Social short video propagation path prediction method based on self-supervised graph model

Abstract

The invention discloses a social short video propagation path prediction method based on a self-supervised graph model, and the method comprises the following steps: collecting data in a social short video platform, and carrying out the preprocessing; constructing a cross-modal propagation graph structure based on the collected data; setting a self-supervision task, constructing self-supervision loss, executing self-supervision training on the cross-modal propagation graph structure, and updating parameters of the propagation graph structure; constructing an energy distribution function, and applying constraints to obtain propagation path energy distribution; introducing an information entropy constraint, and generating a propagation path set optimized by the entropy constraint; executing causal correction to obtain a propagation path set with causal consistency; and executing time continuous evolution modeling to obtain a final social short video propagation path. The method utilizes the self-supervised graph model to predict the short video propagation path, fuses the multi-modal features and introduces the causal time constraint, and has the advantages of high accuracy, strong diversity and good dynamic nature.

Description

Technical Field

[0001] This invention relates to the field of social network analysis, and in particular to a method for predicting the propagation path of short social videos based on a self-supervised graph model. Background Technology

[0002] With the rapid development of social short video platforms, user interactions with videos are becoming increasingly frequent. Predicting propagation paths has become a key aspect of social network analysis and recommendation systems. Existing technologies mostly employ prediction methods based on statistical modeling or traditional graph structures, relying primarily on user interaction frequency, content similarity, and time series features for modeling. While these methods can reflect the patterns of video propagation to some extent, they often struggle to simultaneously consider the fusion effect of multimodal features and have limited ability to model complex propagation chains, making it difficult to truly depict the dynamic propagation process of short videos on social networks.

[0003] While existing research has incorporated graph neural networks, most remain within supervised or semi-supervised paradigms, relying heavily on labeled data and failing to fully exploit potential self-supervised signals. Furthermore, current methods lack energy distribution modeling and entropy constraint mechanisms when optimizing propagation paths, leading to overfitting risks and insufficient path diversity. In addition, the lack of systematic modeling for factors such as user recommendation causality, time decay causality, and interaction-dependent causality makes it difficult to guarantee causal consistency and temporal continuity in prediction results, thus affecting the accuracy and stability of short video propagation path prediction. Summary of the Invention

[0004] One objective of this invention is to propose a method for predicting the propagation path of social short videos based on a self-supervised graph model. This invention utilizes a self-supervised graph model to predict the propagation path of short videos, integrates multimodal features and introduces causal time constraints, and has the advantages of high accuracy, strong diversity and good dynamism.

[0005] A method for predicting the propagation path of social short videos based on a self-supervised graph model according to an embodiment of the present invention includes the following steps:

[0006] Collect user interaction data, video content data, and time series data from social short video platforms, and perform preprocessing to generate multimodal feature data;

[0007] Based on multimodal feature data, a propagation graph structure containing user nodes and video nodes is constructed, social relationship edges and interaction behavior edges are defined, and embedded into the propagation graph structure to form a cross-modal propagation graph structure;

[0008] Set a self-supervised task, construct a self-supervised loss, perform self-supervised training on the cross-modal propagation graph structure, update the propagation graph structure parameters based on the self-supervised loss, and generate a set of propagation representation vectors;

[0009] Energy calculation is performed based on the propagation representation vector set, an energy distribution function is constructed, and energy minimization constraints are applied to the real propagation path, while energy maximization constraints are applied to the random path to obtain the propagation path energy distribution.

[0010] Based on the energy distribution of the propagation path, an information entropy constraint is introduced to generate a set of propagation paths optimized by the entropy constraint.

[0011] The set of propagation paths optimized by entropy constraints is subjected to causal correction based on recommendation mechanism causality, time decay causality, and interaction dependency causality to obtain a set of propagation paths with causal consistency.

[0012] By performing time-continuous evolution modeling on the set of causally consistent propagation paths, the evolution trajectory of the propagation path energy dynamically changes over time, and finally obtaining the predicted propagation path of social short videos.

[0013] Optionally, the generation of the multimodal feature data specifically includes:

[0014] Collect user interaction data, video content data, and time series data, and perform data processing such as format unification, scale standardization, noise removal, and missing value imputation on the collected data;

[0015] Encode user interaction data to generate user interaction feature data;

[0016] The video content data is processed to extract image depth features, audio signal features, and text encoding features, and then fused to generate video content feature data.

[0017] The time series data is normalized and standardized to generate time series feature data, which includes interaction timestamp data and propagation delay parameters.

[0018] By fusing user interaction feature data, video content feature data, and time series feature data, multimodal feature data is generated.

[0019] Optionally, the generation of the cross-modal propagation graph structure specifically includes:

[0020] Based on multimodal feature data, a propagation graph structure is constructed, where each user in the user interaction feature data is represented as a user node, and each video in the video content feature data is represented as a video node, forming a node set composed of user nodes and video nodes.

[0021] Extract social relationships between users from user interaction feature data, establish social relationships as social relationship edges between user nodes, and bind the social relationship edge feature data between users;

[0022] Extract user interaction behavior from user interaction feature data, establish interaction behavior edges between user nodes and video nodes, and bind the corresponding interaction behavior edge feature data.

[0023] Video content feature data is bound to video nodes to obtain video node feature data;

[0024] Bind user interaction feature data to user nodes to obtain user node feature data;

[0025] Bind time series feature data to the interaction behavior edges in the propagation graph structure;

[0026] The user nodes, video nodes, social relationship edges, interaction behavior edges, and their corresponding node feature data and edge feature data are fused in the propagation graph structure to generate a cross-modal propagation graph structure.

[0027] Optionally, the generation of the propagation representation vector set specifically includes:

[0028] Set up an edge masking prediction task to mask the edge feature data of some social relationship edges and interaction behavior edges in the cross-modal propagation graph structure according to a preset ratio, and generate a masked edge set;

[0029] Based on the unmasked node feature data and edge feature data in the cross-modal propagation graph structure, edge feature prediction is performed on each edge in the masked edge set to obtain edge prediction output data. The edge prediction output data is then compared with the real edge feature data to generate the edge mask prediction loss value.

[0030] A node context comparison task is set up. Positive sample node pairs and negative sample node pairs are sampled for user nodes and video nodes in the cross-modal propagation graph structure, respectively. The cosine similarity between node pairs is calculated based on node feature data, and combined with label information, a node context comparison loss value is generated.

[0031] The path reconstruction task is defined. A portion of the propagation path segments are randomly extracted from the cross-modal propagation graph structure to form a path mask set. The path mask set is reconstructed based on the existing propagation representation vector set to obtain the path reconstruction output data. The path reconstruction output data is then compared with the real path data to generate the path reconstruction loss value.

[0032] The edge mask prediction loss, node context comparison loss, and path reconstruction loss are weighted and summed according to preset weights to generate a self-supervised loss.

