Method and system for accurate pushing of media information fused with multi-source data

Abstract

This invention provides a method and system for precise media information push that integrates multi-source data, relating to the field of data processing technology. The method includes collecting user multimodal interaction behavior data from heterogeneous data sources and performing cross-modal semantic decoupling; identifying the contribution of each modality by constructing an intermodal causal dependency graph, and thereby selecting and adjusting features to obtain decoupled feature representations; constructing a user hierarchical interest graph and characterizing the interest evolution path; performing multi-hop inference on the graph based on the content features of the information to be pushed, and dynamically recalibrating the association strength in conjunction with the context to generate a push decision; after executing the decision, driving the causal dependency graph to reorganize based on the temporal fluctuation characteristics of user interaction response data, and predictively reshaping the interest graph evolution path based on the behavior chain, thereby achieving precise and adaptive information push.

Description

Technical Field

[0001] This invention relates to data processing technology, and more particularly to a method and system for accurately pushing media information by integrating multi-source data. Background Technology

[0002] In the field of media information delivery, existing technologies typically rely on the analysis of users' historical behavioral data to build interest models. These methods generally collect multimodal behavioral data from multiple heterogeneous data sources, including user clicks, browsing, searches, and social interactions, and attempt to fuse these data through feature engineering or deep learning models to form a unified user profile. Conventional approaches focus on using collaborative filtering, content filtering, or a hybrid model of both to concatenate or weighted sum data from different modalities in order to capture user preferences. Furthermore, some advanced solutions introduce attention mechanisms or graph neural networks to model the interaction between users and items, thereby improving the accuracy of recommendations.

[0003] However, these conventional approaches have significant drawbacks. On the one hand, existing methods often overlook the complex causal dependencies and interactions between features of different modalities when fusing multimodal data. Simply merging feature vectors from text, images, and videos easily introduces redundancy and noise interference between modalities, leading to biases in the generated user interest representations and an inability to clearly distinguish the independent contributions and synergistic effects of each modality on interest formation. On the other hand, most push systems lack a detailed characterization of the dynamic evolution of user interests. They typically treat interests as static or perform only simple temporal smoothing, failing to effectively model the formation, evolution, and decay patterns of interest nodes at different time scales, and even more so, failing to capture the immediate impact of contextual changes on the strength of interest associations. This results in push results often lagging behind the user's actual interest migration, making it difficult to achieve accurate personalized matching. Summary of the Invention

[0004] The embodiments of the present invention provide a method and system for accurate push of media information by integrating multi-source data, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a method for accurately pushing media information by fusing multi-source data, comprising:

[0006] Collect multimodal interaction behavior data of users from heterogeneous data sources, and perform cross-modal semantic decoupling on the multimodal interaction behavior data;

[0007] By constructing a causal dependency graph between modalities, the independent contribution and interactive contribution of each modality to the formation of user interests are identified. Based on the complementary and competitive relationship between the independent contribution and the interactive contribution, the modal features are differentiated and their intensity is adjusted to obtain a decoupled feature representation that eliminates redundant interference between modalities.

[0008] Based on the decoupled feature representation, a hierarchical interest graph of the user is constructed, and the evolution path between interest nodes at different time scales is characterized by graph structure.

[0009] Based on the content characteristics of the media information to be pushed, multi-hop reasoning is performed on the hierarchical interest graph to calculate the correlation strength distribution between the media information to be pushed and interest nodes at each time scale. The correlation strength distribution is dynamically recalibrated in combination with the current context features to generate a push decision.

[0010] The system executes the push decision and observes the user interaction response data after the push. Based on the temporal fluctuation characteristics of different modal feedbacks in the user interaction response data, it identifies the response preference migration pattern. It uses the deviation between the temporal fluctuation characteristics and the contribution of each modality in the causal dependency graph to drive the topological restructuring of the causal dependency graph. At the same time, it predictively reshapes the evolution path in the hierarchical interest graph based on the interest node trigger sequence and decay cycle of the behavioral chain in the user interaction response data.

[0011] Based on the complementary and competitive relationship between the independent contribution and the interactive contribution, differentiated screening and intensity adjustment are performed on each modal feature to obtain a decoupled feature representation that eliminates redundant interference between modalities, including:

[0012] For each modal node in the causal dependency graph, the difference in the direction and intensity of the independent contribution and the interaction contribution in the process of user interest formation is quantified. When the two have opposite directions and the intensity difference exceeds the cooperation threshold, they are identified as competitive relationships. When the two have the same direction and there is a mutual reinforcement effect, they are identified as complementary relationships.

[0013] For modal feature pairs identified as competing relationships, the dominant contribution mode and the suppressed contribution mode are determined based on the intensity difference. The features of the dominant contribution mode are selectively retained by increasing their expression weight in the decoupled feature representation, and the features of the suppressed contribution mode are selectively suppressed by decreasing their expression weight.

[0014] For modal feature pairs identified as complementary, a synergistic gain coefficient between modalities is constructed based on the strength of the mutual enhancement effect. The synergistic gain coefficient is then used to amplify the expression weights of each modal feature participating in the complementarity, thereby strengthening the complementary relationship in the decoupled feature representation.

[0015] The modal features, after selective retention, selective suppression, and linked amplification, are combined according to the adjusted expression weights to form a decoupled feature representation.

[0016] Based on the decoupled feature representation, a hierarchical interest graph of the user is constructed, and the evolution path between interest nodes at different time scales is characterized by a graph structure, including:

[0017] The time dependence of each modal feature in the decoupled feature representation is measured. The time-scale representation vector of the feature is constructed by analyzing the activation frequency distribution and persistence distribution of each feature in the user interaction sequence. The features are aggregated into interest node levels with different time dependencies according to the topological distribution of the time-scale representation vector in the feature space.

[0018] For each node in the interest node hierarchy, the semantic association tensor and temporal transition probability matrix between nodes are calculated. The semantic association tensor and the temporal transition probability matrix are fused by tensor operation to obtain a joint representation matrix that characterizes the evolutionary association between nodes. Node pairs with association strength exceeding the evolution threshold are extracted from the joint representation matrix as candidate node pairs with evolutionary relationship.

[0019] For the candidate node pair, the starting node and ending node of the evolution path are determined based on the difference in time dependency between the nodes and the direction of semantic drift. When the time dependency of the starting node is less than that of the ending node, a fixed evolution path is established. When the time dependency of the starting node is greater than that of the ending node, an induced evolution path is established.

[0020] Based on the content characteristics of the media information to be pushed, multi-hop inference is performed on the hierarchical interest graph to calculate the association strength distribution between the media information to be pushed and interest nodes at each time scale. The association strength distribution is then dynamically recalibrated in conjunction with the current context features to generate a push decision, including:

[0021] Extract the content features of the media information to be pushed and construct a multi-scale representation of the content features. Then, perform semantic matching between the multi-scale representation and the feature representations of interest nodes at each level in the hierarchical interest graph to obtain the initial matching response distribution.

[0022] Using the initial matching response distribution as the starting activation state for multi-hop inference, information is propagated along the evolution path on the hierarchical interest graph. Each hop attenuates and transfers the activation intensity of the current node through path weights, and modulates and amplifies the propagation intensity according to the semantic association between nodes. After multi-hop propagation, the cumulative activation intensity of interest nodes at each time scale is used as the association intensity distribution.

[0023] Obtain the current context features and construct a spatiotemporal representation matrix of the context features. Perform tensor shrinkage operation on the spatiotemporal representation matrix and the association strength distribution to obtain context-sensitive association strength recalibration results.

[0024] Based on the adjusted association strength of interest nodes at each time scale in the association strength recalibration results, a multi-scale fusion vector is constructed. The multi-scale fusion vector is weighted and aggregated to obtain a comprehensive push score. When the comprehensive push score meets the push conditions, a push decision is generated for the media information to be pushed.

[0025] Obtain the current context features and construct a spatiotemporal representation matrix of the context features. Perform tensor shrinkage operation on the spatiotemporal representation matrix and the association strength distribution to obtain context-sensitive association strength recalibration results, including:

[0026] The current context features are obtained, and the current context features are subjected to spatiotemporal separation transformation to obtain temporal context components and spatial context components. The frequency domain characteristics of the temporal context components are extracted to construct a temporal spectrum representation, and the topological characteristics of the spatial context components are extracted to construct a spatial topological representation. The temporal spectrum representation and the spatial topological representation are subjected to tensor outer product operation to construct the spatiotemporal representation matrix of the context features.

[0027] The association strength values ​​of interest nodes at each time scale in the association strength distribution are extracted and an association strength tensor is constructed. The association strength tensor is then dimensionally expanded so that its dimensional structure forms a contractible correspondence with the dimensional structure of the spatiotemporal representation matrix.

[0028] The spatiotemporal representation matrix and the correlation strength tensor are subjected to tensor shrinkage operation. In the tensor shrinkage operation, the temporal scale dimension of the correlation strength tensor is frequency-matched modulated by the frequency domain component of the temporal spectrum representation, and the node index dimension of the correlation strength tensor is structure-matched modulated by the topological component of the spatial topology representation, so as to obtain the shrinkage result tensor of the fused spatiotemporal characteristics.

