A user preference-aware short video content recommendation and distribution system

Abstract

The application discloses a kind of user preference perceived short video content recommendation distribution systems, comprising: data preprocessing module, user short video interaction data is collected, and user behavior sequence is generated;Preference modeling module, construct time-semantic double scale map, generate user preference representation;Multi-modal encoding module, extract image, speech, caption and so on Feature, generate multi-modal content vector;Score calculation module, improve xDeepFM model, output dynamic recommendation score;Distribution strategy module, build progressive theme quota table, generate recommendation queue;Closed loop update module, distribute short video and record feedback, update behavior sequence and preference.The application realizes the accurate identification of user immediate interest migration and the adaptive optimization of recommendation distribution order in short video recommendation process by constructing dynamic recommendation model of user preference perception and multi-modal content matching.

Description

Technical Field

[0001] This invention relates to the field of intelligent recommendation technology, and in particular to a short video content recommendation and distribution system that is aware of user preferences. Background Technology

[0002] With the rapid development of short video platforms, the user base and content volume have experienced explosive growth. Achieving accurate recommendations amidst this massive amount of content has become a crucial research direction in the field of intelligent recommendation. Existing technologies typically construct user profiles based on historical click behavior, browsing history, or static interest tags, and combine these with collaborative filtering or deep learning models to rank and distribute short videos, thus improving recommendation efficiency and user experience to some extent. However, these methods largely rely on long-term accumulated user behavior data and have limited ability to perceive short-term changes in user interests, making it difficult to adapt to real-time fluctuations in user interests during continuous browsing.

[0003] In practical applications, users often exhibit significant interest migration characteristics while watching short videos, such as rapidly shifting from one topic to another within a short period or showing periodic interest in novel content. Existing technologies typically employ a uniform time window or a single interest representation during modeling, lacking differentiation and collaborative modeling of immediate, periodic, and long-term stable interests. This results in a lag in the response of recommendation results to the user's current interests. The utilization of multimodal features in short videos still relies primarily on simple concatenation or single-feature modeling, failing to fully explore the deep correlation between user preferences and multimodal content, thus impacting recommendation accuracy.

[0004] Existing recommendation systems rely heavily on model scores for content distribution, lacking effective mechanisms to constrain the diversity and novelty of recommendation results. This can easily lead to problems such as severe content homogenization, concentrated exposure of creators, and a monotonous user experience. Furthermore, the closed-loop update mechanism between recommendation results and user feedback is imperfect, and insufficient utilization of viewing completion rates, skipping behavior, and interactive feedback makes it difficult to adjust recommendation strategies in a timely manner.

[0005] Therefore, how to provide a short video content recommendation and distribution system that is aware of user preferences is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a user preference-aware short video content recommendation and distribution system. This invention models user short video interaction behavior by constructing a time-semantic dual-scale variable graph, and combines multimodal content feature encoding and an improved xDeepFM model to achieve dynamic matching and scoring between user preference representations and short video content. At the same time, a progressive topic quota table is introduced to control the distribution of recommendation results, and a closed-loop update mechanism is formed based on user viewing completion rate, skipped status, and interactive feedback. Thus, it has the advantages of high accuracy in user interest recognition, fast preference response speed, and strong diversity of recommended content.

[0007] A user preference-aware short video content recommendation and distribution system according to an embodiment of the present invention includes:

[0008] The data preprocessing module is used to collect user short video interaction data on short video platforms, perform preprocessing on the user short video interaction data, and generate user behavior sequences.

[0009] The preference modeling module is used to construct a time-semantic dual-scale variable graph based on user behavior sequences. It establishes temporal and semantic adjacencies within a sliding window, performs temporal and semantic fusion updates on behavior nodes, aggregates immediate interests, stage interests, and long-term stable interests, and splices them together to form a user preference representation.

[0010] The multimodal coding module is used to perform unified vectorized encoding on the image, audio, subtitle, title and tag features of candidate short videos to generate multimodal content vectors;

[0011] The rating calculation module is used to concatenate user preference representation and multimodal content vector into the improved xDeepFM model. The improved xDeepFM model connects an interest gating layer and a residual feature enhancement layer in parallel after the linear part, the factorization machine part and the compressed cross network. A time-level attention mechanism is introduced in the compressed cross network to output dynamic recommendation ratings for candidate short videos.

[0012] The distribution strategy module is used to label candidate short videos with themes, creators, styles and release times based on dynamic recommendation scores, build a progressive theme quota table, first fill the minimum quota of the main preferred theme, then insert and dynamically adjust the quota of non-mainstream themes according to the novelty threshold, and generate a recommendation queue.

[0013] The closed-loop update module is used to distribute short videos according to the recommendation queue, record the viewing completion rate, skip status and interaction feedback, update the user behavior sequence and user preference representation, and re-input into the improved xDeepFM model for score update, forming a closed-loop iterative recommendation distribution result.

[0014] Optionally, the user short video interaction data includes playback behavior data and interaction behavior data. The playback behavior data includes playback duration, playback completion rate, skipped position, and number of repeated views. The interaction behavior data includes likes, comments, favorites, forwards, and search behavior data.

[0015] Optionally, the preprocessing of user short video interaction data to generate user behavior sequences includes:

[0016] Acquire user short video interaction data, perform correlation and matching based on user identifier, short video identifier and behavior occurrence time, and generate a set of original behavior records;

[0017] Perform missing data detection, duplicate behavior detection, and abnormal behavior filtering on the original set of behavior records to generate a cleaned set of behavior records.

[0018] The cleaned set of behavior records is sorted in ascending order of execution time based on the time of occurrence of the behavior, and adjacent behavior records are sequentially associated to generate a time-series behavior sequence;

[0019] The playback behavior data and interaction behavior data in the time-series behavior sequence are uniformly encoded, and each behavior record is assigned a corresponding behavior type identifier, time index identifier and short video identifier to generate a standardized behavior sequence.

[0020] The standardized behavior sequence is continuously divided according to a preset sliding time window, and the user behavior sequence is output.

[0021] Optionally, the step of performing temporal and semantic fusion updates on behavior nodes, aggregating immediate interests, periodic interests, and long-term stable interests, and concatenating them to form a user preference representation includes:

[0022] The sliding window length is adaptively determined based on the density of user behavior sequences. Behavior records within the same sliding window are used as behavior nodes. Temporal relationships are established based on the order of behavior, and semantic relationships are established based on the interaction object, behavior type, and content theme, generating an initial time-semantic dual-scale variable graph.

[0023] Calculate the activity weight of newly added behavior nodes within each sliding window, perform marginalization processing on nodes with activity below a set threshold, perform elimination processing on nodes that exceed the sliding window, and recalibrate the temporal adjacency relationship between the remaining nodes according to the decay coefficient of the most recent interaction time to generate a sparse temporal adjacency graph.

[0024] Based on the topic similarity, style distance and creator difference of candidate short videos, positive weights or suppression weights are assigned to the semantic association between behavioral nodes. The semantic adjacency with the highest weight is retained and negative suppression edges are added to form a reinforced semantic adjacency graph.

[0025] Dual-path attention propagation is performed on the sparse temporal adjacency graph and the enhanced semantic adjacency graph respectively. The features of near-end nodes are aggregated using a decreasing coefficient on the temporal adjacency side, and the features of heterogeneous nodes are fused using a complementary coefficient on the semantic adjacency side. The two aggregation results are fused and updated through a residual gating unit to obtain the updated behavior node features.

[0026] Based on the updated behavioral node features, the node features in the most recent window are aggregated to generate immediate interests, the node features in the intermediate window are aggregated to generate stage interests, and the node features in the far-term window are aggregated to generate long-term stable interests. The immediate interests, stage interests, and long-term stable interests are then concatenated to generate a user preference representation.

[0027] Optionally, generating the multimodal content vector includes:

[0028] Extract continuous image frames according to the playback timeline of candidate short videos, perform size unification, keyframe filtering and visual region segmentation on the continuous image frames, extract the features of objects, scene types, human actions and color changes, and generate image feature sequences.

[0029] The speech signal is separated from the candidate short video, and noise reduction, frame segmentation and acoustic feature extraction are performed on the speech signal. The semantic content, speech rate changes, emotional state and background sound type of the speech signal are identified, and a speech feature sequence is generated.