[0033] Based on self-supervised loss, parameter updates are performed on user node feature data, video node feature data, social relationship edge feature data, and interaction behavior edge feature data in the cross-modal propagation graph structure to generate an updated propagation graph structure parameter set.

[0034] Based on the updated propagation graph structure parameter set, information aggregation calculation is performed on the cross-modal propagation graph structure to generate a propagation representation vector set.

[0035] Optionally, the generation of the propagation path energy distribution specifically includes:

[0036] Based on the propagation representation vector set, user node representation vectors, video node representation vectors, and corresponding social relationship edge feature data and interaction behavior edge feature data in the real propagation path are extracted to generate real path feature data.

[0037] The real path feature data is processed to calculate the energy value of the real propagation path. The energy value is calculated by calculating the cosine similarity of the combined representation vector of each pair of adjacent nodes in the real path and the edge feature data of their connecting edges, and all cosine similarities are accumulated. The result is defined as the energy value of the real path.

[0038] Negative sample propagation paths are generated by randomly sampling from the propagation representation vector set. User node representation vectors, video node representation vectors, and edge feature data are extracted from these paths to generate random path feature data.

[0039] The random path feature data is processed to calculate the energy value of the random propagation path. The energy value is calculated by calculating the cosine similarity between the combined representation vector of each pair of adjacent nodes in the random path and the edge feature data of their connecting edges. All cosine similarities are accumulated, and the result is defined as the random path energy value.

[0040] Apply an energy minimization constraint to the actual propagation path;

[0041] Imposing an energy maximization constraint on the random propagation path;

[0042] Under the combined influence of energy minimization and energy maximization constraints, an energy distribution function is constructed by calculating the difference between the energy values ​​of the actual path and the energy values ​​of the random path.

[0043] The propagation representation vector set is optimized based on the energy distribution function to generate the propagation path energy distribution.

[0044] Optionally, the generation of the entropy-constrained optimized propagation path set specifically includes:

[0045] Based on the propagation path energy distribution, the path energy weighting value corresponding to each candidate propagation path is extracted to form a path energy weighting value set;

[0046] Normalize the set of path energy weights to obtain the set of candidate propagation path probability values;

[0047] Calculate the propagation path information entropy value based on the set of candidate propagation path probability values;

[0048] The propagation path information entropy value is compared with a preset entropy threshold. When the propagation path information entropy value is less than the preset entropy threshold, the candidate propagation path probability value set is redistributed. The distribution of candidate propagation path probability values ​​is adjusted while keeping the relative magnitude of the path energy weighting values ​​unchanged, thus forming a corrected candidate propagation path probability value set.

[0049] The propagation path information entropy value is recalculated based on the corrected candidate propagation path probability value set. If the corrected propagation path information entropy value is greater than or equal to the preset entropy threshold, the corrected candidate propagation path probability value set is confirmed as the final candidate propagation path probability value set. If the corrected propagation path information entropy value is still less than the preset entropy threshold, the reallocation process continues until the preset entropy threshold constraint is met.

[0050] Output the set of propagation paths optimized by entropy constraints.

[0051] Optionally, the generation of the set of causal consistency propagation paths specifically includes:

[0052] Based on the set of propagation paths optimized by entropy constraints, the probability value of each candidate propagation path is extracted to form a set of candidate propagation path probability values.

[0053] Based on the set of probability values ​​of candidate propagation paths, causal constraints of the recommendation mechanism are constructed, and causal correction factors of the recommendation mechanism are generated according to the frequency of recommendation triggering behavior, thus forming a set of causal correction factors of the recommendation mechanism.

[0054] Based on the probability value set of candidate propagation paths and time series data, a time decay causal constraint is constructed. This constraint is established by calculating the time interval between adjacent interactions in the candidate propagation path. When the time interval is greater than a preset time threshold, a time decay factor is assigned to the candidate propagation path. When the time interval is less than or equal to the preset time threshold, a time preservation factor is assigned to the candidate propagation path. All factors form a time factor set.

[0055] Based on the probability value set of candidate propagation paths and user interaction feature data, interaction dependency causal constraints are constructed. These constraints are established by analyzing multiple interaction behaviors of user nodes in candidate propagation paths. When user nodes exhibit multiple interaction behaviors in candidate propagation paths, interaction dependency factors are assigned to the candidate propagation paths. When user nodes exhibit only a single interaction behavior in candidate propagation paths, interaction independence factors are assigned to the candidate propagation paths. All factors form an interaction factor set.

[0056] For each candidate propagation path, the causal correction factor, time factor and interaction factor of the recommendation mechanism are applied to the corresponding candidate propagation path probability value to generate the corrected candidate propagation path probability value, forming a set of causal corrected candidate propagation path probability values.

[0057] The propagation path information entropy value is recalculated based on the causal correction candidate propagation path probability value set. When the propagation path information entropy value is greater than or equal to the preset entropy threshold, the causal correction candidate propagation path probability value set is confirmed as the final candidate propagation path probability value set. When the propagation path information entropy value is less than the preset entropy threshold, the causal correction factor set, time factor set, and interaction factor set of the recommendation mechanism are redistributed until the propagation path information entropy value is greater than or equal to the preset entropy threshold.

[0058] Output the set of propagation paths that maintain causal consistency.

[0059] Optionally, the generation of the final predicted social short video propagation path specifically includes:

[0060] The set of propagation paths with causal consistency is sorted in chronological order, and the causal-corrected probability value of each candidate propagation path is bound to the corresponding timestamp to obtain a set of path time state vectors.

[0061] The time evolution calculation is performed on the path time state vector set. The time state of each candidate propagation path is iteratively calculated based on the continuous time evolution function. The energy change value of the candidate propagation path at adjacent time scales is output, and the path energy time series is generated.

[0062] In the process of calculating the path energy time series, the energy change value of each candidate propagation path on the continuous time scale is accumulated to form a set of dynamic cumulative path energy values. The dynamic cumulative path energy values ​​are compared with a preset time evolution threshold. When the value is greater than the time evolution threshold, a time enhancement weight is assigned to the propagation path. When the value is less than or equal to the time evolution threshold, a time preservation weight is assigned to the propagation path, thus obtaining a set of path time weights.

[0063] The path time weight set is combined with the path energy time series one by one to generate the path weighted energy time series, and the energy dynamic trajectory of the candidate propagation path is recorded on a continuous time scale to form the energy evolution trajectory set of the candidate propagation path.

[0064] Normalization is performed on the set of energy evolution trajectories to obtain the set of evolution trajectories in which the propagation path energy changes dynamically over time. The set of evolution trajectories is then used as the final predicted propagation path for social short videos.