[0029] Tensor decomposition is performed on the tensor of the contraction result to separate the modulation coefficients of interest nodes at each time scale in the current context. The modulation coefficients are then multiplied element-wise with the association strength distribution to obtain the context-sensitive association strength recalibration result.

[0030] The topological restructuring of the causal dependency graph is driven by the deviation between the temporal fluctuation characteristics and the modal contributions in the causal dependency graph. Simultaneously, the evolutionary path in the hierarchical interest graph is predictively reshaped based on the interest node trigger sequence and decay cycle of the behavioral chain in the user interaction response data, including:

[0031] Construct a mutual information flow tensor of the temporal fluctuation features and the contribution of each mode in the causal dependency graph. Perform manifold embedding on the mutual information flow tensor to obtain a causal deviation manifold. Identify the modal node clusters whose geodesic distance exceeds the stability threshold on the causal deviation manifold. Reconstruct the propagation path of the causal dependency edges along the geodesic direction for the modal node clusters to complete the topological restructuring of the causal dependency graph.

[0032] Behavioral chains are extracted from the user interaction response data and their projection trajectories on the hierarchical interest map are constructed. Lyapunov exponential analysis is performed on the projection trajectories to obtain the chaotic characteristic quantification index of the interest node trigger sequence. A nonlinear decay cycle model is fitted based on the chaotic characteristic quantification index.

[0033] The nonlinear decay cycle model is aligned with the causal deviation manifold, and the conditional activation probability density function is calculated in the aligned joint manifold space. The evolution path in the hierarchical interest map is predictively reshaped based on the weight evolution trajectory predicted by the conditional activation probability density function.

[0034] Construct a mutual information flow tensor of the temporal fluctuation features and the modal contributions in the causal dependency graph, perform manifold embedding on the mutual information flow tensor to obtain a causal deviation manifold, and identify modal node clusters on the causal deviation manifold whose geodesic distances exceed the stability threshold, including:

[0035] The intrinsic mode function set is obtained by performing Hilbert-Huang transform on the time series fluctuation characteristics. For each mode contribution in the causal dependency graph, an attractor trajectory in the phase space is constructed. The transfer entropy flow between the intrinsic mode function set and the attractor trajectory is calculated. The transfer entropy flow is tensorized in the modal dimension, time delay dimension and embedding dimension to construct a mutual information flow tensor. Redundant degrees of freedom are eliminated by performing symplectic geometric reduction on the mutual information flow tensor.

[0036] The reduced mutual information flow tensor is embedded into a Riemannian manifold with curvature constraints. During the embedding process, hyperbolic and elliptical regions of the causal deviation manifold are identified by the sign distribution of curvature. The variational geodesic distance between modal nodes is calculated on the causal deviation manifold using the geodesic equation of hybrid geometry.

[0037] A persistent homology group of modal nodes is constructed on the causal deviation manifold. A candidate set of topologically stable modal node clusters is identified by the Betti number sequence of the persistent homology group. For each cluster in the candidate set of modal node clusters, the entropy rate of its internal variational geodesic distance is calculated. The entropy rate is compared with a stability threshold to identify modal node clusters whose geodesic distance exceeds the stability threshold.

[0038] A second aspect of the present invention provides a media information precision push system that integrates multi-source data, comprising:

[0039] The data collection unit is used to collect multimodal interaction behavior data of users from heterogeneous data sources and to perform cross-modal semantic decoupling on the multimodal interaction behavior data;

[0040] The feature decoupling unit is used to identify the independent contribution and interactive contribution of each modality to the formation of user interests by constructing a causal dependency graph between modalities, and to perform differential screening and intensity adjustment of each modal feature based on the complementary and competitive relationship between the independent contribution and the interactive contribution, so as to obtain a decoupled feature representation that eliminates redundant interference between modalities;

[0041] The graph construction unit is used to construct a hierarchical interest graph of the user based on the decoupled feature representation, and to characterize the evolution path between interest nodes at different time scales through graph structure;

[0042] The push decision unit is used to perform multi-hop reasoning on the hierarchical interest graph based on the content characteristics of the media information to be pushed, calculate the correlation strength distribution between the media information to be pushed and interest nodes at each time scale, dynamically recalibrate the correlation strength distribution in combination with the current context features, and generate a push decision.

[0043] The graph optimization unit is used to execute the push decision and observe the user interaction response data after the push. It identifies the response preference migration pattern based on the temporal fluctuation characteristics of different modal feedbacks in the user interaction response data, drives the topological restructuring of the causal dependency graph by using the deviation of the temporal fluctuation characteristics from the contribution of each modality in the causal dependency graph, and predictively reshapes the evolution path in the hierarchical interest graph based on the interest node trigger sequence and decay cycle of the behavioral chain in the user interaction response data.

[0044] A third aspect of the present invention provides an electronic device, comprising:

[0045] processor;

[0046] Memory used to store processor-executable instructions;

[0047] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0048] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0049] The beneficial effects of this application are as follows:

[0050] This method effectively collects multimodal user interaction data from heterogeneous data sources and performs cross-modal semantic decoupling. This process removes semantic information mixed in different modal data, laying the foundation for subsequent accurate analysis. By constructing a causal dependency graph between modalities, the independent and interactive contributions of each modality to the formation of user interests can be quantitatively identified. Based on the complementary and competitive relationships between contributions, differentiated screening and intensity adjustment of modal features can be implemented, effectively eliminating redundancy and interference between modalities, ultimately obtaining a pure, decoupled feature representation that more realistically reflects the user's core intent.

[0051] A hierarchical interest graph of users is constructed based on decoupled feature representation, and the evolutionary path between interest nodes at different time scales is characterized by a graph structure. This graph representation can intuitively and structurally show the hierarchical relationships and dynamic changes of user interests. Multi-hop inference is performed on the hierarchical interest graph based on the content features of the media information to be pushed, and the distribution of the association strength between the information and interest nodes at each time scale can be calculated. Dynamic recalibration of this distribution, combined with current contextual features, can generate push decisions that better fit the user's immediate state and scenario, significantly improving the real-time relevance and contextual adaptability of push notifications. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the method for precise media information delivery by integrating multi-source data according to an embodiment of the present invention.

[0053] Figure 2 This is a flowchart of the mutual information flow tensor manifold embedding and weight evolution prediction in an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0056] Figure 1 This is a flowchart illustrating the method for precise media information delivery by integrating multi-source data according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0057] Collect multimodal interaction behavior data of users from heterogeneous data sources, and perform cross-modal semantic decoupling on the multimodal interaction behavior data;

[0058] By constructing a causal dependency graph between modalities, the independent contribution and interactive contribution of each modality to the formation of user interests are identified. Based on the complementary and competitive relationship between the independent contribution and the interactive contribution, the modal features are differentiated and their intensity is adjusted to obtain a decoupled feature representation that eliminates redundant interference between modalities.

[0059] Based on the decoupled feature representation, a hierarchical interest graph of the user is constructed, and the evolution path between interest nodes at different time scales is characterized by graph structure.

[0060] Based on the content characteristics of the media information to be pushed, multi-hop reasoning is performed on the hierarchical interest graph to calculate the correlation strength distribution between the media information to be pushed and interest nodes at each time scale. The correlation strength distribution is dynamically recalibrated in combination with the current context features to generate a push decision.

[0061] The system executes the push decision and observes the user interaction response data after the push. Based on the temporal fluctuation characteristics of different modal feedbacks in the user interaction response data, it identifies the response preference migration pattern. It uses the deviation between the temporal fluctuation characteristics and the contribution of each modality in the causal dependency graph to drive the topological restructuring of the causal dependency graph. At the same time, it predictively reshapes the evolution path in the hierarchical interest graph based on the interest node trigger sequence and decay cycle of the behavioral chain in the user interaction response data.

[0062] In one optional implementation, the modal features are differentiated and their intensity adjusted based on the complementary and competitive relationship between the independent contribution and the interactive contribution, resulting in a decoupled feature representation that eliminates redundant interference between modalities, including:

[0063] For each modal node in the causal dependency graph, the difference in the direction and intensity of the independent contribution and the interaction contribution in the process of user interest formation is quantified. When the two have opposite directions and the intensity difference exceeds the cooperation threshold, they are identified as competitive relationships. When the two have the same direction and there is a mutual reinforcement effect, they are identified as complementary relationships.

[0064] For modal feature pairs identified as competing relationships, the dominant contribution mode and the suppressed contribution mode are determined based on the intensity difference. The features of the dominant contribution mode are selectively retained by increasing their expression weight in the decoupled feature representation, and the features of the suppressed contribution mode are selectively suppressed by decreasing their expression weight.

[0065] For modal feature pairs identified as complementary, a synergistic gain coefficient between modalities is constructed based on the strength of the mutual enhancement effect. The synergistic gain coefficient is then used to amplify the expression weights of each modal feature participating in the complementarity, thereby strengthening the complementary relationship in the decoupled feature representation.