[0030] Obtain the subtitles and titles corresponding to the candidate short videos, perform text cleaning, word segmentation and semantic encoding on the subtitles and titles, extract topic words, entity words, sentiment words and content category features, and generate text feature sequences;

[0031] Obtain the tags corresponding to the candidate short videos, perform hierarchical mapping according to the theme, content category, style and target audience of the tags, and perform merge encoding on duplicate tags and semantically similar tags to generate tag feature sequences;

[0032] Based on the candidate short video identifiers, the image feature sequence, speech feature sequence, text feature sequence, and tag feature sequence are associated and aligned, and dimension unification, feature splicing, and normalization are performed to generate a multimodal content vector of the candidate short video.

[0033] Optionally, the dynamic recommendation score for the output candidate short videos includes:

[0034] The linear part includes an embedding mapping layer and a first-order weight layer. The embedding mapping layer maps the high-dimensional feature vector obtained by concatenating the user preference representation vector and the multimodal content vector to a unified low-dimensional embedding space to generate a low-dimensional embedding feature representation. The first-order weight layer performs a weighted summation of the low-dimensional embedding feature representation according to the dimensional weight to generate a first-order recommendation feature.

[0035] The factorization machine part includes a latent vector decomposition layer and a second-order interaction layer. The latent vector decomposition layer decomposes the embedded feature vector into several latent vector groups. The second-order interaction layer performs an inner product operation on any two groups of latent vectors and accumulates them to output a second-order preference matching feature.

[0036] The compressed crossover network consists of multi-level explicit crossover layers and compression mapping layers. The multi-level explicit crossover layers perform explicit crossover on the embedded feature vectors level by level. The compression mapping layer compresses the output according to a set dimension after each level of crossover. A time-level attention mechanism is introduced inside the explicit crossover layer to assign weights to crossover features at different time levels and output time-level high-order crossover features.

[0037] The interest gating layer includes an interest matching calculation unit and a gating weight generation unit. The interest matching calculation unit generates matching vectors based on the matching degree between real-time interest vectors, phased interest vectors, and long-term stable interest vectors and multimodal content vectors. The gating weight generation unit outputs gating weights based on the matching vectors and performs weighted filtering and fusion on first-order recommendation features, second-order preference matching features, and time-level high-order cross features to generate interest gating features.

[0038] The residual feature enhancement layer includes a feature mapping unit and a residual fusion unit. The feature mapping unit performs linear mapping on the concatenated feature vectors. The residual fusion unit performs residual superposition and normalization on the linear mapping result and interest-gated features to output residual-enhanced recommendation features. The first-order recommendation features, second-order preference matching features, time-level high-order cross features and residual-enhanced recommendation features are concatenated and processed by fully connected mapping and normalization to generate dynamic recommendation scores for candidate short videos.

[0039] Optionally, the generation of the recommendation queue includes:

[0040] Candidate short videos are sorted from highest to lowest according to the dynamic recommendation score. The theme, creator, style and release time of each candidate short video are read. The theme, creator, style and release time are associated with the candidate short video identifier and dynamic recommendation score to generate a sorted set of candidate short videos.

[0041] Based on the user's immediate interests, periodic interests, and long-term stable interests, the aggregation strength of different topics in the three types of interests is calculated. The main preferred topics and non-mainstream topics are determined according to the aggregation strength from large to small, and the number of candidate short videos corresponding to each topic is recorded.

[0042] Based on the aggregation strength of the main preference theme and the number of candidate short videos, a minimum filling quantity is set for each main preference theme. The corresponding short videos are selected sequentially from the candidate short video sorting set according to the dynamic recommendation score to fill the minimum quota of the main preference theme and generate the main preference content sequence.

[0043] For candidate short videos that are not included in the main preferred content sequence, the differences between the candidate short video's theme, creator, style, and release time and the already filled short videos are compared. The novelty of the candidate short videos is then summarized, and candidate short videos with a novelty of a preset threshold are identified as non-mainstream theme insertion objects.

[0044] The insertion position of non-mainstream topics is determined by the number of consecutive identical topics in the main preferred content sequence. Non-mainstream topic insertion objects are inserted into the corresponding positions in sequence according to the dynamic recommendation score. The available quota of each non-mainstream topic is adjusted according to the number of repeated topics, repeated creators, and repeated styles of the inserted short videos, and a progressive topic quota table is generated.

[0045] Based on the progressive theme quota table, short videos on the main preferred theme and non-mainstream theme are extracted one by one. The order of short videos with continuous repetition by creators, continuous repetition of style, and excessive concentration of release time is adjusted to generate a recommendation queue of candidate short videos.

[0046] Optionally, the recommendation distribution results forming a closed-loop iteration include:

[0047] Candidate short videos are distributed sequentially according to the recommendation queue. The display time, display position and distribution order of each candidate short video are recorded. The viewing completion rate, skip status and interaction feedback of users during the viewing process are collected to generate new user short video interaction data.

[0048] New user short video interaction data is written into the user behavior sequence according to the occurrence time. Based on the updated user behavior sequence, a time-semantic dual-scale variable graph is reconstructed. The behavior nodes are updated by temporal and semantic fusion. Real-time interests, stage interests and long-term stable interests are re-aggregated to generate an updated user preference representation.

[0049] The updated user preference representation is concatenated with the multimodal content vector corresponding to the undistributed candidate short videos, and then input into the improved xDeepFM model. The model is processed sequentially through the linear part, factorization machine part, compressed cross network, interest gating layer and residual feature enhancement layer, and outputs the updated dynamic recommendation score corresponding to each undistributed candidate short video.

[0050] Based on the updated dynamic recommendation scores, the ranking of each candidate short video is recalculated, the progressive topic quota table is updated, the minimum quota for the main preferred topic and the quota for non-mainstream topics are reallocated, and a new recommendation queue is generated.

[0051] The short video distribution continues according to the new recommendation queue, and the interaction data of newly added users' short videos is continuously recorded. The user behavior sequence update, user preference representation update, dynamic recommendation score update and recommendation queue update are repeatedly executed to form a closed-loop iterative recommendation distribution result.

[0052] The beneficial effects of this invention are:

[0053] This invention constructs a time-semantic dual-scale variable graph to perform hierarchical modeling of user behavior sequences, enabling the differentiation and collaborative characterization of users' immediate, stage-based, and long-term stable interests. Compared to existing recommendation methods that rely solely on a single interest representation, this invention can more accurately reflect changes in users' interests during continuous short video browsing, thereby significantly improving the accuracy of user preference recognition and making recommendation results more aligned with users' current real needs.

[0054] This invention achieves deep matching and dynamic fusion between user preference representation and multimodal content features by uniformly vectorizing the features of short video images, audio, subtitles, titles, and tags, and combining the interest gating layer, residual feature enhancement layer, and time-level attention mechanism introduced in the improved xDeepFM model. This effectively mines the correlation between interests and content at different time scales, improves the accuracy and stability of recommendation scoring, and enhances the system's ability to identify potential interest content.

[0055] This invention constrains and regulates the recommendation queue by constructing a progressive topic quota table, and forms a closed-loop update mechanism by combining user viewing completion rate, skipped status, and interactive feedback. This ensures that the recommendation results, while guaranteeing coverage of the main preferred content, introduce an appropriate amount of non-mainstream topic content and dynamically adjust its exposure ratio. This avoids the problem of content homogenization, improves the diversity and novelty of the recommendation results, and can optimize the recommendation strategy in real time based on user feedback, thereby improving the response speed of the recommendation system and the overall user experience. Attached Figure Description

[0056] 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:

[0057] Figure 1 This is a schematic diagram of the structure of a user preference-aware short video content recommendation and distribution system proposed in this invention;

[0058] Figure 2 This is a diagram showing the adaptive sliding window partitioning and behavior density variation of a user preference-aware short video content recommendation and distribution system proposed in this invention.

[0059] Figure 3 This is a weighted distribution diagram of the activity of behavioral nodes in a user preference-aware short video content recommendation and distribution system proposed in this invention. Detailed Implementation

[0060] 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.

[0061] refer to Figures 1-3 A user preference-aware short video content recommendation and distribution system, comprising:

[0062] The data preprocessing module is used to collect user short video interaction data on short video platforms, perform preprocessing on the user short video interaction data, and generate user behavior sequences.

[0063] The preference modeling module is used to construct a time-semantic dual-scale variable graph based on user behavior sequences. It establishes temporal and semantic adjacencies within a sliding window, performs temporal and semantic fusion updates on behavior nodes, aggregates immediate interests, stage interests, and long-term stable interests, and splices them together to form a user preference representation.