[0065] The beneficial effects of this invention are:

[0066] This invention introduces a self-supervised graph model to establish a novel modeling approach for predicting the propagation path of social short videos. It addresses the shortcomings of existing technologies, such as insufficient ability to characterize propagation chains, inadequate fusion of multimodal features, and a lack of systematic modeling of causal relationships and temporal evolution mechanisms. In this invention, user interaction data, video content data, and time-series data are uniformly incorporated into a multimodal feature data framework. A cross-modal propagation graph structure is constructed to achieve deep association between user nodes and video nodes. This ensures that the propagation path not only reflects the interaction patterns between users and videos but also embodies the fusion characteristics of multimodal information during propagation. Through the design of a self-supervised task, the system can effectively learn the latent semantic representations of nodes and paths even in the absence of large-scale labeled data, guaranteeing the accuracy and completeness of the propagation representation vector set.

[0067] In terms of energy modeling, this invention introduces an energy distribution function into the short video propagation prediction task, and applies energy minimization constraints to real paths and energy maximization constraints to random paths, effectively distinguishing between real propagation paths and noise paths. At the same time, combined with the information entropy constraint optimization mechanism, it not only ensures the diversity of the propagation path set, but also prevents overfitting of the prediction results, thereby improving the flexibility of prediction while ensuring stability. In terms of causal modeling, this invention proposes three types of constraints: recommendation mechanism causality, time decay causality, and interaction dependency causality, and modifies the probability values ​​of candidate propagation paths one by one to generate a causally consistent propagation path set. This enables the prediction results to truly reflect the comprehensive impact of recommendation triggering, time intervals, and multiple user interaction behaviors on the propagation path, overcoming the shortcomings of existing technologies in causal consistency modeling.

[0068] Furthermore, by constructing a time-continuous evolution function, this invention further dynamically models the set of causally consistent propagation paths in the time dimension, generating the evolution trajectory of propagation path energy over time. This more realistically depicts the propagation patterns of short videos on social networks, ensuring that the prediction results possess temporal continuity and dynamic evolution characteristics. The resulting social short video propagation paths not only outperform traditional methods in accuracy but also show significant improvements in maintaining propagation diversity, ensuring reasonable causal relationships, and exhibiting temporal dynamism. Attached Figure Description

[0069] 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:

[0070] Figure 1 This is a flowchart of a social short video propagation path prediction method based on a self-supervised graph model proposed in this invention;

[0071] Figure 2 This is a schematic diagram of the cross-modal propagation graph structure construction for a social short video propagation path prediction method based on a self-supervised graph model proposed in this invention. Detailed Implementation

[0072] 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.

[0073] refer to Figures 1-2 A method for predicting the propagation path of social short videos based on a self-supervised graph model includes the following steps:

[0074] Collect user interaction data, video content data, and time series data from social short video platforms, and perform preprocessing to generate multimodal feature data;

[0075] Based on multimodal feature data, a propagation graph structure containing user nodes and video nodes is constructed, social relationship edges and interaction behavior edges are defined, and embedded into the propagation graph structure to form a cross-modal propagation graph structure;

[0076] Set a self-supervised task, construct a self-supervised loss, perform self-supervised training on the cross-modal propagation graph structure, update the propagation graph structure parameters based on the self-supervised loss, and generate a set of propagation representation vectors;

[0077] Energy calculation is performed based on the propagation representation vector set, an energy distribution function is constructed, and energy minimization constraints are applied to the real propagation path, while energy maximization constraints are applied to the random path to obtain the propagation path energy distribution.

[0078] Based on the energy distribution of propagation paths, information entropy constraints are introduced to ensure that the propagation path set maintains diversity and stability, thereby generating a propagation path set optimized by entropy constraints.

[0079] The set of propagation paths optimized by entropy constraints is subjected to causal correction based on recommendation mechanism causality, time decay causality, and interaction dependency causality to obtain a set of propagation paths with causal consistency.

[0080] By performing time-continuous evolution modeling on the set of causally consistent propagation paths, the evolution trajectory of the propagation path energy dynamically changes over time, and finally obtaining the predicted propagation path of social short videos.

[0081] In this embodiment, the generation of the multimodal feature data specifically includes:

[0082] Collect user interaction data, video content data, and time series data, and perform data processing such as format unification, scale standardization, noise removal, and missing value imputation on the collected data;

[0083] User interaction data is encoded to generate user interaction feature data, which includes like behavior data, comment behavior data, and forwarding behavior data. Like behavior data refers to the interaction data generated when a user node performs a like operation on a video node, including like tag value and like count value. Comment behavior data refers to the interaction data generated when a user node submits a comment on a video node, including comment content count value and comment count value. Forwarding behavior data refers to the interaction data generated when a user node performs a forwarding operation on a video node, including forwarding tag value and forwarding count value.

[0084] The video content data is processed to extract image depth features, audio signal features, and text encoding features, and then fused to generate video content feature data, which includes video frame sequences, audio trajectory data, and text description information.

[0085] The time series data is normalized and standardized to generate time series feature data, which includes interaction timestamp data and propagation delay parameters.

[0086] By fusing user interaction feature data, video content feature data, and time series feature data, multimodal feature data is generated.

[0087] In this embodiment, the generation of the cross-modal propagation graph structure specifically includes:

[0088] Based on multimodal feature data, a propagation graph structure is constructed, where each user in the user interaction feature data is represented as a user node, and each video in the video content feature data is represented as a video node, forming a node set composed of user nodes and video nodes.

[0089] Extract social relationships between users from user interaction feature data, establish social relationships as social relationship edges between user nodes, and bind the social relationship edge feature data between users;

[0090] The extraction of social relationships between users specifically includes: generating a candidate set of user social connections based on user interaction feature data to obtain a candidate user pair set value; for each pair of user nodes in the candidate user pair set, counting the number of direct interactions between the two users to obtain an interaction frequency value; counting the number of mutual friends between the two users to obtain a mutual friend quantity value; performing a weighted summation of the interaction frequency value and the mutual friend quantity value according to a preset weight to obtain a user relationship strength value; further calling the interaction timestamp data in the time series feature data to extract the relationship establishment time value to obtain a relationship establishment time value, and calculating the relationship maintenance duration value to obtain a relationship duration value; using the user relationship strength value, relationship establishment time value, and relationship duration value together as social connection features between users, binding them to the social relationship edges between user nodes to generate social relationship edge feature data;

[0091] Extract user interaction behavior from user interaction feature data, establish interaction behavior edges between user nodes and video nodes, and bind the corresponding interaction behavior edge feature data.

[0092] The extraction of user-video interaction behavior specifically includes: generating a candidate set of user-video interactions based on user interaction feature data to obtain a candidate interaction pair set value; for each user node and video node in the candidate interaction pair set, identifying whether a like behavior exists to obtain a like tag value; counting the number of times a comment behavior occurs to obtain a comment count value; counting the number of times a forwarding behavior occurs to obtain a forwarding count value; further combining the like tag value, comment count value, and forwarding count value to generate user-video interaction behavior feature values; and using the interaction behavior feature values ​​as interaction behavior edge feature data to bind to the interaction behavior edge between the user node and the video node.