[0066] The modal features, after selective retention, selective suppression, and linked amplification, are combined according to the adjusted expression weights to form a decoupled feature representation.

[0067] For each modal node in the causal dependency graph, the system first extracts numerical representations of independent contribution and interaction contribution. Independent contribution is obtained by calculating the marginal effect of a single modal feature on the target user's interest prediction result. Specifically, the feature vector corresponding to a modal node in the causal dependency graph is used as the sole input, while the features of other modalities are either zeroed out or filled with the mean. The processed features are then input into the interest prediction model, and the difference between the predicted output and the baseline output is recorded as the independent contribution value of that modality. Interaction contribution is obtained by calculating the difference between the joint marginal effect and the sum of the individual independent marginal effects of a modal pair. The process involves retaining the feature vectors of two modalities, setting the features of other modalities to zero, inputting the joint features into the prediction model, recording the difference between the predicted output and the baseline output, and subtracting the independent contribution values ​​of the two modalities from the difference. The resulting difference is the interaction contribution. The baseline output is defined as the predicted output value when all modal features are filled with the mean of the training set.

[0068] To quantify the direction and intensity difference of the roles of independent and interactive contributions in the formation of user interest, both need to be vectorized. The direction of influence is determined by a sign function of the contribution values: positive values ​​indicate promotion of user interest formation, while negative values ​​indicate inhibition. The intensity difference is defined as the difference between the absolute values ​​of independent and interactive contributions, divided by the sum of their absolute values, and then standardized to limit the value of the intensity difference to between 0 and 1. When the signs of independent and interactive contributions are opposite, they are considered to have opposite directions of influence. If the standardized intensity difference exceeds a collaboration threshold, a competitive relationship is identified. The recommended range for the collaboration threshold is 0.2 to 0.5, with a default value of 0.3. This threshold is determined through cluster analysis of modal relationships in historical data, ensuring that competitive and complementary relationship samples reach maximum inter-class distance at this threshold. When the signs of independent and interactive contributions are the same and the absolute value of the interactive contribution is greater than 20% of the absolute value of the independent contribution, a mutually reinforcing effect is considered, and the relationship is identified as complementary. The intensity of the mutual reinforcement effect is quantified as the ratio of the absolute value of the interaction contribution to the absolute value of the independent contribution. This ratio reflects the amplification of the intermodal synergy relative to the single-modal effect.

[0069] For modal feature pairs identified as having a competitive relationship, the system determines the dominant and suppressed modes based on their intensity differences. The dominant mode is defined as the mode with the larger absolute value of its independent contribution, while the suppressed mode is defined as the mode with the smaller absolute value of its independent contribution. Selective retention is achieved by increasing the representation weight of the dominant mode in the decoupled feature representation. Specifically, the original feature vector of this mode is multiplied element-wise by a weight factor, which is set to 1 plus the square of the standardized intensity difference. This calculation ensures that the larger the intensity difference, the more significant the weight increase. Selective suppression is achieved by reducing the representation weight of the suppressed mode. The weight factor is set to 1 minus the square of the standardized intensity difference. When the weight factor is less than 0.1, it is truncated to 0.1 to avoid completely eliminating the modality information, as the suppressed mode still retains some useful information even in a competitive relationship. Boundary case handling rules are as follows: when the difference in the absolute value of the independent contributions of the two modes is less than 0.05, no weight adjustment is performed, and the original weight is kept at 1, avoiding the introduction of unnecessary adjustment noise when the contributions are close.

[0070] For modal feature pairs identified as complementary, the system constructs a synergistic gain coefficient between modes based on the strength of the mutual reinforcement effect. The synergistic gain coefficient is calculated as: 1 plus the natural logarithm of the mutual reinforcement effect strength, multiplied by an adjustment parameter. The default value of the adjustment parameter is set to 0.5, with a range of 0.3 to 0.8. This parameter is optimized on the validation set through grid search, ensuring that the joint representation of complementary modes is optimal in the decoupled feature representation. The synergistic gain coefficient is used to amplify the expression weights of the complementary modal features. Specifically, the original feature vectors of the two modes are multiplied by the synergistic gain coefficient, respectively, thus strengthening the complementary relationship in the decoupled feature representation. This amplification mechanism ensures that the synergistic effect of complementary modes in the feature space is fully reflected. When a mode participates in multiple complementary relationships simultaneously, the geometric mean of all relevant synergistic gain coefficients is taken as the final gain coefficient of that mode. The use of the geometric mean avoids the problem of extreme values ​​dominating due to the arithmetic mean, ensuring a balanced expression of multiple complementary relationships.

[0071] The modal features, after selective retention, selective suppression, and linked amplification, are combined according to adjusted expression weights to form a decoupled feature representation. The combination operation employs a weighted concatenation method, concatenating the adjusted feature vectors of all modalities in modality number order into a high-dimensional vector. The dimension of this high-dimensional vector is the sum of the dimensions of the original features of each modality. To ensure the numerical stability of the decoupled feature representation, the concatenated high-dimensional vector is standardized by calculating its mean and standard deviation. The mean is subtracted from the vector, and then the result is divided by the standard deviation, with a numerical stabilization term of 0.00001 added. This stabilization term avoids division-by-zero errors when the standard deviation is close to 0.

[0072] In one optional implementation, constructing a hierarchical interest graph of users based on the decoupled feature representation, and characterizing the evolution path between interest nodes at different time scales through graph structure includes:

[0073] The time dependence of each modal feature in the decoupled feature representation is measured. The time-scale representation vector of the feature is constructed by analyzing the activation frequency distribution and persistence distribution of each feature in the user interaction sequence. The features are aggregated into interest node levels with different time dependencies according to the topological distribution of the time-scale representation vector in the feature space.

[0074] For each node in the interest node hierarchy, the semantic association tensor and temporal transition probability matrix between nodes are calculated. The semantic association tensor and the temporal transition probability matrix are fused by tensor operation to obtain a joint representation matrix that characterizes the evolutionary association between nodes. Node pairs with association strength exceeding the evolution threshold are extracted from the joint representation matrix as candidate node pairs with evolutionary relationship.

[0075] For the candidate node pair, the starting node and ending node of the evolution path are determined based on the difference in time dependency between the nodes and the direction of semantic drift. When the time dependency of the starting node is less than that of the ending node, a fixed evolution path is established. When the time dependency of the starting node is greater than that of the ending node, an induced evolution path is established.

[0076] After obtaining the decoupled feature representation, it needs to be transformed into an interest map with a time-hierarchical structure. For each modality feature in the decoupled feature representation, its activation frequency in the user interaction sequence is calculated. The activation frequency is determined by the number of times the corresponding behavior type of the statistical feature occurs within a sliding time window, with the time window length set to three scales: 7 days, 30 days, and 90 days. The persistence distribution is obtained by calculating the mean and variance of the time interval between feature activation events. Features with a mean interval of less than 3 days and a variance below a threshold are marked as high persistence features. Activation frequency and persistence are used as two dimensions to construct a time-scale representation vector, and the magnitude of this vector reflects the temporal stability of the feature.

[0077] Density clustering analysis is performed on the time-scale representation vectors to divide the feature space into short-term, medium-term, and long-term interest regions. Short-term interest regions contain features with high activation frequencies but durations less than 2 days; medium-term interest regions correspond to features with durations between 2 and 14 days; and long-term interest regions encompass stable features with durations exceeding 14 days. Features located within the same region are aggregated into an interest node. The semantic representation of the node is calculated from the weighted centroid of all feature vectors within the region, with the weights depending on the ratio of the feature's activation frequency to the region's average activation frequency.

[0078] For the constructed hierarchy of interest nodes, a semantic association tensor between nodes is calculated. This tensor is a third-order tensor, with the first and second dimensions corresponding to node pairs, and the third dimension representing the association type, including three types: topic similarity, entity co-occurrence, and attribute complementarity. Topic similarity is measured by the cosine distance of the node semantic vectors; entity co-occurrence is calculated by statistically analyzing the overlap rate of named entities in the corresponding features of the nodes; and attribute complementarity is obtained by analyzing the frequency of collaborative triggering of node features in user behavior decisions. The temporal transition probability matrix records the conditional probability of a user moving from one interest node to another. This probability is obtained by statistically analyzing the frequency of node activation order in the user interaction sequence and then normalizing it.

[0079] Element-wise multiplication and tensor shrinkage operations are performed on the semantic association tensor and the temporal transition probability matrix to obtain a joint representation matrix. Each element of this matrix quantifies the combined strength of semantic association and temporal transition tendency between node pairs. An evolution threshold is set to 1.5 times the mean of all off-diagonal elements in the matrix, and node pairs with values ​​exceeding this threshold are extracted as candidate node pairs.