[0064] The multimodal coding module is used to perform unified vectorized encoding on the image, audio, subtitle, title and tag features of candidate short videos to generate multimodal content vectors;

[0065] The rating calculation module is used to concatenate user preference representation and multimodal content vector into the improved xDeepFM model. The improved xDeepFM model connects an interest gating layer and a residual feature enhancement layer in parallel after the linear part, the factorization machine part and the compressed cross network. A time-level attention mechanism is introduced in the compressed cross network to output dynamic recommendation ratings for candidate short videos.

[0066] The distribution strategy module is used to label candidate short videos with themes, creators, styles and release times based on dynamic recommendation scores, build a progressive theme quota table, first fill the minimum quota of the main preferred theme, then insert and dynamically adjust the quota of non-mainstream themes according to the novelty threshold, and generate a recommendation queue.

[0067] The closed-loop update module is used to distribute short videos according to the recommendation queue, record the viewing completion rate, skip status and interaction feedback, update the user behavior sequence and user preference representation, and re-input into the improved xDeepFM model for score update, forming a closed-loop iterative recommendation distribution result.

[0068] In this embodiment, the user short video interaction data includes playback behavior data and interaction behavior data. The playback behavior data includes playback duration, playback completion rate, skipped position, and number of repeated views. The interaction behavior data includes likes, comments, favorites, forwards, and search behavior data.

[0069] In this embodiment, the step of preprocessing user short video interaction data to generate user behavior sequences includes:

[0070] Acquire user short video interaction data, perform correlation and matching based on user identifier, short video identifier and behavior occurrence time, and generate a set of original behavior records;

[0071] Perform missing data detection, duplicate behavior detection, and abnormal behavior filtering on the original set of behavior records to generate a cleaned set of behavior records.

[0072] The cleaned set of behavior records is sorted in ascending order of execution time based on the time of occurrence of the behavior, and adjacent behavior records are sequentially associated to generate a time-series behavior sequence;

[0073] The playback behavior data and interaction behavior data in the time-series behavior sequence are uniformly encoded, and each behavior record is assigned a corresponding behavior type identifier, time index identifier and short video identifier to generate a standardized behavior sequence.

[0074] The standardized behavior sequence is continuously divided according to a preset sliding time window, and the user behavior sequence is output. The preset sliding time window is 300 seconds.

[0075] In this embodiment, the step of performing temporal and semantic fusion updates on behavior nodes, aggregating immediate interests, periodic interests, and long-term stable interests, and concatenating them to form a user preference representation includes:

[0076] The sliding window length is adaptively determined based on the density of user behavior sequences. Behavior records within the same sliding window are treated as behavior nodes. Temporal relationships are established based on the order of behavior, and semantic relationships are established based on the interaction object, behavior type, and content theme, generating an initial time-semantic dual-scale variable graph. Specifically, the sliding window length is adaptively determined based on the density of user behavior sequences as follows:

[0077] First, set the shortest window length to 30 seconds and the longest window length to 5 minutes, and then sort all behavior records by the time they occurred from earliest to latest.

[0078] Starting from the current time point, accumulate and count behavior records: if no less than 20 behaviors are accumulated within 30 consecutive seconds, then 30 seconds will be used as the length of this window;

[0079] If less than 20 actions are accumulated within 30 consecutive seconds, the time period is extended until at least 5 actions are accumulated or the 5-minute limit is reached. The extended time period is then used as the window length.

[0080] Using the obtained window end point as the new starting time, repeat the statistical and expansion process until the entire behavior sequence has been traversed;

[0081] If the number of behaviors in two adjacent windows differs by more than 50%, the windows are merged or subdivided so that the window length is dynamically adjusted according to the density of behaviors, ultimately resulting in an adaptive sliding window set.

[0082] The generation of the initial time-semantic dual-scale variable graph is specifically as follows:

[0083] Sort the behavior records within the same sliding window by their occurrence time from earliest to latest, and number them sequentially as behavior nodes;

[0084] For any two adjacent action nodes, if the difference between the occurrence time of the latter node and the occurrence time of the former node does not exceed 30 seconds, then add a time edge between the two nodes; otherwise, do not connect the two nodes.

[0085] For any two behavior nodes within the same window, compare their interaction object, behavior type, and content theme. If the similarity of all three items is higher than 70%, add a semantic edge between the two nodes; otherwise, do not connect them.

[0086] The established temporal edges are assigned temporal weights that decrease with time intervals, while the semantic edges are assigned semantic weights that increase with similarity.

[0087] If the number of behavior nodes in two adjacent sliding windows differs by more than 50%, the windows are merged or subdivided, and the temporal and semantic edges are updated. Finally, all behavior nodes, temporal edges, and semantic edges are summarized to obtain a temporal-semantic dual-scale variable graph, which is written to the graph update cache as the initial structure for fusion update.

[0088] For each newly added behavior node within a sliding window, a behavior activity weight is calculated. Nodes with activity below a set threshold are marginalized, and nodes exceeding the sliding window are eliminated. The temporal adjacency relationships between the remaining nodes are recalibrated using the decay coefficient of the most recent interaction time, generating a sparse temporal adjacency graph. Specifically, the behavior activity weight for each newly added behavior node within a sliding window is calculated as follows:

[0089] Record various interactive metrics associated with the nodes: playback completion rate, skip rate, number of likes, number of comments, number of favorites, number of shares, and number of search triggers;

[0090] A fixed contribution percentage is allocated to different metrics: completion rate accounts for 35%, effective viewership rate accounts for 15%, likes account for 15%, comments account for 10%, favorites account for 10%, shares account for 10%, and searches account for 5%.

[0091] Logarithmic normalization was applied to the number of likes, comments, favorites, shares, and searches to eliminate absolute numerical differences; the conversion of views into effective viewership was skipped.

[0092] The activity weight of a node is obtained by multiplying each indicator by its corresponding contribution ratio and summing the results.

[0093] Compared to a threshold of 0.25: nodes with a weight below 0.25 are marked as inactive and marginalized; nodes with a weight greater than or equal to 0.25 are retained.

[0094] Generate a sparsed temporal adjacency graph, specifically as follows:

[0095] Within the current sliding window, retain behavior nodes with an activity weight of at least 0.25, and remove nodes that have been marked as low-activity and marginalized.

[0096] Based on the occurrence time of the remaining nodes, calculate the time interval between two adjacent nodes; if the time interval does not exceed 30 seconds, retain the original time edge between the two nodes; otherwise, delete the time edge.

[0097] The retention time edges are assigned attenuation coefficients according to the time interval size. Time edges with time intervals in the range of 0–10 seconds are assigned a weight of 1, time edges with time intervals in the range of 10–20 seconds are assigned a weight of 0.6, and time edges with time intervals in the range of 20–30 seconds are assigned a weight of 0.3.

[0098] For two nodes in the same window but not directly adjacent, if the time interval exceeds 30 seconds, no time edge is created; if the time interval is less than 30 seconds but all intermediate edges between the two nodes have been deleted, a weak time edge with a weight of 0.1 is added.

[0099] Summarize the remaining nodes and the recalibrated time edges, and output a sparse time adjacency graph with fewer nodes, sparser edges, and weight information.

[0100] Based on the topic similarity, style distance, and creator differences of candidate short videos, positive or suppressive weights are assigned to the semantic relationships between behavioral nodes. The top-ranked semantic adjacencies are retained and negative suppressive edges are added to form a reinforced semantic adjacency graph. Specifically, the assignment of positive or suppressive weights based on the topic similarity, style distance, and creator differences of candidate short videos to the semantic relationships between behavioral nodes is as follows:

[0101] For any two behavior nodes within the same sliding window, read the topic tags, style tags, and creator identifiers of the candidate short videos respectively, and calculate the topic similarity, style distance, and creator difference of the short videos corresponding to the two nodes.

[0102] If the topic similarity between two nodes is ≥80%, the style distance is ≤20%, and the creator difference is ≥50%, then a positive weight of 1 is assigned to the semantic association.

[0103] If only two of the three indicators are met, a positive weight of 0.6 is assigned; if only one indicator is met, a positive weight of 0.3 is assigned.

[0104] If none of the three indicators are met, a suppression weight of -0.3 is assigned to the semantic association, and the association is marked as a semantic edge to be suppressed.

[0105] All semantic associations are sorted from high to low according to positive weights. The top 25% of positive semantic edges are retained, and negative suppression edges are added to the bottom 25% of suppression semantic edges to form a reinforced semantic adjacency graph.

