Similar event analysis method based on subgraph feature reconstruction and multi-modal fusion

Abstract

The invention relates to the technical field of event analysis, and discloses a similar event analysis method based on subgraph feature reconstruction and multi-modal fusion, which comprises the following steps: generating an event knowledge graph according to event text data, calculating the extended Jaccard similarity of the event knowledge graph, screening associated event node pairs through a self-adaptive threshold value and adding hidden edges, generating a reconstructed implicit event knowledge graph; performing global bidirectional breadth-first search on the event knowledge graph to obtain a candidate extended event node set, dynamically screening extended event nodes through attention weight and semantic matching score, and evaluating and merging extended event knowledge graphs based on path importance; generating multi-granularity embedding features for the event knowledge graph, the hidden event knowledge graph and the extended event knowledge graph through a graph convolutional network; and carrying out weighted aggregation on the multi-granularity embedded features, carrying out high-order interaction through a neural tensor network, and carrying out dimensionality reduction through a multi-layer perceptron to output an event similarity score. According to the invention, accurate calculation of event similarity is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of event analysis, and in particular to a similar event analysis method based on subgraph feature reconstruction and multimodal fusion. Background Art

[0002] In knowledge graph-driven intelligent applications, event similarity analysis is one of the core technologies supporting personalized recommendations, risk warnings, and decision reasoning. With the explosive growth of multimodal data, the definition of an event has expanded from a single structure to a complex subgraph consisting of entities, relationships, attributes, and text descriptions. For example, in social networks, user behavior events may involve user nodes, interaction relationships, temporal attributes, and dynamic comment text; in e-commerce scenarios, product transaction events may involve product entities, user profiles, supply chain paths, and product descriptions. Accurately calculating the similarity of such heterogeneous event subgraphs has become a key challenge in improving the performance of intelligent systems.

[0003] Existing graph similarity calculation methods primarily extract graph structural features based on graph neural networks and calculate similarity using embedding vectors. For example, SimGNN (AAAI 2019) utilizes graph attention networks to aggregate node features and calculate graph-level embedding similarity, GraphSim (CIKM 2020) extracts multi-order neighborhood information through multi-layer graph convolutional networks, and NA-GSL (KDD 2021) introduces a self-attention mechanism to optimize cross-graph interaction modeling. These methods perform well on homogeneous graphs (i.e., graphs with a single node and edge type), but have significant limitations when processing heterogeneous graphs containing multiple node and edge types within knowledge graphs. On the one hand, they rely on explicit graph structure and cannot effectively mine potential connections between directly connected nodes of the same type. On the other hand, analyzing event subgraphs in isolation results in a loss of contextual semantics in the global knowledge graph. Furthermore, the lack of dynamic integration of graph structure, textual semantics, and extended neighborhood features limits the collaborative optimization capabilities of multimodal information, ultimately resulting in insufficient accuracy in similarity analysis.

[0004] Therefore, there is an urgent need for a similar event analysis method for heterogeneous knowledge graphs to solve the problems of implicit association mining, multi-granularity feature fusion and noise suppression, so as to meet the needs of high-precision and high-robustness event analysis in actual scenarios. Summary of the Invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a similar event analysis method based on subgraph feature reconstruction and multimodal fusion.

[0006] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: A similar event analysis method based on subgraph feature reconstruction and multimodal fusion includes the following steps: Generate an event knowledge graph based on event text data, calculate the extended Jaccard similarity of the event knowledge graph, filter related event node pairs through adaptive thresholds and add implicit edges to generate a reconstructed implicit event knowledge graph; Perform a global bidirectional breadth-first search on the event knowledge graph to obtain a set of candidate extended event nodes, dynamically filter extended event nodes by attention weight and semantic matching score, and merge the extended event knowledge graph based on path importance evaluation; Generate multi-granularity embedding features for event knowledge graph, implicit event knowledge graph and extended event knowledge graph through graph convolutional network; Multi-granularity embedded features are weightedly aggregated, high-order interactions are performed through neural tensor networks, and event similarity scores are output through dimensionality reduction using a multi-layer perceptron.

[0007] Furthermore, the extended Jaccard similarity of the event knowledge graph is calculated as follows:

[0008] in, is the extended Jaccard similarity of event nodes (u, v) of the same category that are not directly connected in the event knowledge graph, For event nodes of The set of neighbors, For event nodes of The set of neighbors, is the attenuation coefficient.