[0080] For each extracted candidate node pair, the temporal dependency difference between the nodes is calculated, defined as the difference in magnitude of the temporal scale representation vectors of the two nodes. The semantic drift direction is calculated by comparing the projection positions of the node semantic vectors in the topic space. A convergent drift is characterized by a trend of refinement in the topic distribution of the endpoint node relative to the starting node, while a divergent drift is characterized by a trend of diffusion. When the temporal dependency value of the starting node is less than that of the endpoint node, it indicates that user interest is shifting from a short-term to a long-term state, establishing a solidified evolutionary path that reflects the sedimentation process of interest. When the temporal dependency value of the starting node is greater than that of the endpoint node, it indicates that long-term interest has triggered new short-term exploratory behavior, establishing an stimulating evolutionary path that characterizes the expansion mechanism of interest. The evolutionary path is represented in the graph using directed edges, with edge weights set to the values ​​of the corresponding elements in the joint representation matrix.

[0081] In one optional implementation, multi-hop inference is performed on the hierarchical interest graph based on the content characteristics of the media information to be pushed, calculating the association strength distribution between the media information to be pushed and interest nodes at each time scale, and dynamically recalibrating the association strength distribution in combination with current context features to generate a push decision, including:

[0082] Extract the content features of the media information to be pushed and construct a multi-scale representation of the content features. Then, perform semantic matching between the multi-scale representation and the feature representations of interest nodes at each level in the hierarchical interest graph to obtain the initial matching response distribution.

[0083] Using the initial matching response distribution as the starting activation state for multi-hop inference, information is propagated along the evolution path on the hierarchical interest graph. Each hop attenuates and transfers the activation intensity of the current node through path weights, and modulates and amplifies the propagation intensity according to the semantic association between nodes. After multi-hop propagation, the cumulative activation intensity of interest nodes at each time scale is used as the association intensity distribution.

[0084] Obtain the current context features and construct a spatiotemporal representation matrix of the context features. Perform tensor shrinkage operation on the spatiotemporal representation matrix and the association strength distribution to obtain context-sensitive association strength recalibration results.

[0085] Based on the adjusted association strength of interest nodes at each time scale in the association strength recalibration results, a multi-scale fusion vector is constructed. The multi-scale fusion vector is weighted and aggregated to obtain a comprehensive push score. When the comprehensive push score meets the push conditions, a push decision is generated for the media information to be pushed.

[0086] The content feature extraction of the media information to be pushed adopts a multimodal feature fusion architecture, decomposing the media information into modal components such as text, image, video, and audio. For the text modality, a word embedding layer is used to convert the text sequence into a 512-dimensional dense vector. The word embedding layer is pre-trained on a corpus of 1 billion words and supports a vocabulary of 32,000 words. For the image modality, a convolutional neural network is used to extract 2048-dimensional visual features. The network depth is 50 layers, and the input image size is normalized to 224 x 224 pixels. For the video modality, 3 keyframes are sampled per second in the temporal dimension. After extracting visual features from each frame, they are aggregated into a 1024-dimensional temporal feature vector through a temporal pooling layer. The pooling operation uses a weighted combination of max pooling and average pooling, with a weight ratio of 0.6 to 0.4. For the audio modality, Mel spectral features are extracted and encoded into 256-dimensional acoustic features through a recurrent neural network. The number of Mel filter banks is set to 128.

[0087] Multi-scale representation construction is achieved by performing multi-level projection transformations on the feature vectors of each modality. The number of projection levels is set to three, corresponding to the semantic granularity of short-term, medium-term, and long-term time scales, respectively. Short-term projection maps the original feature vectors to a 128-dimensional space through a linear transformation. Medium-term projection maps to a 256-dimensional space after applying a modified linear unit activation function to the short-term representation, with the momentum parameter for batch normalization set to 0.9. Long-term projection maps to a 512-dimensional space through residual connections and dimensionality reduction transformations based on the medium-term representation, with the weight coefficient of the residual connections set to 0.3. The outputs of each level of the multi-scale representation are concatenated to form an 896-dimensional comprehensive feature vector, which serves as the query representation for subsequent semantic matching.

[0088] In the hierarchical interest graph, each level of interest node maintains a 512-dimensional feature vector as a semantic anchor. The semantic matching operation calculates the cosine similarity between the query representation and the feature representations of each interest node, with the value range limited to -1 to +1. An approximate nearest neighbor index is constructed for the hierarchical interest graph. The index structure adopts a hierarchical navigable small-world graph. The index construction parameters include 12 levels, a maximum of 32 connections per level, and a search depth of 200 during the construction phase. During a query, the top 50 interest nodes with the highest similarity and their similarity values ​​are returned. The similarity values ​​are normalized using a softmax function to obtain the initial matching response distribution, with the temperature parameter set to 0.1. Nodes with response values ​​below 0.001 in the initial matching response distribution are filtered out, and the response values ​​of the remaining nodes are re-normalized and used as the initial activation state for multi-hop inference.

[0089] Multi-hop inference propagates information along the evolutionary path of the hierarchical interest graph, defined as a directed edge connecting interest nodes at different time scales. The propagation process is set to a 3-hop iteration. In each hop, the activation intensity of the current node is distributed to neighboring nodes according to the path weight, with a path weight decay coefficient set to 0.85. The cumulative activation increment received by neighboring nodes is summed with their own current activation intensity to update the new activation state. Before the summation operation, a semantic association modulation coefficient is applied to the activation increment. When the similarity is greater than 0.7, the modulation coefficient is set to 1.5 to achieve an amplification effect; when the similarity is between 0.4 and 0.7, the modulation coefficient is set to 1.0 to maintain the original value; when the similarity is less than 0.4, the modulation coefficient is set to 0.5 to achieve a suppression effect. The first hop propagation starts from the initial activation state and transmits the activation intensity to intermediate interest nodes, with the propagation range limited to a maximum of 5 neighboring nodes per node. The second hop propagation starts from the updated activation state of the intermediate interest nodes and propagates to long-term interest nodes, with the path weight decay coefficient for lateral propagation adjusted to 0.7. The third hop propagation integrates the activation states of all nodes at all levels for global propagation. During the propagation process, an attention mechanism is introduced to adaptively weight the propagation intensity at different levels.

[0090] After multi-hop propagation, the cumulative activation intensity of interest nodes at each time scale is used as the initial value of the association intensity distribution, which is represented as a dictionary structure containing the identifiers of all interest nodes and their corresponding activation intensities. Nodes with activation intensities below 0.005 in the association intensity distribution are marked as weakly associated nodes. A hyperbolic tangent function constraint is applied to the cumulative activation intensities to map the activation intensities to the range of -1 to +1. Before mapping, the activation intensities are multiplied by a scaling factor of 2.

[0091] The acquisition of current contextual features encompasses multi-source information across time and space dimensions. Time-dimensional features include the current hour, day of the week, month, and holiday identifiers. The hour identifier ranges from 0 to 23 and is encoded as a 24-dimensional one-hot vector. Spatial-dimensional features include the user's current geographical location (latitude and longitude coordinates), region type identifier, ambient noise level, and network connection type. The latitude and longitude coordinates are converted into a 4-dimensional continuous vector using sine and cosine coding. The region type identifier covers five categories: home, office, transportation, entertainment, and outdoor, encoded as a 5-dimensional one-hot vector. The spatiotemporal representation matrix is ​​constructed by performing an outer product operation on the time-dimensional and spatial-dimensional features. The outer product result is a matrix with 45 rows equal to the total dimension of the time-dimensional features and 15 columns equal to the total dimension of the spatial-dimensional features.

[0092] The tensor shrinkage operation of the spatiotemporal representation matrix and the association strength distribution achieves context-sensitive recalibration through pattern multiplication. After the association strength distribution is converted into a column vector form, it is expanded into a three-dimensional tensor, resulting in a tensor with a shape of 45 x 15 x the number of nodes. Tensor shrinkage is performed by weighted summation along the time and spatial dimensions. The time dimension weight is calculated based on the matching degree between the current moment and the user's historical active periods: periods with a matching degree higher than 0.6 are weighted at 1.2, and those with a matching degree lower than 0.3 are weighted at 0.6. The spatial dimension weight is calculated based on the distance between the current location and the user's usual location: distances less than 1 kilometer are weighted at 1.3, and distances greater than 5 kilometers are weighted at 0.7. After the shrinkage operation, the adjusted association strength column vector is obtained.

[0093] In the recalibrated association strength results, the adjusted association strengths of interest nodes at each time scale are aggregated in groups according to the time scale. The association strength of short-term interest nodes is represented by the maximum value within the group, indicating the immediate interest peak; the association strength of medium-term interest nodes is represented by the average value within the group, indicating the sustained interest level; and the association strength of long-term interest nodes is represented by the weighted average value within the group, indicating the stable interest baseline. The weighting coefficient is positively correlated with the historical interaction frequency of the node. The multi-scale fusion vector is constructed by concatenating the short-term peak, medium-term level, and long-term baseline, and the vector dimension is 3-dimensional. The weighted aggregation of the fusion vector uses learnable weight parameters, which are obtained by training with historical push effect data. Click-through rate accounts for 0.7% and dwell time accounts for 0.3%. The weighted aggregation calculation is to multiply each dimension of the fusion vector with its corresponding weight parameter element-wise and then sum them to obtain a scalar form of comprehensive push score. The score value is normalized to the interval of 0 to 1 using the sigmoid function.