[0093] Video content feature data is bound to video nodes, giving the video nodes visual, audio, and text features, thus obtaining video node feature data.

[0094] By binding user interaction feature data to user nodes, user nodes acquire features such as liking behavior, commenting behavior, and forwarding behavior, thus obtaining user node feature data.

[0095] Bind time series feature data to the interaction behavior edges in the propagation graph structure, so that the interaction behavior edges have interaction timestamp and propagation delay parameter features;

[0096] The user nodes, video nodes, social relationship edges, interaction behavior edges, and their corresponding node feature data and edge feature data are fused in the propagation graph structure to generate a cross-modal propagation graph structure.

[0097] In this embodiment, the generation of the propagation representation vector set specifically includes:

[0098] Set up an edge masking prediction task to mask the edge feature data of some social relationship edges and interaction behavior edges in the cross-modal propagation graph structure according to a preset ratio, and generate a masked edge set;

[0099] Based on the unmasked node feature data and edge feature data in the cross-modal propagation graph structure, edge feature prediction is performed on each edge in the masked edge set to obtain edge prediction output data. The edge prediction output data is then compared with the real edge feature data to generate the edge mask prediction loss value.

[0100] The specific steps of performing edge feature prediction on each edge in the masked edge set include: for each masked edge, determining its associated user node or video node, extracting the corresponding node feature data to obtain node input data; using the connection relationship between the masked edge and adjacent nodes in the cross-modal propagation graph structure, extracting the unmasked adjacent edge feature data to obtain adjacent edge input data; concatenating the node input data and the adjacent edge input data to generate edge feature input data; using the edge feature input data to perform feature propagation and aggregation in the cross-modal propagation graph structure to generate edge feature estimates; and using the edge feature estimates as the edge prediction output data for the masked edge to replace the masked real edge feature data.

[0101] A node context comparison task is set up. Positive sample node pairs and negative sample node pairs are sampled for user nodes and video nodes in the cross-modal propagation graph structure, respectively. The cosine similarity between node pairs is calculated based on node feature data, and combined with label information, a node context comparison loss value is generated.

[0102] The generation of the node context comparison loss value specifically includes: sampling node pairs for user nodes and video nodes respectively in the cross-modal propagation graph structure to obtain positive sample node pair set values ​​and negative sample node pair set values; extracting node feature data for each node in the positive sample node pair set and negative sample node pair set to generate node feature input data; calculating the cosine similarity between node pairs based on the node feature input data to obtain node similarity distribution values; comparing the node similarity distribution values ​​with the label information of the positive sample node pair set and negative sample node pair set to generate node comparison error values; and summarizing all node comparison error values ​​to obtain the node context comparison loss value.

[0103] The path reconstruction task is defined. A portion of the propagation path segments are randomly extracted from the cross-modal propagation graph structure to form a path mask set. The path mask set is reconstructed based on the existing propagation representation vector set to obtain the path reconstruction output data. The path reconstruction output data is then compared with the real path data to generate the path reconstruction loss value.

[0104] The generation of the path reconstruction loss value specifically includes: randomly selecting a portion of the propagation path segments in the cross-modal propagation graph structure to obtain a path mask set value; extracting the node feature data and edge feature data of each propagation path segment in the path mask set to generate path feature input data; calling an existing propagation representation vector set to perform a completion operation on the path feature input data to generate path reconstruction output data; comparing the path reconstruction output data with the real path data corresponding to the path mask set one by one to obtain the path reconstruction error value; and summarizing all path reconstruction error values ​​to generate the path reconstruction loss value.

[0105] The existing set of propagation representation vectors refers to the set of user node representation vectors, video node representation vectors, and propagation path representation vectors generated by information aggregation calculation after updating the parameters of user node feature data, video node feature data, social relationship edge feature data, and interaction behavior edge feature data based on self-supervised loss in a cross-modal propagation graph structure.

[0106] The edge mask prediction loss, node context comparison loss, and path reconstruction loss are weighted and summed according to preset weights to generate a self-supervised loss. The preset weights correspond to the contribution ratios of the three types of self-supervised tasks in the training process, respectively.

[0107] Based on self-supervised loss, parameter updates are performed on user node feature data, video node feature data, social relationship edge feature data, and interaction behavior edge feature data in the cross-modal propagation graph structure to generate an updated propagation graph structure parameter set.

[0108] Based on the updated propagation graph structure parameter set, information aggregation calculation is performed on the cross-modal propagation graph structure to generate a propagation representation vector set. The propagation representation vector set includes user node representation vectors, video node representation vectors, and propagation path representation vectors. The propagation path representation vector is a path-level feature representation obtained by aggregating the user node representation vectors, video node representation vectors, and corresponding edge feature data on the propagation path. It is a component of the propagation representation vector set and is used to describe the propagation mode of the entire propagation path.

[0109] In this embodiment, the generation of the propagation path energy distribution specifically includes:

[0110] Based on the propagation representation vector set, user node representation vectors, video node representation vectors, and corresponding social relationship edge feature data and interaction behavior edge feature data in the real propagation path are extracted to generate real path feature data.

[0111] The real path feature data is processed to calculate the energy value of the real propagation path. The energy value is calculated by calculating the cosine similarity of the combined representation vector of each pair of adjacent nodes in the real path and the edge feature data of their connecting edges, and all cosine similarities are accumulated. The result is defined as the energy value of the real path.

[0112] Negative sample propagation paths are generated by randomly sampling from the propagation representation vector set. User node representation vectors, video node representation vectors, and edge feature data are extracted from these paths to generate random path feature data.

[0113] The random path feature data is processed to calculate the energy value of the random propagation path. The energy value is calculated by calculating the cosine similarity between the combined representation vector of each pair of adjacent nodes in the random path and the edge feature data of their connecting edges. All cosine similarities are accumulated, and the result is defined as the random path energy value.

[0114] Apply an energy minimization constraint to the real propagation path, requiring the energy value of the real path to converge in the direction of energy minimization;

[0115] Apply an energy maximization constraint to the random propagation path, requiring the energy value of the random path to converge in the direction of energy maximization;

[0116] Under the combined influence of energy minimization and energy maximization constraints, an energy distribution function is constructed by calculating the difference between the energy values ​​of the real path and the random path, which is used to characterize the energy difference between the two.

[0117] The propagation path energy distribution is generated by optimizing the propagation representation vector set based on the energy distribution function. Specifically, this includes: calculating the initial path energy value for each candidate propagation path in the propagation representation vector set to obtain a set of initial path energy values; inputting the set of initial path energy values ​​into the energy distribution function to generate a set of path energy differences, and performing scaling to generate corresponding normalized path energy values; comparing the normalized path energy values ​​with a preset energy threshold, assigning a weight value of 1 to paths with normalized values ​​less than the energy threshold, and assigning a weight value of 0 to paths with normalized values ​​greater than or equal to the energy threshold, to obtain a set of path energy weight values; for each candidate propagation path, multiplying the normalized energy value of the path with the corresponding energy weight value to generate a path energy weight value, and summing all path energy weight values ​​to form the propagation path energy distribution.