[0106] Dual-path attention propagation is performed on both the sparse temporal adjacency graph and the enhanced semantic adjacency graph. For the temporal adjacency side, a decreasing coefficient is used to aggregate near-end node features, while for the semantic adjacency side, complementary coefficients are used to fuse heterogeneous node features. The two aggregation results are then fused and updated using a residual gating unit to obtain the updated behavior node features, where:

[0107] For temporal adjacency, a decreasing coefficient is used to aggregate near-end node features, specifically:

[0108] For each target node, neighboring nodes are selected in descending order of time edge weight;

[0109] The features of the target node and its time neighbors within one hop are summed by a weighted sum of 1.

[0110] The features of adjacent nodes with remaining time intervals of 10–20 seconds are weighted and summed with a coefficient of 0.6.

[0111] The features of adjacent nodes with time intervals of 20–30 seconds are weighted and summed with a coefficient of 0.3.

[0112] The weighted sums of the above three layers are added together to form the time-side aggregated feature;

[0113] For the semantic adjacency side, complementary coefficients are used to fuse heterogeneous node features, specifically:

[0114] For each target node, the features of semantically adjacent nodes with the same theme but different creators are summed with a weighted average of 0.7.

[0115] Features that are complementary in style and have a similarity of 50%–80% to the theme are summed by a weighting factor of 0.5.

[0116] Node features that are the same as the creator but have a theme difference greater than 50% are summed using a weighted average of 0.2.

[0117] The three weighted sums are added together to form a semantic aggregate feature;

[0118] Based on the updated behavioral node features, the node features within the most recent window are aggregated to generate immediate interest, the node features within the intermediate window are aggregated to generate stage-specific interest, and the node features within the far-term window are aggregated to generate long-term stable interest. The immediate interest, stage-specific interest, and long-term stable interest are then concatenated to generate a user preference representation, where:

[0119] The nearest window is the collection of behavioral nodes accumulated from the current moment back 10 minutes. If there are fewer than 20 nodes in 10 minutes, the window continues to expand forward until there are 20 nodes or the 30-minute limit is reached. All nodes within this time period are then used as the nearest window.

[0120] Mid-term window: Continue accumulating from the most recent window forward until it covers behavioral nodes from 30 minutes to 3 hours; if the total number of nodes is less than 100 within 3 hours, all nodes are merged into a mid-term window;

[0121] Forward window: refers to all historical behavior nodes that are more than 3 hours ago and no more than 7 days ago; if there are fewer than 200 nodes more than 3 hours ago, they are merged with the intermediate window; if there are still fewer than 200 data nodes more than 7 days ago, all available nodes are retained as the forward window.

[0122] Generate user preference representations, specifically as follows:

[0123] The weights of all behavioral node features within the most recent window are assigned in decreasing order of occurrence time. Nodes 0-2 minutes from the current time are assigned a weight of 1, nodes 2-5 minutes from the current time are assigned a weight of 0.7, and nodes 5-10 minutes from the current time are assigned a weight of 0.4. The weighted node features are averaged to obtain the instantaneous interest vector.

[0124] The weights of all behavioral node features within the intermediate window are assigned in decreasing order of occurrence time. The weight of nodes 10-30 minutes from the current time is set to 0.6, the weight of nodes 30 minutes-1 hour is set to 0.4, and the weight of nodes 1-3 hours is set to 0.2. The average of the weighted node features is calculated to obtain the stage-specific interest vector.

[0125] The weights of all behavioral node features within the long-term window are assigned in decreasing order of occurrence time. The weight of nodes 3-12 hours from the current time is set to 0.3, the weight of nodes 12 hours-1 day is set to 0.2, and the weight of nodes 1-7 days is set to 0.1. The weighted node features are averaged to obtain the long-term stable interest vector.

[0126] L2 normalization is performed sequentially on the immediate interest vector, the periodic interest vector, and the long-term stable interest vector so that the magnitude of the three types of interest vectors is 1.

[0127] The normalized real-time interest vector, the phased interest vector, and the long-term stable interest vector are concatenated in sequence. The concatenation result is then compressed to a fixed dimension using a linear mapping, and the user preference representation vector is output.

[0128] In this embodiment, generating the multimodal content vector includes:

[0129] Extract continuous image frames according to the playback timeline of candidate short videos, perform size unification, keyframe filtering and visual region segmentation on the continuous image frames, extract the features of objects, scene types, human actions and color changes, and generate image feature sequences.

[0130] The speech signal is separated from the candidate short video, and noise reduction, frame segmentation and acoustic feature extraction are performed on the speech signal. The semantic content, speech rate changes, emotional state and background sound type of the speech signal are identified, and a speech feature sequence is generated.

[0131] Obtain the subtitles and titles corresponding to the candidate short videos, perform text cleaning, word segmentation and semantic encoding on the subtitles and titles, extract topic words, entity words, sentiment words and content category features, and generate text feature sequences;

[0132] Obtain the tags corresponding to the candidate short videos, perform hierarchical mapping according to the theme, content category, style and target audience of the tags, and perform merge encoding on duplicate tags and semantically similar tags to generate tag feature sequences;

[0133] Based on the candidate short video identifiers, the image feature sequence, speech feature sequence, text feature sequence, and tag feature sequence are associated and aligned, and dimension unification, feature splicing, and normalization are performed to generate a multimodal content vector of the candidate short video.

[0134] In this embodiment, the dynamic recommendation score for the output candidate short video includes:

[0135] The linear part includes an embedding mapping layer and a first-order weighting layer. The embedding mapping layer maps the high-dimensional feature vector obtained by concatenating the user preference representation vector and the multimodal content vector to a unified low-dimensional embedding space, generating a low-dimensional embedding feature representation. The first-order weighting layer performs a weighted summation of the low-dimensional embedding feature representation according to the dimensional weights, generating a first-order recommendation feature. Specifically, the generation of the first-order recommendation feature is as follows:

[0136] After concatenating the user preference representation vector with the multimodal content vector, the concatenation is input into the embedding mapping layer, which maps the high-dimensional feature vector to a unified low-dimensional embedding space of dimension 128, thus obtaining the embedded feature vector.

[0137] In the first-order weight layer, a weight of 0.8 is assigned to the 1st to 32nd dimensions of the embedded feature vector, a weight of 0.6 is assigned to the 33rd to 64th dimensions, a weight of 0.4 is assigned to the 65th to 96th dimensions, and a weight of 0.2 is assigned to the 97th to 128th dimensions.

[0138] The weighted feature vector is obtained by multiplying each dimension of the embedded feature vector by its corresponding weight dimension by dimension.

[0139] Summing each dimension of the weighted feature vector yields a first-order linear score.

[0140] Add the first-order linear score to the all-zero bias vector and output it as the first-order recommendation feature;

[0141] The factorization machine consists of a latent vector decomposition layer and a second-order interaction layer. The latent vector decomposition layer decomposes the embedded feature vectors into several latent vector groups. The second-order interaction layer performs an inner product operation on any two groups of latent vectors and sums them to output second-order preference matching features. Specifically, the output second-order preference matching features are as follows:

[0142] The embedded feature vectors are divided into 8 groups, each with 16 dimensions, resulting in 8 latent vectors;

[0143] Perform vector dot product operations on any two sets of implicit vectors in sequence, and calculate a total of 28 dot product results;

[0144] Multiply each inner product result by a coefficient of 0.5, and sum the 28 results to obtain the combined score;

[0145] The combined score is added to the constant bias of 0.1 to obtain the second-order matching score;

[0146] The second-order matching score is extended to 32 dimensions through a linear mapping, and the output is used as a second-order preference matching feature.

[0147] The compressed cross-connect network consists of multi-level explicit cross-connect layers and compression mapping layers. The multi-level explicit cross-connect layers perform explicit cross-connection on the embedded feature vectors level by level. The compression mapping layer compresses the output according to a set dimension after each cross-connection level. A time-level attention mechanism is introduced within the explicit cross-connect layers to assign weights to cross-connect features at different time levels, outputting time-level high-order cross-connect features. Specifically, the weight assignment and output of time-level high-order cross-connect features are as follows:

[0148] The higher-order features of the explicit cross-output of the previous layer are divided into three categories based on their time source: cross-features from the most recent window, cross-features from the intermediate window, and cross-features from the far-term window.

[0149] Set the initial weight of the most recent window cross feature to 0.6, the initial weight of the intermediate window cross feature to 0.3, and the initial weight of the long-term window cross feature to 0.1.