[0094] The determination of whether a push notification meets the conditions based on the overall push score is based on a dynamic threshold strategy. The dynamic threshold is calculated based on the push capacity and user fatigue level of the current time period. The push capacity is set to 8 push notifications per hour during periods of high user activity, and reduced to 2 push notifications per hour during periods of low user activity. User fatigue level is calculated based on the push response rate over the past 24 hours. When the response rate is higher than 50%, the fatigue level is set to 0.2, and when the response rate is lower than 20%, the fatigue level is set to 0.8. The dynamic threshold is calculated by multiplying the baseline threshold of 0.6 by 1 and adding the fatigue level. A push decision is generated when the overall push score exceeds the dynamic threshold and the remaining push capacity for the current time period is greater than 0. The push decision includes the identifier of the media information to be pushed, the push time, the push channel, and the push priority fields. The push priority is divided into three levels: high, medium, and low, based on the difference between the overall push score and the dynamic threshold.

[0095] In one optional implementation, the current context features are obtained and a spatiotemporal representation matrix of the context features is constructed. Tensor contraction is then performed on the spatiotemporal representation matrix and the association strength distribution to obtain a context-sensitive association strength recalibration result, including:

[0096] The current context features are obtained, and the current context features are subjected to spatiotemporal separation transformation to obtain temporal context components and spatial context components. The frequency domain characteristics of the temporal context components are extracted to construct a temporal spectrum representation, and the topological characteristics of the spatial context components are extracted to construct a spatial topological representation. The temporal spectrum representation and the spatial topological representation are subjected to tensor outer product operation to construct the spatiotemporal representation matrix of the context features.

[0097] The association strength values ​​of interest nodes at each time scale in the association strength distribution are extracted and an association strength tensor is constructed. The association strength tensor is then dimensionally expanded so that its dimensional structure forms a contractible correspondence with the dimensional structure of the spatiotemporal representation matrix.

[0098] The spatiotemporal representation matrix and the correlation strength tensor are subjected to tensor shrinkage operation. In the tensor shrinkage operation, the temporal scale dimension of the correlation strength tensor is frequency-matched modulated by the frequency domain component of the temporal spectrum representation, and the node index dimension of the correlation strength tensor is structure-matched modulated by the topological component of the spatial topology representation, so as to obtain the shrinkage result tensor of the fused spatiotemporal characteristics.

[0099] Tensor decomposition is performed on the tensor of the contraction result to separate the modulation coefficients of interest nodes at each time scale in the current context. The modulation coefficients are then multiplied element-wise with the association strength distribution to obtain the context-sensitive association strength recalibration result.

[0100] When acquiring current contextual features, raw contextual data such as the user's current time marker, geographic location coordinates, ambient light intensity, and network connection status are collected from device sensors. Spatiotemporal separation transformation is performed on this raw contextual data, mapping the time marker to periodic coordinates on a continuous time axis. Periodic phases at three time scales—hourly, daily, and weekly—are extracted and combined to form a temporal contextual component. Simultaneously, the geographic location coordinates are converted into a regional identifier using a spatial encoding algorithm, and combined with ambient light intensity and network connection status to construct a spatial contextual component.

[0101] For the temporal context component, a Fast Fourier Transform (FFT) is used to perform frequency domain analysis on each periodic phase to obtain the dominant frequency components and frequency energy distribution representing the user's daily routine. The frequency components and their corresponding energy values ​​are organized into a two-dimensional frequency-energy spectrum, which is the temporal spectrum representation. For the spatial context component, a pre-constructed scene topology network is queried based on the region identifier code. In this network, nodes represent specific scene types, and edge weights represent the transition frequency between scenes. The current region node and its first-order neighbor nodes are extracted to form a subgraph structure. The adjacency matrix of the subgraph is merged with the environmental feature vector to form a spatial topology representation.

[0102] The temporal spectral representation is constructed as an M×N matrix, where M is the number of frequency components and N is the number of time scales. The spatial topological representation is constructed as a P×Q matrix, where P is the number of scene nodes and Q is the topological feature dimension. Tensor outer product operations are performed on these two matrices to generate a four-dimensional spatiotemporal representation matrix with a dimensional structure of M×N×P×Q. This matrix simultaneously encodes the temporal frequency characteristics and spatial topological characteristics of the context.

[0103] The association strength values ​​of interest nodes at each time scale in the association strength distribution are extracted. These values ​​record the matching degree between the media information to be pushed and short-term, medium-term, and long-term interest nodes. These association strength values ​​are organized into a three-dimensional association strength tensor according to time scale type and node index, with initial dimensions of time scale number K, number of nodes L, and feature dimension D. A dimensionality expansion operation is performed on this tensor, inserting new dimensions into the first and third dimensions respectively, making the expanded tensor dimension M×K×L×P, thus establishing a correspondence with the dimensional structure of the spatiotemporal representation matrix.

[0104] During tensor shrinkage, the shrinkage operation is performed along the first and third dimensions of the spatiotemporal representation matrix and the corresponding dimensions of the association strength tensor. Specifically, the frequency components in the temporal spectrum representation are matched with the frequency characteristics of the time scale dimension in the association strength tensor through inner product matching, thereby strengthening the association strength of short-term interest nodes in high-frequency scenarios and enhancing the weight of long-term interest nodes in low-frequency scenarios. The topological adjacency relationships in the spatial topology representation are structurally aligned with the node index dimension in the association strength tensor, so that interest nodes with similar topology to the current scene receive higher activation intensity. The shrinkage operation outputs a shrinkage result tensor with dimensions N×L×Q×D.

[0105] The shrinkage result tensor is decomposed using CP decomposition, which decomposes it into a weighted sum of multiple rank-tensors. Factor vectors corresponding to the node index dimensions are extracted from the factor matrix obtained from the decomposition. Each element of this vector represents the modulation coefficient of the interest node at each time scale in the current context. The modulation coefficient vector is then element-wise multiplied with the original association strength distribution. This amplifies the association strength of interest nodes with high context fit and suppresses the association strength of nodes with mismatched contexts, thus obtaining a recalibrated association strength result that reflects the influence of the current context. This result serves as a key basis for generating push decisions.

[0106] In one optional implementation, the deviation between the temporal fluctuation characteristics and the modal contributions in the causal dependency graph drives the topological restructuring of the causal dependency graph. Simultaneously, predictive reshaping of the evolutionary path in the hierarchical interest graph is performed based on the interest node trigger sequence and decay cycle of the behavioral chain in the user interaction response data, including:

[0107] Construct a mutual information flow tensor of the temporal fluctuation features and the contribution of each mode in the causal dependency graph. Perform manifold embedding on the mutual information flow tensor to obtain a causal deviation manifold. Identify the modal node clusters whose geodesic distance exceeds the stability threshold on the causal deviation manifold. Reconstruct the propagation path of the causal dependency edges along the geodesic direction for the modal node clusters to complete the topological restructuring of the causal dependency graph.

[0108] Behavioral chains are extracted from the user interaction response data and their projection trajectories on the hierarchical interest map are constructed. Lyapunov exponential analysis is performed on the projection trajectories to obtain the chaotic characteristic quantification index of the interest node trigger sequence. A nonlinear decay cycle model is fitted based on the chaotic characteristic quantification index.

[0109] The nonlinear decay cycle model is aligned with the causal deviation manifold, and the conditional activation probability density function is calculated in the aligned joint manifold space. The evolution path in the hierarchical interest map is predictively reshaped based on the weight evolution trajectory predicted by the conditional activation probability density function.

[0110] The mutual information flow tensor of temporal fluctuation features and modal contributions in the causal dependency graph is constructed by calculating the joint probability distribution and marginal probability distribution of both within the time window. Temporal fluctuation features are represented as time series vectors, with the vector length equal to the number of sampling points within the observation time window. The sampling frequency is set to once per hour, and the time window length is set to 168 hours to cover the entire period. Modal contributions in the causal dependency graph are represented as multidimensional vectors, with the number of dimensions equal to the number of modes. Each dimension corresponds to the current contribution value of a mode.

[0111] The construction of the mutual information flow tensor combines each time point of the temporal fluctuation feature with the modal contribution vector to form a three-dimensional tensor. The tensor shape is the product of the number of time points and the number of modalities. The tensor element value is the logarithm of the ratio of the joint distribution of the temporal fluctuation feature component and the modal contribution component to the marginal distribution at the corresponding time point. The joint probability distribution is estimated by statistically analyzing the co-occurrence frequency of temporal fluctuation features and modal contributions in historical data. Frequency statistics are performed using a sliding window method with a window size of 24 hours and a sliding step size of 1 hour. The calculation precision of the mutual information flow tensor element values ​​is maintained to four decimal places. During the calculation, Laplace smoothing is applied to cases with a probability value of 0, and the smoothing coefficient is set to 0.0001 to avoid numerical overflow in logarithmic operations.