[0009] Furthermore, the adaptive threshold is calculated as follows:

[0010] in, is the adaptive threshold, is the global similarity mean, is the global similarity standard deviation, is the sensitivity coefficient.

[0011] Furthermore, a global bidirectional breadth-first search is performed on the event knowledge graph to obtain a set of candidate extended event nodes, including: Perform a bidirectional breadth-first search on the event node set of the event knowledge graph to obtain its k-hop external neighbor event node set; The external neighbor event node set is detected to form a bridge structure with relation paths connected to multiple event nodes, and a candidate extended event node set is generated.

[0012] Furthermore, the attention weight is calculated as follows:

[0013] in, is the attention weight of the event node pair (u,v), is the embedding vector of event node u,v, is the learning parameter, T is the transpose, is the activation function.

[0014] Furthermore, the semantic matching score is calculated as follows:

[0015] in, is the semantic matching score of event node u, is the extended event node set, v is the event node, Embed function for text.

[0016] Furthermore, the extended event nodes are dynamically screened by attention weight and semantic matching score, including: The event node score is calculated based on the attention weight and semantic matching score, specifically:

[0017] in, Score event node u, is the weighting coefficient, is the attention weight of the event node pair (u,v), is the semantic matching score of event node u; The extended event nodes whose event node u score is greater than the set threshold are retained, and the retained extended event nodes form a refined event extension set.

[0018] Furthermore, the event knowledge graph is merged and extended based on the path importance evaluation, including: Perform meta-path-guided subgraph stitching on the filtered extended event nodes, extract all relationship paths between the filtered extended event nodes and the event knowledge graph, and calculate the path weights based on the PathRank path importance evaluation model, specifically:

[0019] in, For path weight, For relationship type The prior weight of is the path length attenuation factor, For path Number of relationships; Select the optimal path, merge the relay event nodes and edges involved into the event knowledge graph, and obtain the extended event knowledge graph.

[0020] Furthermore, the multi-granularity embedding features are weighted and aggregated as follows:

[0021] in, is the weighted aggregation vector, is the original embedding vector of the event knowledge graph, To extend the event knowledge graph, is the weight coefficient, is the embedding vector of the implicit event knowledge graph n, and k is the number of implicit event knowledge graphs.

[0022] Furthermore, high-order interactions are performed through neural tensor networks:

[0023] in, is the interaction function, is the activation function, is the weight set, c is the dimension hyperparameter, is the comprehensive embedding vector of event knowledge graph j, is a node set, Event Knowledge Graph i The comprehensive embedding vector of To learn the weights.

[0024] The present invention has the following beneficial effects: The present invention addresses the problems in the prior art of ignoring implicit associations in heterogeneous graphs and underutilizing the global structural information of knowledge graphs. It achieves accurate calculation of event similarity through implicit subgraph mining, subgraph neighborhood expansion and multi-granularity feature fusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a schematic diagram of the architecture of a similar event analysis method based on subgraph feature reconstruction and multimodal fusion; Figure 2 This is a flowchart of a similar event analysis method based on subgraph feature reconstruction and multimodal fusion. DETAILED DESCRIPTION

[0026] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0027] This invention solves the problem of insufficient mining of implicit relations of event subgraphs and difficulty in global semantic integration in heterogeneous graph scenarios through subgraph feature reconstruction and multimodal fusion. The core is the three-stage framework of "endogenous reconstruction-external expansion-multi-source fusion". Figure 1 shown.

[0028] like Figure 2 As shown, an embodiment of the present invention provides a similar event analysis method based on subgraph feature reconstruction and multimodal fusion, comprising the following steps S1 to S4: S1. Generate an event knowledge graph based on event text data, calculate the extended Jaccard similarity of the event knowledge graph, filter related event node pairs through adaptive thresholds and add implicit edges to generate a reconstructed implicit event knowledge graph; In an optional embodiment of the present invention, step S1 first generates an event knowledge graph based on event text data. The time text data here can be the metadata of the book and the user's rating information included in the book event, or it can be the user's rating and basic attribute information of the movie included in the movie recommendation event.