[0150] If a user exhibits consistently high activity within the last 5 minutes (average viewing completion rate greater than 80% and more than 3 interactions triggered), the weight of the recent window cross feature will be increased to 0.8, while the weights of the mid-term and long-term window cross features will be decreased to 0.15 and 0.05, respectively.

[0151] When user activity levels stabilize within the last 30 minutes (average viewing completion rate between 40% and 80% and no more than 3 interactions), maintain the initial weighting.

[0152] If a user’s activity level drops significantly in the last 30 minutes (average completion rate below 40% and skip rate above 60%), the weight of the cross-feature of the long-term window will be increased to 0.3, the weight of the cross-feature of the medium-term window will be increased to 0.5, and the weight of the cross-feature of the recent window will be decreased to 0.2.

[0153] The three types of cross features are multiplied by the adjusted weight coefficients and then weighted and summed to output the fused time-level high-order cross features.

[0154] The interest gating layer includes an interest matching calculation unit and a gating weight generation unit. The interest matching calculation unit generates matching vectors based on the matching degree between real-time interest vectors, periodic interest vectors, and long-term stable interest vectors and multimodal content vectors. The gating weight generation unit outputs gating weights based on the matching vectors and performs weighted filtering and fusion on first-order recommendation features, second-order preference matching features, and time-level high-order cross features to generate interest-gated features, where:

[0155] Generate the matching vector, specifically:

[0156] Calculate the cosine similarity between the user's real-time interest vector and the multimodal content vector to obtain the real-time matching score;

[0157] Calculate the cosine similarity between the user's staged interest vector and the multimodal content vector to obtain the stage matching score;

[0158] Calculate the cosine similarity between the user's long-term stable interest vector and the multimodal content vector to obtain the long-term matching score;

[0159] The immediate matching score, the phase matching score, and the long-term matching score are weighted and summed with weights of 0.6, 0.3, and 0.1, respectively, to obtain the comprehensive matching score.

[0160] The overall matching score is concatenated with the three individual matching scores in sequence to form a matching vector of length 4;

[0161] Generate interest-gated features, specifically:

[0162] The matching vector is input into the gate weight generation unit, and after passing through a fully connected layer and sigmoid activation, the four-dimensional gate weight is output.

[0163] The gating weights are normalized so that the sum of the weights is 1;

[0164] The weighted feature vectors are multiplied sequentially by the first-order recommendation feature, the second-order preference matching feature, and the time-level high-order cross feature, and then output as a weighted feature vector set.

[0165] The weighted feature vectors are aligned by dimension and then summed element by element to obtain the fused feature vector.

[0166] Perform layer normalization and ReLU activation on the fused feature vector to output interest-gated features;

[0167] The residual feature enhancement layer includes a feature mapping unit and a residual fusion unit. The feature mapping unit performs a linear mapping on the concatenated feature vector. The residual fusion unit performs residual superposition and normalization on the linear mapping result and interest-gated features, outputting residual-enhanced recommendation features. The first-order recommendation features, second-order preference matching features, time-level high-order cross features, and residual-enhanced recommendation features are concatenated, and after fully connected mapping and normalization, dynamic recommendation scores for candidate short videos are generated. Specifically, the generation of dynamic recommendation scores for candidate short videos is as follows:

[0168] The first-order recommendation features, second-order preference matching features, time-level high-order cross features, and residual enhancement recommendation features are aligned by dimension and then concatenated sequentially to obtain the concatenated feature vector.

[0169] The concatenated feature vector is input into the first fully connected mapping layer and mapped to a 128-dimensional hidden vector.

[0170] The hidden vectors are normalized and ReLU activation is applied to obtain normalized hidden vectors;

[0171] The normalized hidden vector is input into the second fully connected mapping layer and mapped to a 1-dimensional score vector;

[0172] Perform Sigmoid normalization on the score vector and output dynamic recommendation scores as candidate short videos.

[0173] This invention structurally optimizes the traditional xDeepFM model, proposing a dynamic scoring method for short video recommendation scenarios. Regarding feature input, the user preference representation vector is concatenated with a multimodal content vector, enabling the model to simultaneously receive user immediate interests, staged interests, long-term stable interests, and image, audio, subtitle, title, and tag information from candidate short videos. First-order recommendation features and second-order preference matching features are extracted in the linear and factorization parts, respectively, allowing the model to retain basic preference matching relationships while capturing the shallow interaction between user interests and short video content, thus improving the accuracy of the initial scoring of candidate short videos.

[0174] To address the issues of rapidly changing user interests and varying degrees of influence from different time-level preferences in short video platforms, this invention introduces a time-level attention mechanism within the compressed cross-network. This mechanism assigns different weights to cross-features from the most recent, intermediate, and long-term windows. This allows the immediate interests corresponding to the most recent window to gain a higher impact when users are actively browsing, the phased interests corresponding to the intermediate window to maintain recommendation continuity when interests are stable, and the long-term stable interests corresponding to the long-term window to provide stable support when user feedback is insufficient or interests fluctuate significantly. The model can dynamically adjust the contribution of features from different time levels based on user behavior, improving the responsiveness of dynamic recommendation scoring to changes in the user's current interests.

[0175] To enhance the deep matching capability between user preferences and short video content, this invention sets up an interest gating layer and a residual feature enhancement layer in parallel in the improved xDeepFM model. The interest gating layer performs weighted filtering and fusion of first-order recommendation features, second-order preference matching features, and time-level high-order cross features based on the matching degree between immediate interest vectors, stage-based interest vectors, long-term stable interest vectors, and multimodal content vectors. The residual feature enhancement layer performs residual superposition and normalization processing on the linear mapping results and interest gating features to reduce the decay of effective preference features in the multi-layer mapping process. The model can simultaneously enhance the ability to capture short-term interests, the ability to identify multimodal content, and the stability of scoring, thereby improving the accuracy, real-time performance, and personalization of short video recommendations.

[0176] In this embodiment, the generation of the recommendation queue includes:

[0177] Candidate short videos are sorted from highest to lowest according to the dynamic recommendation score. The theme, creator, style and release time of each candidate short video are read. The theme, creator, style and release time are associated with the candidate short video identifier and dynamic recommendation score to generate a sorted set of candidate short videos.

[0178] Based on the user's immediate interests, periodic interests, and long-term stable interests, the aggregation strength of different topics in the three types of interests is calculated. The main preferred topics and non-mainstream topics are determined according to the aggregation strength from large to small, and the number of candidate short videos corresponding to each topic is recorded.

[0179] Based on the aggregation strength of the main preference topic and the number of candidate short videos, a minimum filling quantity is set for each main preference topic. Short videos are then selected sequentially from the candidate short video ranking set according to the dynamic recommendation score to fill the minimum quota for each main preference topic, generating a main preference content sequence. Specifically, the generation of the main preference content sequence is as follows:

[0180] The number of behavioral nodes for each main preference topic within the recent window, medium-term window, and long-term window is counted. The number of nodes in the three types of windows is multiplied by the weights of 0.5, 0.3, and 0.2, respectively, and then summed to obtain the aggregation intensity score of the topic.

[0181] All preferred topics are sorted from high to low based on aggregation intensity. The minimum number of candidate short videos for each topic is calculated based on the number of available short videos for each topic, multiplied by 20% of the aggregation intensity percentage. The minimum number of short videos to fill is no less than 3 and no more than 40% of the total number of candidate short videos for each topic.

[0182] Videos belonging to the first priority main preference topic from the candidate short video sorting set are selected in descending order of dynamic recommendation score until the minimum number of topics to be filled is met.

[0183] Repeat the selection process, filling the minimum quantity for the second priority and subsequent main preference topics in turn, until all main preference topics have reached the set quota;

[0184] The selected videos are merged according to the filling order to form a main preference content sequence that meets the minimum quota requirements of the main preference theme;

[0185] For candidate short videos not included in the main preferred content sequence, the differences between the candidate short video's theme, creator, style, and publication time and the already filled short videos are compared. The novelty of the candidate short videos is then calculated. Candidate short videos with a novelty reaching a preset threshold are identified as non-mainstream theme insertion objects. The specific calculation of the novelty of the candidate short videos is as follows:

[0186] Calculate the similarity between the candidate short video topic and each topic in the main preferred content sequence. If the similarity is less than 40%, the topic difference score is 0.4; otherwise, it is 0.

[0187] Calculate whether the candidate short video creator is appearing in the current sliding window for the first time: if it is, score the creator difference score as 0.3; otherwise, score it as 0.