[0112] The manifold embedding of the mutual information flow tensor employs an isometric mapping algorithm to project the high-dimensional tensor onto a low-dimensional causal deviation manifold space. The isometric mapping algorithm achieves dimensionality reduction by constructing a geodesic distance matrix between tensor elements, where the geodesic distance is defined as the shortest path length between two points in the tensor space. The geodesic distance matrix is ​​constructed using a k-nearest neighbor graph method, where each tensor element is connected to its k nearest neighbors, with k set to 15. The geodesic distance is calculated on the k-nearest neighbor graph using Dijkstra's algorithm, with an upper limit of 1000 iterations. The convergence criterion is that the distance change between two consecutive iterations is less than 0.001. The target dimension of the manifold embedding is set to 3D, and the embedding coordinates are obtained by eigenvalue decomposition of the geodesic distance matrix, retaining the eigenvectors corresponding to the top three largest eigenvalues ​​as the embedding coordinate axes. The causal deviation manifold is represented as a point cloud distribution in 3D space, where each point corresponds to the projected coordinates of an element in the mutual information flow tensor in the low-dimensional space.

[0113] The identification of modal node clusters on causal deviation manifolds where geodesic distances exceed a stability threshold is achieved using a density clustering algorithm. The density clustering algorithm scans all points in the causal deviation manifold, calculating the number of neighboring points within a radius for each point. The radius parameter is set to 0.3 times the average distance between points in the manifold space. Points with more than 8 neighboring points are marked as core points. After clustering, the variance of the geodesic distance within each cluster is calculated. Clusters with geodesic distance variances exceeding a stability threshold are identified as anomalous modal node clusters. The stability threshold is set to the mean of the geodesic distance variances of all clusters plus 1.5 times the standard deviation.

[0114] The reconstruction of the propagation path of causal dependency edges along the geodesic direction of modal node clusters is achieved by calculating the geodesic direction vector from the cluster center to each modal node in the causal dependency graph. The cluster center is defined as the mean point of the coordinates of all points in the cluster in the manifold space. The geodesic direction vector is calculated as the normalized vector of the difference between the coordinates of the cluster center and the target modal node in the manifold space, and the normalization operation is to divide the vector by its Euclidean modulus. The reconstruction of the propagation path of causal dependency edges adjusts the weights of the original dependency edges according to the geodesic direction vector. The adjustment coefficient is the cosine similarity between the geodesic direction vector and the original dependency edge direction vector. When the cosine similarity is greater than 0.5, the weights of the original dependency edges remain unchanged. When the cosine similarity is between 0 and 0.5, the weights of the dependency edges are multiplied by the similarity value to achieve attenuation. When the cosine similarity is less than 0, the weights of the dependency edges are reset to zero, indicating that the dependency relationship is invalid. The reconstructed causal dependency graph retains dependency edges with weights greater than 0.1, and removes dependency edges with weights less than 0.1 from the topology.

[0115] The extraction of behavioral chains from user interaction response data is based on a sequence of interaction events sorted by timestamps. Each interaction event includes fields such as event type, trigger time, associated media information identifier, dwell time, and interaction depth. Event types cover five operations: click, swipe, share, favorite, and comment. A behavioral chain is defined as a continuous sequence of interaction events with a time interval less than a threshold, set to 300 seconds. The projection trajectory of the behavioral chain onto the hierarchical interest graph is constructed by associating each interaction event with the most matching interest node. The matching degree is calculated as the cosine similarity between the content feature vector of the media information associated with the interaction event and the feature vector of the interest node. The interest node with the highest similarity is selected as the projection node. The projection trajectory is represented as a sequence of interest nodes. When there is an evolutionary path connection between adjacent nodes in the sequence, the path connection is preserved; when no path connection exists, a virtual path is added and marked with a weight of 0.

[0116] Lyapunov exponent analysis of projected trajectories is achieved by calculating the expansion rate of the trajectory in the interest graph space. The Lyapunov exponent is defined as the average of the logarithmic growth rate of the distance between trajectory points at adjacent time steps over the time series. The distance between trajectory points is calculated as the Euclidean distance between interest nodes in the feature space, using a 512-dimensional node feature vector. The logarithmic growth rate is calculated as the natural logarithm of the distance at the current time step divided by the distance at the previous time step, with the time step length set as the average interval between individual interaction events. The calculation window length for the Lyapunov exponent is set to 20 interaction events, and the arithmetic mean of the logarithmic growth rate within the window is used as the estimated value of the Lyapunov exponent for that window. A Lyapunov exponent greater than 0 indicates that the trajectory exhibits chaotic divergent characteristics, an exponent less than 0 indicates that the trajectory converges to a stable attractor, and an exponent close to 0 indicates that the trajectory is in a critical state. The quantification index of chaotic characteristics is defined as the absolute value of the Lyapunov exponent; the larger the absolute value, the higher the degree of chaos.

[0117] The nonlinear decay cycle model is fitted using the product of an exponential decay function and a periodic modulation function. The exponential decay function represents the overall decreasing trend of interest intensity over time. The decay rate parameter is obtained by fitting the historical activation intensity sequence of interest nodes using the least squares method, with a fitting window length set to 30 days. The periodic modulation function uses a superposition of sine functions to represent the periodic fluctuations of interest intensity. The period parameter is obtained by extracting the dominant frequency components of the activation intensity sequence using a fast Fourier transform, retaining the top three frequency components by energy percentage as the period parameter. The parameters of the nonlinear decay cycle model include initial intensity, decay rate, periodic frequency, and phase shift. The initial parameter values ​​are determined by fitting historical data, and the fitting error is evaluated using the root mean square error. When the error is less than 0.05, the model parameters are fixed. The model predicts that the decay cycle of the interest node activation intensity is the time required for the activation intensity to decay to 50% of its initial value. The cycle length is obtained by solving the decay model equation using a bisection method of numerical iteration, with an iteration precision set to 0.1 days.

[0118] The manifold alignment of the nonlinear decay cycle model and the causal deviation manifold is achieved through a rotational transformation of the spatial coordinate system using Prouk analysis. Prouk analysis calculates the covariance matrix of corresponding point sets in the two manifold spaces, and the singular value decomposition of the covariance matrix yields the rotation matrix and scaling factor. The rotation matrix is ​​applied to the trajectory coordinates of the nonlinear decay cycle model in the time-intensity space, and the rotated trajectory coordinates are projected onto the 3D space of the causal deviation manifold. The scaling factor is used to adjust the scale consistency between the two spaces; it is calculated as the standard deviation of the point cloud distribution in the causal deviation manifold space divided by the standard deviation of the trajectory point distribution in the decay model. The aligned joint manifold space merges the point cloud data of the two manifolds. The merging operation associates the trajectory points of the decay model with the points of the causal deviation manifold through nearest neighbor matching, with the matching distance threshold set to 0.5 times the average point spacing in the joint space.

[0119] The conditional activation probability density function is calculated in the joint manifold space using a kernel density estimation method. The kernel density estimation employs a Gaussian kernel function, with the bandwidth parameter set to 0.2 times the standard deviation of the point cloud distribution in the joint space. The conditional activation probability is defined as the probability of activating a specific interest node in a future time step given the current state of the interest node. This probability is calculated as the frequency of the target node appearing in the neighborhood of the current state point in the joint manifold space divided by the total number of neighborhood points. The neighborhood is defined as a spherical region centered on the current state point with a radius equal to twice the bandwidth parameter. The output of the conditional activation probability density function is a vector representing the activation probability distribution of each interest node in the future time step. The vector length is equal to the total number of interest nodes, and the sum of the vector elements is normalized to 1.

[0120] The weighted evolution trajectory prediction is based on the time series integral of the conditional activation probability density function, with a prediction time span of 7 days and a time resolution of 1 day. The weighted evolution trajectory is calculated as a weighted average of the conditional activation probability and the current evolution path weight at each time step, with the weighting coefficient being the conditional activation probability value. The evolution path weight is updated by multiplying the current weight by 1, subtracting the decay rate, and adding the conditional activation probability multiplied by the growth rate. The decay rate is extracted from a nonlinear decay cycle model, and the growth rate is set to 0.3 times the decay rate. The predicted weighted evolution trajectory is represented as a two-dimensional time-weight curve, with the number of curve sampling points being the prediction time span divided by the time resolution plus 1. Predictive reshaping of the evolution path adjusts the path connectivity in the hierarchical interest map based on the trend of the weighted evolution trajectory. Paths with an upward trend in weight are retained and their weight values ​​are enhanced; paths with a downward trend in weight are marked as awaiting elimination; and paths with balanced weights maintain their existing weights. The weight enhancement operation multiplies the original weight by 1 and adds the absolute value of the upward slope, which is obtained by fitting the weighted evolution trajectory using linear regression. Paths marked as pending elimination are removed from the graph when their weights decrease for three consecutive prediction periods. The removal operation triggers path reconnection of adjacent nodes to maintain graph connectivity.