[0118] The candidate propagation path refers to multiple possible propagation sequences composed of user nodes and video nodes arranged in the order of their connection relationships in the propagation graph structure. These sequences are used as comparison objects between real propagation paths and random propagation paths during the energy calculation process.

[0119] In this embodiment, the generation of the entropy-constrained optimized propagation path set specifically includes:

[0120] Based on the propagation path energy distribution, the path energy weighting value corresponding to each candidate propagation path is extracted to form a path energy weighting value set;

[0121] Normalization is performed on the path energy weighted value set to obtain the candidate propagation path probability value set. Each candidate propagation path probability value in the candidate propagation path probability value set is between 0 and 1, and the sum of all candidate propagation path probability values ​​is equal to 1.

[0122] The propagation path information entropy value is calculated based on the set of candidate propagation path probability values. The propagation path information entropy value is obtained by multiplying the candidate propagation path probability value with the corresponding logarithm of the candidate propagation path probability value and then summing them. It is used to measure the diversity level of the candidate propagation path set.

[0123] The propagation path information entropy value is compared with a preset entropy threshold. When the propagation path information entropy value is less than the preset entropy threshold, the candidate propagation path probability value set is redistributed. The distribution of candidate propagation path probability values ​​is adjusted while keeping the relative magnitude of the path energy weighting values ​​unchanged, thus forming a corrected candidate propagation path probability value set.

[0124] During the adjustment of the candidate propagation path probability value set, the candidate propagation paths are sorted according to the magnitude of their path energy weights to obtain a path sorting sequence. Based on the path sorting sequence, an initial proportional coefficient is assigned to each candidate propagation path according to the magnitude of its path energy weights, forming a path proportional coefficient set. The candidate propagation path probability value set is then balanced based on the path proportional coefficient set. While ensuring the path sorting sequence remains unchanged, the candidate propagation path probability values ​​are redistributed to new probability values, generating a corrected candidate propagation path probability value set. The corrected candidate propagation path probability value set is then normalized to ensure that the sum of all corrected candidate propagation path probability values ​​equals 1, and the corrected candidate propagation path probability value set is output.

[0125] The propagation path information entropy value is recalculated based on the corrected candidate propagation path probability value set. If the corrected propagation path information entropy value is greater than or equal to the preset entropy threshold, the corrected candidate propagation path probability value set is confirmed as the final candidate propagation path probability value set. If the corrected propagation path information entropy value is still less than the preset entropy threshold, the reallocation process continues until the preset entropy threshold constraint is met.

[0126] The output is a set of propagation paths optimized by entropy constraints. Each candidate propagation path corresponds to a candidate propagation path probability value after being corrected by entropy constraints. The set of propagation paths maintains the diversity and stability of the distribution.

[0127] In this embodiment, the generation of the set of causal consistency propagation paths specifically includes:

[0128] Based on the set of propagation paths optimized by entropy constraints, the probability value of each candidate propagation path is extracted to form a set of candidate propagation path probability values.

[0129] Based on the set of probability values ​​of candidate propagation paths, causal constraints of the recommendation mechanism are constructed, and causal correction factors of the recommendation mechanism are generated according to the frequency of recommendation triggering behavior, thus forming a set of causal correction factors of the recommendation mechanism.

[0130] In constructing the causal constraints of the recommendation mechanism, user node and video node pairs are extracted from each candidate propagation path based on the candidate propagation path probability value set, forming a user-video node pair set. Based on this set, the frequency of recommendation triggering behavior for each user node towards the target video node is statistically analyzed. This triggering behavior includes recommendation exposure and recommendation click behavior, yielding recommendation exposure frequency and recommendation click frequency values, respectively. The recommendation exposure frequency and recommendation click frequency values ​​are then weighted to generate a causal correction factor for each user-video node pair, forming a causal correction factor set. Finally, this causal correction factor set is bound to the candidate propagation path probability value set one by one.

[0131] Based on the probability value set of candidate propagation paths and time series data, a time decay causal constraint is constructed. This constraint is established by calculating the time interval between adjacent interactions in the candidate propagation path. When the time interval is greater than a preset time threshold, a time decay factor is assigned to the candidate propagation path. When the time interval is less than or equal to the preset time threshold, a time preservation factor is assigned to the candidate propagation path. All factors form a time factor set.

[0132] In constructing the time decay causal constraint, the timestamps of each adjacent interaction in the candidate propagation path are extracted sequentially from the time series data. The time interval between adjacent interactions is calculated and compared with a preset time threshold. When the time interval is greater than the preset time threshold, a time decay factor is assigned to the corresponding candidate propagation path, indicating that the propagation intensity of the path decays over time. When the time interval is less than or equal to the preset time threshold, a time retention factor is assigned to the corresponding candidate propagation path, indicating that the propagation intensity of the path remains stable over time. The time decay factors and time retention factors assigned to all candidate propagation paths are recorded one by one to form a time factor set, and the time factor set is bound to the candidate propagation path probability value set.

[0133] The time decay factor value is obtained by dividing a preset time threshold by the time interval value, and the time retention factor value is fixed at 1.

[0134] Based on the probability value set of candidate propagation paths and user interaction feature data, interaction dependency causal constraints are constructed. These constraints are established by analyzing multiple interaction behaviors of user nodes in candidate propagation paths. When user nodes exhibit multiple interaction behaviors in candidate propagation paths, interaction dependency factors are assigned to the candidate propagation paths. When user nodes exhibit only a single interaction behavior in candidate propagation paths, interaction independence factors are assigned to the candidate propagation paths. All factors form an interaction factor set.

[0135] In constructing the interaction dependency causal constraints, the interaction behavior types of each user node in the candidate propagation path are extracted from the user interaction feature data, and the number of interaction behaviors of each user node in the candidate propagation path is recorded. When the number of interaction behaviors of a user node in the candidate propagation path is greater than 1, an interaction dependency factor is assigned to the candidate propagation path, indicating that the path is affected by multiple interaction behaviors. When the number of interaction behaviors of a user node in the candidate propagation path is equal to 1, an interaction independence factor is assigned to the candidate propagation path, indicating that the path is affected by only a single interaction behavior. The interaction dependency factor and interaction independence factor values ​​assigned to all candidate propagation paths are recorded one by one to form an interaction factor set, and the interaction factor set is correspondingly bound to the candidate propagation path probability value set.

[0136] The interaction dependency factor and interaction independence factor are respectively assigned the value of the interaction behavior quantity value divided by 3 and the fixed value of 0.5;

[0137] For each candidate propagation path, the causal correction factor, time factor and interaction factor of the recommendation mechanism are applied to the corresponding candidate propagation path probability value to generate the corrected candidate propagation path probability value, forming a set of causal corrected candidate propagation path probability values.