[0188] Calculate the overlap rate between the style tags of candidate short videos and the style distribution of the main preferred content sequence. If the overlap rate is less than 30%, score 0.2 for style difference; otherwise, score 0.

[0189] Calculate the time difference between the release time of the candidate short video and the current time. If the time difference is less than 2 hours, score 0.1 for freshness of release time; otherwise, score 0.

[0190] The novelty score of candidate short videos is obtained by adding the theme difference score, creator difference score, style difference score and release time freshness score, and then filtering is performed with a preset threshold of 0.7.

[0191] The insertion position of non-mainstream topics is determined based on the number of consecutive identical topics in the main preferred content sequence. Non-mainstream topics are then inserted sequentially according to dynamic recommendation scores. The available quota for each non-mainstream topic is adjusted based on the number of repeated topics, creators, and styles in the already inserted short videos, generating a progressive topic quota table. Specifically, the generation of the progressive topic quota table is as follows:

[0192] Scan the main preference content sequence sequentially and count the number n of consecutive identical topics. If n≥3, mark a non-mainstream topic insertion slot at the end of the consecutive segment; if n=2, mark an insertion slot between the consecutive segment and the next topic; if n=1, skip and do not mark an insertion slot.

[0193] All non-mainstream topic insertion objects are sorted from high to low according to the dynamic recommendation rating. Starting from the first insertion slot, the process is traversed sequentially, and the top-ranked non-mainstream topic short videos are inserted into adjacent insertion slots. If there are not enough insertion slots, insertions are added at the end of the main preferred content sequence.

[0194] After insertion, statistics are collected on the theme, creator, and style of the inserted short videos, and the theme duplication rate, creator duplication rate, and style duplication rate are calculated. When the theme duplication rate exceeds 30%, the creator duplication rate exceeds 25%, or the style duplication rate exceeds 40%, the available quota for the corresponding non-mainstream theme is reduced by 1. When all three duplication rates are below their respective thresholds, the available quota for the corresponding non-mainstream theme is increased by 1.

[0195] Iterate through the list of non-mainstream theme insertion objects again, and try to fill the remaining insertion slots in turn according to the updated available quota. If there are still empty slots, keep the empty slots and wait for the next round of recommendation to fill them; if the available quota is 0, skip the corresponding theme.

[0196] The main preferred content sequence, non-mainstream topic insertion objects and their dynamic quota information are summarized to generate a progressive topic quota table;

[0197] Based on the progressive theme quota table, short videos on the main preferred theme and non-mainstream theme are extracted one by one. The order of short videos with continuous repetition by creators, continuous repetition of style, and excessive concentration of release time is adjusted to generate a recommendation queue of candidate short videos.

[0198] In this embodiment, the recommendation distribution result that forms a closed-loop iteration includes:

[0199] Candidate short videos are distributed sequentially according to the recommendation queue. The display time, display position and distribution order of each candidate short video are recorded. The viewing completion rate, skip status and interaction feedback of users during the viewing process are collected to generate new user short video interaction data.

[0200] New user short video interaction data is written into the user behavior sequence according to the occurrence time. Based on the updated user behavior sequence, a time-semantic dual-scale variable graph is reconstructed. The behavior nodes are updated by temporal and semantic fusion. Real-time interests, stage interests and long-term stable interests are re-aggregated to generate an updated user preference representation.

[0201] The updated user preference representation is concatenated with the multimodal content vector corresponding to the undistributed candidate short videos, and then input into the improved xDeepFM model. The model is processed sequentially through the linear part, factorization machine part, compressed cross network, interest gating layer and residual feature enhancement layer, and outputs the updated dynamic recommendation score corresponding to each undistributed candidate short video.

[0202] Based on the updated dynamic recommendation scores, the ranking of each candidate short video is recalculated, the progressive topic quota table is updated, the minimum quota for the main preferred topic and the quota for non-mainstream topics are reallocated, and a new recommendation queue is generated.

[0203] The short video distribution continues according to the new recommendation queue, and the interaction data of newly added users' short videos is continuously recorded. The user behavior sequence update, user preference representation update, dynamic recommendation score update and recommendation queue update are repeatedly executed to form a closed-loop iterative recommendation distribution result.

[0204] Example 1: In a continuous short video recommendation session, the platform receives a batch of short video interaction data from anonymous users, including 96 original behavior records and 620 candidate short videos. The original behavior records include exposure location, playback duration, total video duration, skipped location, likes, comments, favorites, shares, search triggers, content theme, creator identifier, style tags, and posting time type. The candidate short videos contain 620 sets of image frames, 620 audio clips, 620 subtitle texts, 620 title texts, and 620 sets of tag fields. The original data contained 11 duplicate exposures, 7 invalid records with a playback duration of less than 1 second, and 3 missing theme fields. After deduplication, anomaly removal, and field completion, 78 valid behavior records were retained, representing an effective rate of 81.25%.

[0205] After the data enters the processing flow, the system generates a sequence of user behaviors according to their order of occurrence. In the first browsing segment, the user generates 22 swipes and views within 30 seconds, meeting the dense condition of at least 20 behaviors within 30 seconds, so this segment is divided into a 30-second sliding window. In the second browsing segment, the user generates only 4 behaviors within 30 seconds. The system continues to expand, reaching a total of 6 behaviors by the 265th second, so this segment is divided into a 265-second sliding window. Another low-frequency browsing segment, after reaching the 5-minute limit, contains only 5 behaviors and is still saved as a separate window. Finally, 78 valid behaviors are divided into 6 sliding windows, with the number of behavior nodes in each window being 22, 6, 13, 17, 5, and 15, respectively.

[0206] The system establishes a time-semantic dual-scale variable graph within each sliding window. Taking the first window as an example, the 22 behavior nodes are numbered in the order of occurrence. Node 1 and Node 2 are separated by 4 seconds, so a time edge is created and weighted 1.0; Node 5 and Node 6 are separated by 16 seconds, so a time edge is created and weighted 0.6; Node 13 and Node 14 are separated by 27 seconds, so a time edge is created and weighted 0.3; Node 18 and Node 19 are separated by 34 seconds. Since the interval exceeds 30 seconds, no time edge is created. Initially, 31 time edges are generated for this window, and 24 are retained after interval filtering. On the semantic side, any two behavior nodes are compared in terms of theme, style, and creator. Node 4 and Node 9 both belong to light food production, with a style distance of 18%, a creator difference of 72%, and a theme similarity of 86%, so a positive weight of 1 is assigned. Node 3 and Node 15 have a theme similarity of 42%, a style distance of 61%, and a creator difference of 20%, all three of which fail to meet the criteria, so a suppression weight of -0.3 is assigned. A total of 312 semantic associations were generated across all windows. The top 25% of positive semantic edges (78) were retained, and 74 negative suppression edges were added.

[0207] The system continues to calculate the activity weight of newly added behavioral nodes. Taking the short video of light food preparation corresponding to node 12 as an example, the video has a total duration of 42 seconds, with 40 seconds actually watched, a completion rate of 95.2%, no skips, and 1 collection, 1 share, and 1 comment. Based on the contribution ratios of 35% completion rate, 15% effective viewing rate, 15% likes, 10% comments, 10% collections, 10% shares, and 5% searches, the activity weight of this node is 0.89. The short video of digital product review corresponding to node 17 has a total duration of 38 seconds, with 6 seconds actually watched, the first part skipped, and no interactive feedback. Its activity weight is 0.19, which is lower than 0.25, so it is marked as a low-activity node and marginalized. Of the 78 behavioral nodes, 9 have a weight lower than 0.25. The remaining 69 nodes undergo time adjacency recalibration, reducing the number of time edges from the original 146 to 91.

[0208] During the dual-path attention propagation phase, the system aggregates near-end node features on the sparse temporal adjacency graph. The one-hop temporal adjacency nodes of node 12 include nodes 10, 11, and 13. Node 11 is aggregated at an 8-second interval (1.0), node 10 at an 18-second interval (0.6), and node 13 at a 26-second interval (0.3), resulting in temporal aggregated features with a mean response of 0.64. Semantic adjacency features are also aggregated: nodes with the same theme but different creators are aggregated at 0.7; nodes with theme similarity between 50% and 80% and complementary styles are aggregated at 0.5; and nodes with the same creator but theme differences exceeding 50% are aggregated at 0.2. After semantic aggregation, the mean feature response of node 12 increases from 0.58 to 0.73. After residual gating fusion, the output update behavior node features are used. The average response of light food preparation nodes increases from 0.49 to 0.76, while the average response of continuously skipped nodes decreases to 0.22.