[0121] In one optional implementation, a mutual information flow tensor is constructed between the temporal fluctuation characteristics and the modal contributions in the causal dependency graph. Manifold embedding is performed on the mutual information flow tensor to obtain a causal deviation manifold. On the causal deviation manifold, clusters of modal nodes whose geodesic distances exceed a stability threshold are identified, including:

[0122] The intrinsic mode function set is obtained by performing Hilbert-Huang transform on the time series fluctuation characteristics. For each mode contribution in the causal dependency graph, an attractor trajectory in the phase space is constructed. The transfer entropy flow between the intrinsic mode function set and the attractor trajectory is calculated. The transfer entropy flow is tensorized in the modal dimension, time delay dimension and embedding dimension to construct a mutual information flow tensor. Redundant degrees of freedom are eliminated by performing symplectic geometric reduction on the mutual information flow tensor.

[0123] The reduced mutual information flow tensor is embedded into a Riemannian manifold with curvature constraints. During the embedding process, hyperbolic and elliptical regions of the causal deviation manifold are identified by the sign distribution of curvature. The variational geodesic distance between modal nodes is calculated on the causal deviation manifold using the geodesic equation of hybrid geometry.

[0124] A persistent homology group of modal nodes is constructed on the causal deviation manifold. A candidate set of topologically stable modal node clusters is identified by the Betti number sequence of the persistent homology group. For each cluster in the candidate set of modal node clusters, the entropy rate of its internal variational geodesic distance is calculated. The entropy rate is compared with a stability threshold to identify modal node clusters whose geodesic distance exceeds the stability threshold.

[0125] like Figure 2 As shown, the method includes:

[0126] When identifying anomalous modal clusters on causal deviation manifolds, the time-series fluctuation characteristics in the collected user interaction response data are first processed using Hilbert-Huang transform. This transform decomposes the non-stationary signal into a finite number of intrinsic mode functions (IMFs) through empirical mode decomposition, with each IMF representing an oscillation mode at a specific time scale. The decomposition process employs a screening iterative algorithm to identify all local extrema of the original time-series fluctuation characteristic signal $x(t)$, construct upper and lower envelopes through cubic spline interpolation, calculate the envelope mean and subtract it from the original signal, repeating this process until the residual satisfies the monotonicity condition of the IMFs, ultimately obtaining a set of multiple IMFs: IMF_1, IMF_2, ..., IMF_n.

[0127] For the independent and interactive contributions of each mode in the causal dependency graph, attractor trajectories are constructed in phase space. Specifically, a time-delay embedding method is used, selecting a delay time τ and an embedding dimension m to reconstruct the one-dimensional contribution time series into an m-dimensional phase space trajectory. The delay time is determined by the first zero point of the autocorrelation function, and the optimal embedding dimension is determined using the spurious nearest neighbor method to ensure that the attractor trajectory can accurately characterize the dynamics of the modal contributions.

[0128] When calculating the transfer entropy flow between the intrinsic mode functions (IMFs) and attractor trajector trajector trajector 'A_i', a conditional mutual information quantization method is employed. For the i-th mode, the transfer entropy between its IMF_i and the contributing attractor trajectory A_i is defined as the information content of A_i predicting the future state of IMF_i. In actual calculations, continuous trajectories are discretized using symbolic encoding, and the transfer entropy values ​​are estimated through the conditional probability distribution of the statistical symbol sequence. The transfer entropy values ​​corresponding to all modes are arranged according to the mode dimension, time delay dimension, and embedding dimension to form a third-order mutual information flow tensor τ. This tensor has a large number of redundant degrees of freedom, which are eliminated using a symplectic geometric reduction method. The symplectic reduction process preserves the symplectic structure invariance of the tensor, and the conservation laws of Hamiltonian systems are used to identify and reduce redundant dimensions, resulting in a rank-reduced mutual information flow tensor.

[0129] Riemannian manifold embedding is applied to the reduced mutual information flow tensor, with curvature constraints imposed during the embedding process to preserve the geometric information of causal deviations. Hyperbolic and elliptical regions are identified by calculating the sign of the local cross-sectional curvature of the manifold. Hyperbolic regions correspond to manifolds with negative curvature, representing divergent deviations in modal contributions; elliptical regions correspond to manifolds with positive curvature, representing convergent deviation modes. On causal deviation manifolds with mixed geometry, the geodesic equations are solved using variational principles, and the shortest path length between any two modal nodes is calculated as the variational geodesic distance.

[0130] To identify topologically stable modal clusters, a persistent homology group is constructed on a causal deviating manifold. This algebraic topology tool generates a series of simplicial complexes by continuously changing a threshold parameter, tracking the appearance and disappearance of topological features at different scales. The Betti number sequence of each dimension of the persistent homology group is calculated, where the zero-dimensional Betti number represents the number of connected components, and the one-dimensional Betti number represents the number of loops. Betti numbers with long persistence correspond to topologically stable structures, thereby determining the candidate set of modal clusters.

[0131] For each modal node cluster in the candidate set, the variational geodesic distance of all node pairs within the cluster is calculated, a distance sequence is constructed, and its entropy rate is calculated. The entropy rate characterizes the complexity and uncertainty of the distance distribution. When the entropy rate exceeds a preset stability threshold, it is determined that there is a significant causal deviation within the modal node cluster, and it is marked as an anomalous cluster that requires topological reorganization.

[0132] A second aspect of the present invention provides a media information precision push system that integrates multi-source data, comprising:

[0133] The data collection unit is used to collect multimodal interaction behavior data of users from heterogeneous data sources and to perform cross-modal semantic decoupling on the multimodal interaction behavior data;

[0134] The feature decoupling unit is used to identify the independent contribution and interactive contribution of each modality to the formation of user interests by constructing a causal dependency graph between modalities, and to perform differential screening and intensity adjustment of each modal feature based on the complementary and competitive relationship between the independent contribution and the interactive contribution, so as to obtain a decoupled feature representation that eliminates redundant interference between modalities;

[0135] The graph construction unit is used to construct a hierarchical interest graph of the user based on the decoupled feature representation, and to characterize the evolution path between interest nodes at different time scales through graph structure;

[0136] The push decision unit is used to perform multi-hop reasoning on the hierarchical interest graph based on the content characteristics of the media information to be pushed, calculate the correlation strength distribution between the media information to be pushed and interest nodes at each time scale, dynamically recalibrate the correlation strength distribution in combination with the current context features, and generate a push decision.

[0137] The graph optimization unit is used to execute the push decision and observe the user interaction response data after the push. It identifies the response preference migration pattern based on the temporal fluctuation characteristics of different modal feedbacks in the user interaction response data, drives the topological restructuring of the causal dependency graph by using the deviation of the temporal fluctuation characteristics from the contribution of each modality in the causal dependency graph, and predictively reshapes the evolution path in the hierarchical interest graph based on the interest node trigger sequence and decay cycle of the behavioral chain in the user interaction response data.

[0138] A third aspect of the present invention provides an electronic device, comprising:

[0139] processor;

[0140] Memory used to store processor-executable instructions;

[0141] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0142] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0143] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for precise delivery of media information by integrating multi-source data, characterized in that, include: Collect multimodal interaction behavior data of users from heterogeneous data sources, and perform cross-modal semantic decoupling on the multimodal interaction behavior data; By constructing a causal dependency graph between modalities, the independent contribution and interactive contribution of each modality to the formation of user interests are identified. Based on the complementary and competitive relationship between the independent contribution and the interactive contribution, the modal features are differentiated and their intensity is adjusted to obtain a decoupled feature representation that eliminates redundant interference between modalities. Based on the decoupled feature representation, a hierarchical interest graph of the user is constructed, and the evolution path between interest nodes at different time scales is characterized by graph structure. Based on the content characteristics of the media information to be pushed, multi-hop reasoning is performed on the hierarchical interest graph to calculate the correlation strength distribution between the media information to be pushed and interest nodes at each time scale. The correlation strength distribution is dynamically recalibrated in combination with the current context features to generate a push decision. The system executes the push decision and observes the user interaction response data after the push. Based on the temporal fluctuation characteristics of different modal feedbacks in the user interaction response data, it identifies the response preference migration pattern. It uses the deviation between the temporal fluctuation characteristics and the contribution of each modality in the causal dependency graph to drive the topological restructuring of the causal dependency graph. At the same time, it predictively reshapes the evolution path in the hierarchical interest graph based on the interest node trigger sequence and decay cycle of the behavioral chain in the user interaction response data.

2. The method according to claim 1, characterized in that, Based on the complementary and competitive relationship between the independent contribution and the interactive contribution, differentiated screening and intensity adjustment are performed on each modal feature to obtain a decoupled feature representation that eliminates redundant interference between modalities, including: For each modal node in the causal dependency graph, the difference in the direction and intensity of the independent contribution and the interaction contribution in the process of user interest formation is quantified. When the two have opposite directions and the intensity difference exceeds the cooperation threshold, they are identified as competitive relationships. When the two have the same direction and there is a mutual reinforcement effect, they are identified as complementary relationships. For modal feature pairs identified as competing relationships, the dominant contribution mode and the suppressed contribution mode are determined based on the intensity difference. The features of the dominant contribution mode are selectively retained by increasing their expression weight in the decoupled feature representation, and the features of the suppressed contribution mode are selectively suppressed by decreasing their expression weight. For modal feature pairs identified as complementary, a synergistic gain coefficient between modalities is constructed based on the strength of the mutual enhancement effect. The synergistic gain coefficient is then used to amplify the expression weights of each modal feature participating in the complementarity, thereby strengthening the complementary relationship in the decoupled feature representation. The modal features, after selective retention, selective suppression, and linked amplification, are combined according to the adjusted expression weights to form a decoupled feature representation.

