights based on the exposure time, exposure order, and exposure interval of similar short videos in the exposure sequence data; Configure the feedback change weight based on real-time feedback data, and configure the continuous push density weight based on continuous exposure characteristics; Interest enhancement weight and fatigue decay weight are written into the interest enhancement edge and fatigue decay edge, which are the core propagation edges, respectively. Exposure interval weight and continuous push density weight are written into the continuous stimulus edge. Feedback change weight is written into the feedback write-back edge. Graph propagation data is generated by combining behavioral feedback edge, content ownership edge, node features and relationship type.

6. The short video recommendation method based on graph neural networks according to claim 1, characterized in that, The generation of the user interest enhancement representation and user fatigue decay representation specifically includes: The graph propagation data is input into the dual-channel propagation model. The node features of each graph node in the interest fatigue propagation graph are read through the node feature input layer, and the relationship type and edge state parameters corresponding to the edge of the target graph are read. The dual-channel propagation model includes an interest enhancement propagation branch and a fatigue decay propagation branch. Both the interest enhancement propagation branch and the fatigue decay propagation branch adopt a relation graph convolutional network structure. By dividing the target graph edges in the graph propagation data into branches through a relation type partitioning layer, we obtain an interest-enhancing relation set and a fatigue-decaying relation set. By using the relation-specific weight transformation layer in the interest enhancement propagation branch and the fatigue decay propagation branch, relation-specific transformations are performed on the node features within the corresponding propagation range according to the relation type and edge state parameters to obtain interest enhancement propagation information and fatigue decay propagation information. By using a relational neighbor aggregation layer, interest enhancement propagation information and fatigue decay propagation information are aggregated by relational sub-relationship to generate positive interest aggregation features and fatigue decay aggregation features. The self-loop feature retention layer preserves the node features of the target user node, and the positive interest aggregation feature and fatigue decay aggregation feature corresponding to the target user node are fused with the node features of the target user node respectively. The node representation output layer outputs the user interest enhancement representation and the user fatigue decay representation.

7. The short video recommendation method based on graph neural networks according to claim 1, characterized in that, The determination of the recommended net income specifically includes: Obtain a set of candidate short videos and read the short video identifier and content affiliation of each candidate short video. Map each candidate short video to the corresponding short video interest cluster and generate the candidate affiliation relationship between the candidate short video and the short video interest cluster. Based on the candidate affiliation relationship, the target user’s interest enhancement component for the short video interest cluster to which each candidate short video belongs is extracted from the user interest enhancement representation, and the real interest enhancement state value is calculated based on the interest enhancement component. Based on the candidate attribution relationship, the fatigue decay component of the target user to the short video interest cluster to which each candidate short video belongs is extracted from the user fatigue decay representation, and the interest fatigue state value is obtained by combining the continuous stimulus intensity and real-time feedback data. Subtracting the interest fatigue value from the genuine interest enhancement value yields the net recommendation benefit for each candidate short video.

8. The short video recommendation method based on graph neural networks according to claim 1, characterized in that, The short video recommendation list includes the recommendation position, short video identifier, short video theme, semantic tags, author identifier, and net recommendation revenue for each candidate short video.

Images