[0209] The system generates user preference representations based on the updated behavioral node characteristics. The most recent window covers behavioral nodes within the last 10 minutes, totaling 31 nodes: 14 for light meal preparation, 8 for commuting outfits, 5 for life skills, and 4 for other items. The intermediate window covers historical nodes from 30 minutes to 3 hours, totaling 126 nodes: 39 for commuting outfits, 34 for life skills, and 28 for digital product assessments. The far-term window covers historical nodes from more than 3 hours ago but no more than 7 days ago, totaling 486 nodes, with life skills and digital product assessments having the highest proportion. The system assigns a weight of 1.0 to nodes within the most recent window (0–2 minutes), 0.7 to nodes within the last 2–5 minutes, and 0.4 to nodes within the last 5–10 minutes, resulting in an immediate interest vector. For the intermediate window, weights decrease by 0.6, 0.4, and 0.2, resulting in a phased interest vector. For the far-term window, weights decrease by 0.3, 0.2, and 0.1, resulting in a long-term stable interest vector. After normalizing the three types of vectors, they are concatenated to form a 384-dimensional concatenated vector, which is then compressed into a 128-dimensional user preference representation vector.

[0210] On the candidate short video side, the system performs multimodal encoding on 620 candidate short videos. Taking candidate video C219 as an example, the system extracts features from the image side, including food images, operational actions, and finished product images; from the audio side, it extracts the rhythm features of the explanation; from the subtitles and titles, it extracts word vectors such as "low-fat," "breakfast," and "quick preparation"; and from the tags, it extracts light meal preparation, step demonstrations, and close-up shots, ultimately generating a 128-dimensional multimodal content vector. The system concatenates the user preference representation vector with the multimodal content vector and inputs it into the improved xDeepFM model. The linear part is mapped to the 128-dimensional embedding space and generates first-order recommendation features according to the following weighting: 0.80 for dimensions 1–32, 0.60 for dimensions 33–64, 0.40 for dimensions 65–96, and 0.20 for dimensions 97–128. The factorization machine part divides the 128-dimensional embedding features into 8 groups of 16 dimensions each, calculates 28 inner product results, and forms second-order preference matching features. The compressed cross-network performs hierarchical weighting on the cross features of the nearest window, intermediate window, and long-term window. Because the average completion rate of light meal preparation videos reached 91.6% and the number of interactions was 5 within the last 5 minutes, the system increased the weight of the recent window cross feature to 0.8, adjusted the weight of the mid-term window to 0.15, and adjusted the weight of the long-term window to 0.05. The dynamic recommendation score of C219 increased from the initial 0.62 to 0.91, while the score of another video with the same creator and style, C304, decreased from 0.77 to 0.58.

[0211] In the training samples, Sample 001 is a commuting outfit video, displayed in the first position, with an 86% completion rate, 1 like, and an activity weight of 0.72; the training label is positive. Sample 014 is a light meal preparation video, displayed in the fourth position, with a 95% completion rate, 1 favorite, and 1 share; the activity weight is 0.89; the training label is positive. Sample 027 is a digital product review video, displayed in the third position, with an 18% completion rate, skipped at the beginning, and an activity weight of 0.19; the training label is negative. Sample 041 is a life skills video, with a 74% completion rate, 1 comment, and an activity weight of 0.63; the training label is positive. The training set contains 156,000 samples, the validation set contains 28,600 samples, and the test set contains 30,000 samples, with a positive to negative sample ratio of approximately 1.18:1.

[0212] During the recommendation queue generation phase, the system arranges candidate short videos according to dynamic recommendation scores and calculates the aggregation strength of the main preferred themes. The aggregation strength is 0.84 for light food preparation, 0.58 for commuting outfits, 0.47 for life skills, and 0.31 for digital product reviews. The system sets a minimum of 6 videos to be filled for light food preparation, 4 for commuting outfits, and 3 for life skills, with the number of videos filled for any theme not exceeding 40% of the total number of candidates. For videos not included in the main preferred content sequence, novelty is calculated. Candidate videos with a C488 theme difference score of 0.4, creator difference score of 0.3, style difference score of 0.2, and release date freshness score of 0.1 have a novelty score of 1.0, reaching the 0.7 threshold, and are therefore identified as non-mainstream theme insertion targets. When the system scans the main preferred content sequence and finds three consecutive light food preparation videos, it marks the end of these segments with an insertion slot. After inserting C488, the number of consecutive videos with the same theme decreases from three to two.

[0213] In the closed-loop distribution, the system first outputs a queue of 20 recommendations. After the first round of distribution, the average completion rate of the first 10 videos was 64.8%, the skip rate in the first 3 seconds was 21.0%, and the number of interactions was 6. In the traditional method, the first 10 videos were distributed according to static ratings in the same candidate pool, with an average completion rate of 52.3%, a skip rate in the first 3 seconds of 32.7%, and the number of interactions being 3. In the second round of updates, this invention writes 18 newly generated behaviors into the user behavior sequence, regenerates user preference representations and dynamic recommendation ratings, and the average ranking of light food preparation videos improved from 8.7th to 3.1st, while the average ranking of videos by repeat creators decreased from 5.4th to 12.6th.

[0214] On the same batch of test data, the traditional xDeepFM direct sorting method and the method of this invention were compared. The recommended click-through rate was 7.8% for the traditional method and 10.9% for this invention; the average view completion rate was 52.6% for the traditional method and 61.8% for this invention; the first 3-second skip rate was 31.5% for the traditional method and 22.4% for this invention; the combined likes and favorites interaction rate was 5.6% for the traditional method and 8.4% for this invention; the topic duplication rate was 68.2% for the traditional method and 43.5% for this invention; the creator duplication rate was 34.7% for the traditional method and 21.2% for this invention; the Top 10 hit rate was 30.4% for the traditional method and 38.9% for this invention; and the average time for updating the recommendation queue was 2.8 seconds for the traditional method and 0.9 seconds for this invention.

[0215] The above description is only a preferred embodiment 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 user preference-aware short video content recommendation and distribution system, characterized in that, include: The data preprocessing module is used to collect user short video interaction data on short video platforms, perform preprocessing on the user short video interaction data, and generate user behavior sequences. The preference modeling module is used to construct a time-semantic dual-scale variable graph based on user behavior sequences. It establishes temporal and semantic adjacencies within a sliding window, performs temporal and semantic fusion updates on behavior nodes, aggregates immediate interests, stage interests, and long-term stable interests, and splices them together to form a user preference representation. The multimodal coding module is used to perform unified vectorized encoding on the image, audio, subtitle, title and tag features of candidate short videos to generate multimodal content vectors; The rating calculation module is used to concatenate user preference representation and multimodal content vector into the improved xDeepFM model. The improved xDeepFM model connects an interest gating layer and a residual feature enhancement layer in parallel after the linear part, the factorization machine part and the compressed cross network. A time-level attention mechanism is introduced in the compressed cross network to output dynamic recommendation ratings for candidate short videos. The distribution strategy module is used to label candidate short videos with themes, creators, styles and release times based on dynamic recommendation scores, build a progressive theme quota table, first fill the minimum quota of the main preferred theme, then insert and dynamically adjust the quota of non-mainstream themes according to the novelty threshold, and generate a recommendation queue. The closed-loop update module is used to distribute short videos according to the recommendation queue, record the viewing completion rate, skip status and interaction feedback, update the user behavior sequence and user preference representation, and re-input into the improved xDeepFM model for score update, forming a closed-loop iterative recommendation distribution result; The process of performing temporal and semantic fusion updates on behavior nodes, aggregating immediate interests, stage-based interests, and long-term stable interests, and concatenating them to form a user preference representation includes: The sliding window length is adaptively determined based on the density of user behavior sequences. Behavior records within the same sliding window are used as behavior nodes. Temporal relationships are established based on the order of behavior, and semantic relationships are established based on the interaction object, behavior type, and content theme, generating an initial time-semantic dual-scale variable graph. Calculate the activity weight of newly added behavior nodes within each sliding window, perform marginalization processing on nodes with activity below a set threshold, perform elimination processing on nodes that exceed the sliding window, and recalibrate the temporal adjacency relationship between the remaining nodes according to the decay coefficient of the most recent interaction time to generate a sparse temporal adjacency graph. Based on the topic similarity, style distance and creator difference of candidate short videos, positive weights or suppression weights are assigned to the semantic association between behavioral nodes. The semantic adjacency with the highest weight is retained and negative suppression edges are added to form a reinforced semantic adjacency graph. Dual-path attention propagation is performed on the sparse temporal adjacency graph and the enhanced semantic adjacency graph respectively. The features of near-end nodes are aggregated using a decreasing coefficient on the temporal adjacency side, and the features of heterogeneous nodes are fused using a complementary coefficient on the semantic adjacency side. The two aggregation results are fused and updated through a residual gating unit to obtain the updated behavior node features. Based on the updated behavioral node features, the node features in the most recent window are aggregated to generate immediate interests, the node features in the intermediate window are aggregated to generate stage interests, and the node features in the far-term window are aggregated to generate long-term stable interests. The immediate interests, stage interests, and long-term stable interests are then concatenated to generate a user preference representation.