[0138] For each candidate propagation path, the corresponding causal correction factor of the recommendation mechanism is extracted from the causal correction factor set of the recommendation mechanism, and the causal correction factor of the recommendation mechanism is multiplied by the probability value of the candidate propagation path to obtain the recommendation correction probability; the corresponding time factor is extracted from the time factor set, and the time factor is multiplied by the recommendation correction probability value to obtain the time correction probability; the corresponding interaction factor is extracted from the interaction factor set, and the interaction factor is multiplied by the time correction probability value to obtain the interaction correction probability; the interaction correction probabilities of all candidate propagation paths are recorded one by one to form a set of causal correction candidate propagation path probability values.

[0139] The propagation path information entropy value is recalculated based on the causal correction candidate propagation path probability value set. When the propagation path information entropy value is greater than or equal to the preset entropy threshold, the causal correction candidate propagation path probability value set is confirmed as the final candidate propagation path probability value set. When the propagation path information entropy value is less than the preset entropy threshold, the causal correction factor set, time factor set, and interaction factor set of the recommendation mechanism are redistributed until the propagation path information entropy value is greater than or equal to the preset entropy threshold.

[0140] Output a set of causal consistent propagation paths, which satisfy the causal constraints of the recommendation mechanism, the time decay causal constraints, and the interactive dependency causal constraints, while maintaining the diversity and stability of the propagation path set.

[0141] In this embodiment, the generation of the final predicted social short video propagation path specifically includes:

[0142] The set of propagation paths with causal consistency is sorted in chronological order, and the causal-corrected probability value of each candidate propagation path is bound to the corresponding timestamp to obtain a set of path time state vectors.

[0143] The time evolution calculation is performed on the path time state vector set. The time state of each candidate propagation path is iteratively calculated based on the continuous time evolution function. The energy change value of the candidate propagation path at adjacent time scales is output, and the path energy time series is generated.

[0144] The aforementioned continuous evolution function is a computational function used to characterize the dynamic energy changes of a set of propagation paths with causal consistency across continuous time scales. It takes the path time state vector of a candidate propagation path at its initial moment as input. This path time state vector consists of the causally corrected probability value of the candidate propagation path and its corresponding timestamp, and is calculated in conjunction with a set of time factors and a set of causal correction factors. The time factor set includes a time decay factor and a time preservation factor, while the causal correction factor set includes a recommendation mechanism causal correction factor, a time factor, and an interaction factor. The continuous evolution function iteratively calculates and outputs the energy change values ​​of the candidate propagation path at adjacent time scales, forming a path energy time series. This path energy time series consists of the path energy value corresponding to each time scale, serving as the basis for generating the propagation path energy evolution trajectory.

[0145] In the specific implementation process, the time continuous evolution function is constructed through neural differential equations. The path time state vector of the candidate propagation path is used as the dynamic system state variable, and the time factor set and the causal correction factor set are used as the differential equation constraints. The energy change value of the candidate propagation path on the continuous time scale is calculated by numerical integration method, thereby generating the propagation path energy evolution trajectory that can reflect the dynamic characteristics of social short video propagation.

[0146] In the process of calculating the path energy time series, the energy change value of each candidate propagation path on the continuous time scale is accumulated to form a set of dynamic cumulative path energy values. The dynamic cumulative path energy values ​​are compared with a preset time evolution threshold. When the value is greater than the time evolution threshold, a time enhancement weight is assigned to the propagation path. When the value is less than or equal to the time evolution threshold, a time preservation weight is assigned to the propagation path, thus obtaining a set of path time weights.

[0147] The value of the time enhancement weight is set to 1.5, and the value of the time retention weight is set to 1;

[0148] The path time weight set is combined with the path energy time series one by one to generate the path weighted energy time series, and the energy dynamic trajectory of the candidate propagation path is recorded on a continuous time scale to form the energy evolution trajectory set of the candidate propagation path.

[0149] Normalization is performed on the set of energy evolution trajectories to obtain the set of evolution trajectories in which the propagation path energy changes dynamically over time. The set of evolution trajectories is then used as the final predicted propagation path for social short videos.

[0150] Example 1:

[0151] To verify the feasibility of this invention in practice, it was applied to a study on the propagation path prediction of a social short video platform. This platform has a large user base, with an extremely high number of videos uploaded and interacted with daily. The propagation chain is complex and dynamically changing. The platform needs to comprehensively model multiple aspects such as recommendation mechanisms, user interactions, video content, and time series. However, existing technologies often only focus on single-modal data or only speculate on propagation paths at fixed time segments, which fails to effectively reflect the multi-dimensional causal relationships and dynamic evolutionary characteristics in the social propagation process. Especially when facing rapidly spreading video content, conventional methods are prone to problems such as prediction lag, single path, and insufficient diversity, thus affecting the accuracy of the platform's decision-making in recommendation distribution, information governance, and content evaluation. This invention introduces a cross-modal propagation graph structure through a self-supervised graph model, combining user nodes, video nodes, social relationship edges, and interaction behavior edges. With deep fusion of multi-source data, a complete set of propagation representation vectors is formed. Then, through energy calculation and causal correction mechanisms, accurate prediction of short video propagation paths is achieved, effectively solving the above-mentioned problems.

[0152] In practical applications, the research team collects user interaction data, including likes, comments, and reposts, in the platform's data center. Combining video image features, audio signals, and textual descriptions, along with timestamps and propagation delay parameters, unified multimodal feature data is generated after preprocessing. Subsequently, a cross-modal propagation graph structure is constructed based on these features, abstracting user relationships as social relationship edges and user-video interactions as interaction behavior edges. By combining node features and temporal information, a graph representation of the propagation chain is achieved. During the training phase, this invention sets three self-supervised tasks: edge mask prediction, node context comparison, and path reconstruction. A propagation representation vector set is generated by continuously updating the propagation graph structure parameters. Then, an energy distribution function is introduced to impose energy constraints on real and random paths, and information entropy is used to ensure the diversity and stability of path distribution. Further corrections are made through recommendation mechanism causality, time decay causality, and interaction dependency causality constraints to ensure that the prediction results match the actual propagation logic. A time-continuous evolution function is used to dynamically model the causally consistent propagation path set, obtaining the temporal evolution trajectory of the propagation path energy, thus outputting the final prediction of the short video propagation path.

[0153] In a validation study, the research team selected video propagation data from the platform within a certain time window, applied the method of this invention for modeling and prediction, and compared the results with those of existing algorithms on the platform. Experiments showed that the propagation paths generated by this invention better reflect the interaction between users and videos, accurately capturing the time decay effect and diverse propagation characteristics during the propagation process. Through dynamic modeling of temporal evolution, this invention not only improves overall prediction accuracy but also demonstrates significant advantages in maintaining path stability and diversity. In application scenarios, this means that the platform can identify potentially high-propagation videos earlier, effectively control the risk of abnormal propagation, and balance content quality and user interests in recommendation and distribution.