3. The method according to claim 1, characterized in that, Based on the decoupled feature representation, a hierarchical interest graph of the user is constructed, and the evolution path between interest nodes at different time scales is characterized by a graph structure, including: The time dependence of each modal feature in the decoupled feature representation is measured. The time-scale representation vector of the feature is constructed by analyzing the activation frequency distribution and persistence distribution of each feature in the user interaction sequence. The features are aggregated into interest node levels with different time dependencies according to the topological distribution of the time-scale representation vector in the feature space. For each node in the interest node hierarchy, the semantic association tensor and temporal transition probability matrix between nodes are calculated. The semantic association tensor and the temporal transition probability matrix are fused by tensor operation to obtain a joint representation matrix that characterizes the evolutionary association between nodes. Node pairs with association strength exceeding the evolution threshold are extracted from the joint representation matrix as candidate node pairs with evolutionary relationship. For the candidate node pair, the starting node and ending node of the evolution path are determined based on the difference in time dependency between the nodes and the direction of semantic drift. When the time dependency of the starting node is less than that of the ending node, a fixed evolution path is established. When the time dependency of the starting node is greater than that of the ending node, an induced evolution path is established.

4. The method according to claim 1, characterized in that, Based on the content characteristics of the media information to be pushed, multi-hop inference is performed on the hierarchical interest graph to calculate the association strength distribution between the media information to be pushed and interest nodes at each time scale. The association strength distribution is then dynamically recalibrated in conjunction with the current context features to generate a push decision, including: Extract the content features of the media information to be pushed and construct a multi-scale representation of the content features. Then, perform semantic matching between the multi-scale representation and the feature representations of interest nodes at each level in the hierarchical interest graph to obtain the initial matching response distribution. Using the initial matching response distribution as the starting activation state for multi-hop inference, information is propagated along the evolution path on the hierarchical interest graph. Each hop attenuates and transfers the activation intensity of the current node through path weights, and modulates and amplifies the propagation intensity according to the semantic association between nodes. After multi-hop propagation, the cumulative activation intensity of interest nodes at each time scale is used as the association intensity distribution. Obtain the current context features and construct a spatiotemporal representation matrix of the context features. Perform tensor shrinkage operation on the spatiotemporal representation matrix and the association strength distribution to obtain context-sensitive association strength recalibration results. Based on the adjusted association strength of interest nodes at each time scale in the association strength recalibration results, a multi-scale fusion vector is constructed. The multi-scale fusion vector is weighted and aggregated to obtain a comprehensive push score. When the comprehensive push score meets the push conditions, a push decision is generated for the media information to be pushed.

5. The method according to claim 4, characterized in that, Obtain the current context features and construct a spatiotemporal representation matrix of the context features. Perform tensor shrinkage operation on the spatiotemporal representation matrix and the association strength distribution to obtain context-sensitive association strength recalibration results, including: The current context features are obtained, and the current context features are subjected to spatiotemporal separation transformation to obtain temporal context components and spatial context components. The frequency domain characteristics of the temporal context components are extracted to construct a temporal spectrum representation, and the topological characteristics of the spatial context components are extracted to construct a spatial topological representation. The temporal spectrum representation and the spatial topological representation are subjected to tensor outer product operation to construct the spatiotemporal representation matrix of the context features. The association strength values ​​of interest nodes at each time scale in the association strength distribution are extracted and an association strength tensor is constructed. The association strength tensor is then dimensionally expanded so that its dimensional structure forms a contractible correspondence with the dimensional structure of the spatiotemporal representation matrix. The spatiotemporal representation matrix and the correlation strength tensor are subjected to tensor shrinkage operation. In the tensor shrinkage operation, the temporal scale dimension of the correlation strength tensor is frequency-matched modulated by the frequency domain component of the temporal spectrum representation, and the node index dimension of the correlation strength tensor is structure-matched modulated by the topological component of the spatial topology representation, so as to obtain the shrinkage result tensor of the fused spatiotemporal characteristics. Tensor decomposition is performed on the tensor of the contraction result to separate the modulation coefficients of interest nodes at each time scale in the current context. The modulation coefficients are then multiplied element-wise with the association strength distribution to obtain the context-sensitive association strength recalibration result.

6. The method according to claim 1, characterized in that, The topological restructuring of the causal dependency graph is driven by the deviation between the temporal fluctuation characteristics and the modal contributions in the causal dependency graph. Simultaneously, the evolutionary path in the hierarchical interest graph is predictively reshaped based on the interest node trigger sequence and decay cycle of the behavioral chain in the user interaction response data, including: Construct a mutual information flow tensor of the temporal fluctuation features and the contribution of each mode in the causal dependency graph. Perform manifold embedding on the mutual information flow tensor to obtain a causal deviation manifold. Identify the modal node clusters whose geodesic distance exceeds the stability threshold on the causal deviation manifold. Reconstruct the propagation path of the causal dependency edges along the geodesic direction for the modal node clusters to complete the topological restructuring of the causal dependency graph. Behavioral chains are extracted from the user interaction response data and their projection trajectories on the hierarchical interest map are constructed. Lyapunov exponential analysis is performed on the projection trajectories to obtain the chaotic characteristic quantification index of the interest node trigger sequence. A nonlinear decay cycle model is fitted based on the chaotic characteristic quantification index. The nonlinear decay cycle model is aligned with the causal deviation manifold, and the conditional activation probability density function is calculated in the aligned joint manifold space. The evolution path in the hierarchical interest map is predictively reshaped based on the weight evolution trajectory predicted by the conditional activation probability density function.

7. The method according to claim 6, characterized in that, Construct a mutual information flow tensor of the temporal fluctuation features and the modal contributions in the causal dependency graph, perform manifold embedding on the mutual information flow tensor to obtain a causal deviation manifold, and identify modal node clusters on the causal deviation manifold whose geodesic distances exceed the stability threshold, including: The intrinsic mode function set is obtained by performing Hilbert-Huang transform on the time series fluctuation characteristics. For each mode contribution in the causal dependency graph, an attractor trajectory in the phase space is constructed. The transfer entropy flow between the intrinsic mode function set and the attractor trajectory is calculated. The transfer entropy flow is tensorized in the modal dimension, time delay dimension and embedding dimension to construct a mutual information flow tensor. Redundant degrees of freedom are eliminated by performing symplectic geometric reduction on the mutual information flow tensor. The reduced mutual information flow tensor is embedded into a Riemannian manifold with curvature constraints. During the embedding process, hyperbolic and elliptical regions of the causal deviation manifold are identified by the sign distribution of curvature. The variational geodesic distance between modal nodes is calculated on the causal deviation manifold using the geodesic equation of hybrid geometry. A persistent homology group of modal nodes is constructed on the causal deviation manifold. A candidate set of topologically stable modal node clusters is identified by the Betti number sequence of the persistent homology group. For each cluster in the candidate set of modal node clusters, the entropy rate of its internal variational geodesic distance is calculated. The entropy rate is compared with a stability threshold to identify modal node clusters whose geodesic distance exceeds the stability threshold.

8. A media information precision push system integrating multi-source data, used to implement the method of any one of claims 1-7, characterized in that, include: The data collection unit is used to collect multimodal interaction behavior data of users from heterogeneous data sources and to perform cross-modal semantic decoupling on the multimodal interaction behavior data; The feature decoupling unit is used to identify the independent contribution and interactive contribution of each modality to the formation of user interests by constructing a causal dependency graph between modalities, and to perform differential screening and intensity adjustment of each modal feature based on the complementary and competitive relationship between the independent contribution and the interactive contribution, so as to obtain a decoupled feature representation that eliminates redundant interference between modalities; The graph construction unit is used to construct a hierarchical interest graph of the user based on the decoupled feature representation, and to characterize the evolution path between interest nodes at different time scales through graph structure; The push decision unit is used to perform multi-hop reasoning on the hierarchical interest graph based on the content characteristics of the media information to be pushed, calculate the correlation strength distribution between the media information to be pushed and interest nodes at each time scale, dynamically recalibrate the correlation strength distribution in combination with the current context features, and generate a push decision. The graph optimization unit is used to execute the push decision and observe the user interaction response data after the push. It identifies the response preference migration pattern based on the temporal fluctuation characteristics of different modal feedbacks in the user interaction response data, drives the topological restructuring of the causal dependency graph by using the deviation of the temporal fluctuation characteristics from the contribution of each modality in the causal dependency graph, and predictively reshapes the evolution path in the hierarchical interest graph based on the interest node trigger sequence and decay cycle of the behavioral chain in the user interaction response data.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images