2. The user preference-aware short video content recommendation and distribution system according to claim 1, characterized in that, The user short video interaction data includes playback behavior data and interaction behavior data. Playback behavior data includes playback duration, playback completion rate, skipped position, and number of repeated views. Interaction behavior data includes likes, comments, favorites, forwards, and search behavior data.

3. The user preference-aware short video content recommendation and distribution system according to claim 1, characterized in that, The step of preprocessing user short video interaction data to generate user behavior sequences includes: Acquire user short video interaction data, perform correlation and matching based on user identifier, short video identifier and behavior occurrence time, and generate a set of original behavior records; Perform missing data detection, duplicate behavior detection, and abnormal behavior filtering on the original set of behavior records to generate a cleaned set of behavior records. The cleaned set of behavior records is sorted in ascending order of execution time based on the time of occurrence of the behavior, and adjacent behavior records are sequentially associated to generate a time-series behavior sequence; The playback behavior data and interaction behavior data in the time-series behavior sequence are uniformly encoded, and each behavior record is assigned a corresponding behavior type identifier, time index identifier and short video identifier to generate a standardized behavior sequence. The standardized behavior sequence is continuously divided according to a preset sliding time window, and the user behavior sequence is output.

4. The user preference-aware short video content recommendation and distribution system according to claim 1, characterized in that, The generation of multimodal content vectors includes: Extract continuous image frames according to the playback timeline of candidate short videos, perform size unification, keyframe filtering and visual region segmentation on the continuous image frames, extract the features of objects, scene types, human actions and color changes, and generate image feature sequences. The speech signal is separated from the candidate short video, and noise reduction, frame segmentation and acoustic feature extraction are performed on the speech signal. The semantic content, speech rate changes, emotional state and background sound type of the speech signal are identified, and a speech feature sequence is generated. Obtain the subtitles and titles corresponding to the candidate short videos, perform text cleaning, word segmentation and semantic encoding on the subtitles and titles, extract topic words, entity words, sentiment words and content category features, and generate text feature sequences; Obtain the tags corresponding to the candidate short videos, perform hierarchical mapping according to the theme, content category, style and target audience of the tags, and perform merge encoding on duplicate tags and semantically similar tags to generate tag feature sequences; Based on the candidate short video identifiers, the image feature sequence, speech feature sequence, text feature sequence, and tag feature sequence are associated and aligned, and dimension unification, feature splicing, and normalization are performed to generate a multimodal content vector of the candidate short video.

5. A user preference-aware short video content recommendation and distribution system according to claim 1, characterized in that, The dynamic recommendation score for the output candidate short videos includes: The linear part includes an embedding mapping layer and a first-order weighting layer. The embedding mapping layer maps the high-dimensional feature vector obtained by concatenating the user preference representation vector and the multimodal content vector to a unified low-dimensional embedding space. The first-order weighting layer performs a weighted summation of the embedded feature vectors according to the dimensional weights to generate first-order recommendation features. The factorization machine part includes a latent vector decomposition layer and a second-order interaction layer. The latent vector decomposition layer decomposes the embedded feature vector into several latent vector groups. The second-order interaction layer performs an inner product operation on any two groups of latent vectors and accumulates them to output a second-order preference matching feature. The compressed crossover network consists of multi-level explicit crossover layers and compression mapping layers. The multi-level explicit crossover layers perform explicit crossover on the embedded feature vectors level by level. The compression mapping layer compresses the output according to a set dimension after each level of crossover. A time-level attention mechanism is introduced inside the explicit crossover layer to assign weights to crossover features at different time levels and output time-level high-order crossover features. The interest gating layer includes an interest matching calculation unit and a gating weight generation unit. The interest matching calculation unit generates matching vectors based on the matching degree between real-time interest vectors, phased interest vectors, and long-term stable interest vectors and multimodal content vectors. The gating weight generation unit outputs gating weights based on the matching vectors and performs weighted filtering and fusion on first-order recommendation features, second-order preference matching features, and time-level high-order cross features to generate interest gating features. The residual feature enhancement layer includes a feature mapping unit and a residual fusion unit. The feature mapping unit performs linear mapping on the concatenated feature vectors. The residual fusion unit performs residual superposition and normalization on the linear mapping result and interest-gated features to output residual-enhanced recommendation features. The first-order recommendation features, second-order preference matching features, time-level high-order cross features and residual-enhanced recommendation features are concatenated and processed by fully connected mapping and normalization to generate dynamic recommendation scores for candidate short videos.

6. A user preference-aware short video content recommendation and distribution system according to claim 1, characterized in that, The generated recommendation queue includes: Candidate short videos are sorted from highest to lowest according to the dynamic recommendation score. The theme, creator, style and release time of each candidate short video are read. The theme, creator, style and release time are associated with the candidate short video identifier and dynamic recommendation score to generate a sorted set of candidate short videos. Based on the user's immediate interests, periodic interests, and long-term stable interests, the aggregation strength of different topics in the three types of interests is calculated. The main preferred topics and non-mainstream topics are determined according to the aggregation strength from large to small, and the number of candidate short videos corresponding to each topic is recorded. Based on the aggregation strength of the main preference theme and the number of candidate short videos, a minimum filling quantity is set for each main preference theme. The corresponding short videos are selected sequentially from the candidate short video sorting set according to the dynamic recommendation score to fill the minimum quota of the main preference theme and generate the main preference content sequence. For candidate short videos that are not included in the main preferred content sequence, the differences between the candidate short video's theme, creator, style, and release time and the already filled short videos are compared. The novelty of the candidate short videos is then summarized, and candidate short videos with a novelty of a preset threshold are identified as non-mainstream theme insertion objects. The insertion position of non-mainstream topics is determined by the number of consecutive identical topics in the main preferred content sequence. Non-mainstream topic insertion objects are inserted into the corresponding positions in sequence according to the dynamic recommendation score. The available quota of each non-mainstream topic is adjusted according to the number of repeated topics, repeated creators, and repeated styles of the inserted short videos, and a progressive topic quota table is generated. Based on the progressive theme quota table, short videos on the main preferred theme and non-mainstream theme are extracted one by one. The order of short videos with continuous repetition by creators, continuous repetition of style, and excessive concentration of release time is adjusted to generate a recommendation queue of candidate short videos.

7. A user preference-aware short video content recommendation and distribution system according to claim 1, characterized in that, The recommendation distribution results that form a closed-loop iteration include: Candidate short videos are distributed sequentially according to the recommendation queue. The display time, display position and distribution order of each candidate short video are recorded. The viewing completion rate, skip status and interaction feedback of users during the viewing process are collected to generate new user short video interaction data. New user short video interaction data is written into the user behavior sequence according to the occurrence time. Based on the updated user behavior sequence, a time-semantic dual-scale variable graph is reconstructed. The behavior nodes are updated by temporal and semantic fusion. Real-time interests, stage interests and long-term stable interests are re-aggregated to generate an updated user preference representation. The updated user preference representation is concatenated with the multimodal content vector corresponding to the undistributed candidate short videos, and then input into the improved xDeepFM model. The model is processed sequentially through the linear part, factorization machine part, compressed cross network, interest gating layer and residual feature enhancement layer, and outputs the updated dynamic recommendation score corresponding to each undistributed candidate short video. Based on the updated dynamic recommendation scores, the ranking of each candidate short video is recalculated, the progressive topic quota table is updated, the minimum quota for the main preferred topic and the quota for non-mainstream topics are reallocated, and a new recommendation queue is generated. The short video distribution continues according to the new recommendation queue, and the interaction data of newly added users' short videos is continuously recorded. The user behavior sequence update, user preference representation update, dynamic recommendation score update and recommendation queue update are repeatedly executed to form a closed-loop iterative recommendation distribution result.

Images