[0154] To verify the performance of the present invention in practice, it was compared with traditional methods, and the results are shown in Table 1.

[0155] Table 1. Comparison of the effectiveness of the method of the present invention and the traditional method in estimating the water depth and storage capacity of ponds.

[0156] method Prediction accuracy (%) Path diversity index Path stability index Average computation delay (s) Information entropy preservation value Single-modal feature method 72.8 0.41 0.46 12.3 1.82 Traditional machine learning methods 75.6 0.44 0.49 10.7 1.95 The method proposed in this invention 86.9 0.63 0.71 8.4 2.74

[0157] The experimental results show that the method of the present invention is superior to traditional methods in terms of prediction accuracy, path diversity, path stability and information entropy preservation.

[0158] The single-modal feature method and traditional machine learning method achieved prediction accuracies of 72.8% and 75.6%, respectively, indicating that they can capture propagation patterns to some extent. However, due to the lack of cross-modal feature fusion and causal correction mechanisms, their overall prediction level is limited. In contrast, this invention constructs a propagation representation vector set through a self-supervised graph model and combines it with an energy distribution function and entropy constraint mechanism, thereby improving the prediction accuracy to 86.9%, which is about 11 percentage points higher than traditional methods. This improvement shows that this invention can better characterize the complex interaction relationship between users and videos.

[0159] In terms of path diversity and stability, this invention achieves scores of 0.63 and 0.71, respectively, while traditional methods fluctuate only between 0.41 and 0.49. This indicates that traditional methods tend to favor single-path prediction and lack an understanding of the diversity of propagation, which can easily lead to discrepancies between recommendations and actual propagation. In contrast, this invention introduces information entropy constraints based on energy modeling to ensure a more balanced path distribution, thereby effectively avoiding the problem of path homogenization. Furthermore, it enhances the logical consistency of propagation paths through a causal correction mechanism, enabling path prediction results to maintain high stability in different scenarios.

[0160] From the perspective of computational efficiency, the average computational latency of this invention is 8.4 seconds, which is lower than the 12.3 seconds and 10.7 seconds of traditional methods. This is due to the fact that this invention adopts structured energy distribution calculation and time continuous evolution function in propagation path modeling, avoiding redundant calculation steps, thereby improving prediction accuracy while taking into account computational efficiency, and meeting the real-time prediction needs of large-scale short video platforms.

[0161] Regarding the information entropy preservation value, the value of this invention reaches 2.74, while the traditional methods are 1.82 and 1.95 respectively. The higher the information entropy preservation value, the more diverse and information-rich the propagation path set is. This invention can effectively maintain the balance and information content of the propagation path distribution, ensuring that the platform can identify both highly disseminated videos and long-tail videos in the recommendation and content governance stages, thereby avoiding overly concentrated information flow distribution.

[0162] 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 method for predicting the propagation path of social short videos based on a self-supervised graph model, characterized in that, Includes the following steps: Collect user interaction data, video content data, and time series data from social short video platforms, and perform preprocessing to generate multimodal feature data; Based on multimodal feature data, a propagation graph structure containing user nodes and video nodes is constructed, social relationship edges and interaction behavior edges are defined, and embedded into the propagation graph structure to form a cross-modal propagation graph structure; Set a self-supervised task, construct a self-supervised loss, perform self-supervised training on the cross-modal propagation graph structure, update the propagation graph structure parameters based on the self-supervised loss, and generate a set of propagation representation vectors; Energy calculation is performed based on the propagation representation vector set, an energy distribution function is constructed, and energy minimization constraints are applied to the real propagation path, while energy maximization constraints are applied to the random path to obtain the propagation path energy distribution. Based on the energy distribution of the propagation path, an information entropy constraint is introduced to generate a set of propagation paths optimized by the entropy constraint. The set of propagation paths optimized by entropy constraints is subjected to causal correction based on recommendation mechanism causality, time decay causality, and interaction dependency causality to obtain a set of propagation paths with causal consistency. By performing time-continuous evolution modeling on the set of causally consistent propagation paths, the evolution trajectory of the propagation path energy dynamically changes over time, and finally obtaining the predicted propagation path of social short videos.

2. The method for predicting the propagation path of social short videos based on a self-supervised graph model according to claim 1, characterized in that, The generation of the multimodal feature data specifically includes: Collect user interaction data, video content data, and time series data, and perform data processing such as format unification, scale standardization, noise removal, and missing value imputation on the collected data; Encode user interaction data to generate user interaction feature data; The video content data is processed to extract image depth features, audio signal features, and text encoding features, and then fused to generate video content feature data. The time series data is normalized and standardized to generate time series feature data, which includes interaction timestamp data and propagation delay parameters. By fusing user interaction feature data, video content feature data, and time series feature data, multimodal feature data is generated.

3. The method for predicting the propagation path of social short videos based on a self-supervised graph model according to claim 1, characterized in that, The generation of the cross-modal propagation graph structure specifically includes: Based on multimodal feature data, a propagation graph structure is constructed, where each user in the user interaction feature data is represented as a user node, and each video in the video content feature data is represented as a video node, forming a node set composed of user nodes and video nodes. Extract social relationships between users from user interaction feature data, establish social relationships as social relationship edges between user nodes, and bind the social relationship edge feature data between users; Extract user interaction behavior from user interaction feature data, establish interaction behavior edges between user nodes and video nodes, and bind the corresponding interaction behavior edge feature data. Video content feature data is bound to video nodes to obtain video node feature data; Bind user interaction feature data to user nodes to obtain user node feature data; Bind time series feature data to the interaction behavior edges in the propagation graph structure; The user nodes, video nodes, social relationship edges, interaction behavior edges, and their corresponding node feature data and edge feature data are fused in the propagation graph structure to generate a cross-modal propagation graph structure.

4. The method for predicting the propagation path of social short videos based on a self-supervised graph model according to claim 1, characterized in that, The generation of the propagation representation vector set specifically includes: Set up an edge masking prediction task to mask the edge feature data of some social relationship edges and interaction behavior edges in the cross-modal propagation graph structure according to a preset ratio, and generate a masked edge set; Based on the unmasked node feature data and edge feature data in the cross-modal propagation graph structure, edge feature prediction is performed on each edge in the masked edge set to obtain edge prediction output data. The edge prediction output data is then compared with the real edge feature data to generate the edge mask prediction loss value. A node context comparison task is set up. Positive sample node pairs and negative sample node pairs are sampled for user nodes and video nodes in the cross-modal propagation graph structure, respectively. The cosine similarity between node pairs is calculated based on node feature data, and combined with label information, a node context comparison loss value is generated. The path reconstruction task is defined. A portion of the propagation path segments are randomly extracted from the cross-modal propagation graph structure to form a path mask set. The path mask set is reconstructed based on the existing propagation representation vector set to obtain the path reconstruction output data. The path reconstruction output data is then compared with the real path data to generate the path reconstruction loss value. The edge mask prediction loss, node context comparison loss, and path reconstruction loss are weighted and summed according to preset weights to generate a self-supervised loss. Based on self-supervised loss, parameter updates are performed on user node feature data, video node feature data, social relationship edge feature data, and interaction behavior edge feature data in the cross-modal propagation graph structure to generate an updated propagation graph structure parameter set. Based on the updated propagation graph structure parameter set, information aggregation calculation is performed on the cross-modal propagation graph structure to generate a propagation representation vector set.

5. The method for predicting the propagation path of social short videos based on a self-supervised graph model according to claim 1, characterized in that, The generation of the propagation path energy distribution specifically includes: Based on the propagation representation vector set, user node representation vectors, video node representation vectors, and corresponding social relationship edge feature data and interaction behavior edge feature data in the real propagation path are extracted to generate real path feature data. The real path feature data is processed to calculate the energy value of the real propagation path. The energy value is calculated by calculating the cosine similarity of the combined representation vector of each pair of adjacent nodes in the real path and the edge feature data of their connecting edges, and all cosine similarities are accumulated. The result is defined as the energy value of the real path. Negative sample propagation paths are generated by randomly sampling from the propagation representation vector set. User node representation vectors, video node representation vectors, and edge feature data are extracted from these paths to generate random path feature data. The random path feature data is processed to calculate the energy value of the random propagation path. The energy value is calculated by calculating the cosine similarity between the combined representation vector of each pair of adjacent nodes in the random path and the edge feature data of their connecting edges. All cosine similarities are accumulated, and the result is defined as the random path energy value. Apply an energy minimization constraint to the actual propagation path; Imposing an energy maximization constraint on the random propagation path; Under the combined influence of energy minimization and energy maximization constraints, an energy distribution function is constructed by calculating the difference between the energy values ​​of the actual path and the energy values ​​of the random path. The propagation representation vector set is optimized based on the energy distribution function to generate the propagation path energy distribution.

6. The method for predicting the propagation path of social short videos based on a self-supervised graph model according to claim 1, characterized in that, The generation of the entropy-constrained optimized propagation path set specifically includes: Based on the propagation path energy distribution, the path energy weighting value corresponding to each candidate propagation path is extracted to form a path energy weighting value set; Normalize the set of path energy weights to obtain the set of candidate propagation path probability values; Calculate the propagation path information entropy value based on the set of candidate propagation path probability values; The propagation path information entropy value is compared with a preset entropy threshold. When the propagation path information entropy value is less than the preset entropy threshold, the candidate propagation path probability value set is redistributed. The distribution of candidate propagation path probability values ​​is adjusted while keeping the relative magnitude of the path energy weighting values ​​unchanged, thus forming a corrected candidate propagation path probability value set. The propagation path information entropy value is recalculated based on the corrected candidate propagation path probability value set. If the corrected propagation path information entropy value is greater than or equal to the preset entropy threshold, the corrected candidate propagation path probability value set is confirmed as the final candidate propagation path probability value set. If the corrected propagation path information entropy value is still less than the preset entropy threshold, the reallocation process continues until the preset entropy threshold constraint is met. Output the set of propagation paths optimized by entropy constraints.

7. The method for predicting the propagation path of social short videos based on a self-supervised graph model according to claim 1, characterized in that, The generation of the set of propagation paths with causal consistency specifically includes: Based on the set of propagation paths optimized by entropy constraints, the probability value of each candidate propagation path is extracted to form a set of candidate propagation path probability values. Based on the set of probability values ​​of candidate propagation paths, causal constraints of the recommendation mechanism are constructed, and causal correction factors of the recommendation mechanism are generated according to the frequency of recommendation triggering behavior, thus forming a set of causal correction factors of the recommendation mechanism. Based on the probability value set of candidate propagation paths and time series data, a time decay causal constraint is constructed. This constraint is established by calculating the time interval between adjacent interactions in the candidate propagation path. When the time interval is greater than a preset time threshold, a time decay factor is assigned to the candidate propagation path. When the time interval is less than or equal to the preset time threshold, a time preservation factor is assigned to the candidate propagation path. All factors form a time factor set. Based on the probability value set of candidate propagation paths and user interaction feature data, interaction dependency causal constraints are constructed. These constraints are established by analyzing multiple interaction behaviors of user nodes in candidate propagation paths. When user nodes exhibit multiple interaction behaviors in candidate propagation paths, interaction dependency factors are assigned to the candidate propagation paths. When user nodes exhibit only a single interaction behavior in candidate propagation paths, interaction independence factors are assigned to the candidate propagation paths. All factors form an interaction factor set. For each candidate propagation path, the causal correction factor, time factor and interaction factor of the recommendation mechanism are applied to the corresponding candidate propagation path probability value to generate the corrected candidate propagation path probability value, forming a set of causal corrected candidate propagation path probability values. The propagation path information entropy value is recalculated based on the causal correction candidate propagation path probability value set. When the propagation path information entropy value is greater than or equal to the preset entropy threshold, the causal correction candidate propagation path probability value set is confirmed as the final candidate propagation path probability value set. When the propagation path information entropy value is less than the preset entropy threshold, the causal correction factor set, time factor set, and interaction factor set of the recommendation mechanism are redistributed until the propagation path information entropy value is greater than or equal to the preset entropy threshold. Output the set of propagation paths that maintain causal consistency.

8. The method for predicting the propagation path of social short videos based on a self-supervised graph model according to claim 1, characterized in that, The generation of the final predicted social short video dissemination path specifically includes: The set of propagation paths with causal consistency is sorted in chronological order, and the causal-corrected probability value of each candidate propagation path is bound to the corresponding timestamp to obtain a set of path time state vectors. The time evolution calculation is performed on the path time state vector set. The time state of each candidate propagation path is iteratively calculated based on the continuous time evolution function. The energy change value of the candidate propagation path at adjacent time scales is output, and the path energy time series is generated. In the process of calculating the path energy time series, the energy change value of each candidate propagation path on the continuous time scale is accumulated to form a set of dynamic cumulative path energy values. The dynamic cumulative path energy values ​​are compared with a preset time evolution threshold. When the value is greater than the time evolution threshold, a time enhancement weight is assigned to the propagation path. When the value is less than or equal to the time evolution threshold, a time preservation weight is assigned to the propagation path, thus obtaining a set of path time weights. The path time weight set is combined with the path energy time series one by one to generate the path weighted energy time series, and the energy dynamic trajectory of the candidate propagation path is recorded on a continuous time scale to form the energy evolution trajectory set of the candidate propagation path. Normalization is performed on the set of energy evolution trajectories to obtain the set of evolution trajectories in which the propagation path energy changes dynamically over time. The set of evolution trajectories is then used as the final predicted propagation path for social short videos.

Images