[0029] After the event knowledge graph is generated in step S1, the extended Jaccard similarity of the two event knowledge graphs is calculated respectively:

[0030] in, is the extended Jaccard similarity of event nodes (u, v) of the same category that are not directly connected in the event knowledge graph, For event nodes of The set of neighbors, For event nodes of The set of neighbors, is the attenuation coefficient. This formula assigns a higher weight to the first-order neighborhood while retaining the complementary information of the second-order neighborhood. This not only avoids the noise interference caused by over-expansion, but also enhances the ability to discover associations across community nodes.

[0031] For each node u in the similar subgraph, extract its first-order neighborhood With the second-order neighborhood , build an extended neighborhood set Then, we traverse all pairs of nodes (u,v) of the same category that are not directly connected, and calculate based on their extended neighborhood After calculation, when The model can show the best effect.

[0032] After completing the above steps in step S1, an adaptive threshold determination strategy is adopted to perform quantile truncation on the similarity matrix:

[0033] in, is the adaptive threshold, is the global similarity mean, is the global similarity standard deviation, is the sensitivity coefficient. Compared with a fixed threshold, this method can better adapt to the distribution characteristics of different subgraphs, thereby adaptively screening significantly associated node pairs and avoiding the distribution bias of a fixed threshold.

[0034] Step S1 finally constructs the implicit event knowledge graph. , then add a hidden edge between u and v to generate the reconstructed hidden subgraph: .

[0035] S2. Perform a global bidirectional breadth-first search on the event knowledge graph to obtain a set of candidate extended event nodes, dynamically filter the extended event nodes based on attention weights and semantic matching scores, and merge the extended event knowledge graph based on path importance evaluation; In an optional embodiment of the present invention, step S2 first performs a bidirectional breadth-first search in the global knowledge graph G based on the event subgraph node set V. That is, for each event node , traverse its k-hop neighbors (usually k=2) and obtain the set of directly related external nodes ; For external nodes , detect whether it is connected to multiple event nodes through a specific relationship path to form a bridge structure. This process generates a set of candidate expansion nodes:

[0036] And retain the connection path P(u,v) as a potential extension edge.

[0037] Subsequently, in order to overcome the noise problem caused by simple structural expansion, a dynamic screening strategy that integrates the attention mechanism and embedding similarity is adopted. The association strength between event nodes and candidate nodes is calculated through the pre-trained graph attention network:

[0038] in, is the attention weight of the event node pair (u,v), is the embedding vector of event node u,v, is the learning parameter, T is the transpose, is the activation function.

[0039] Then the semantic matching score is calculated:

[0040] in, is the semantic matching score of event node u, is the extended event node set, v is the event node, Embed function for text.

[0041] Finally, dynamic weighted screening is performed:

[0042] in, Score event node u, is the weighting coefficient, is the attention weight of the event node pair (u,v), is the semantic matching score of event node u; Retention satisfaction Nodes, where is the empirical threshold, which is generally selected as 0.6. The retained nodes form a refined expansion set To prevent the extended subgraph from becoming an isolated connected component, meta-path-guided subgraph stitching is performed on the filtered nodes to extract all the relationship paths between the nodes in the refined extended set and the event subgraph. , and calculate the path weight based on the PathRank path importance evaluation model:

[0043] in, For path weight, For relationship type The prior weight of is the path length attenuation factor, For path Number of relationships; Finally, the optimal path is selected and the relay nodes and edges involved are merged into the extended subgraph to obtain the extended subgraph G'.

[0044] S3. Generate multi-granularity embedding features for event knowledge graph, implicit event knowledge graph and extended event knowledge graph through graph convolutional network; In an optional embodiment of the present invention, step S3 first extracts k implicit subgraphs from the target event subgraph G by using an implicit graph decomposition algorithm. And construct its extended subgraph G', where the latent subgraph is used to capture local structural features, and the extended subgraph achieves semantic enhancement by merging neighborhood nodes.

[0045] To ensure the accuracy of the embedded representation, all subgraphs (including G, and G') perform edge attribute expansion operations. The edge attribute expansion operations are as follows: set up G =( V , E ) is an n-order graph, where V is a node set, E is an edge set. u 1, u 2∈ V yes G The two nodes in v ∈ E is an edge such that v ={ u 1, u 2}, in Figure G is represented as a triple ( u 1, v , u 2) Change the triple v Modify to node u ′, the original triple ( u 1, v , u 2) Convert to two edges ( u 1, v ′, u ') and( u ′, v ′, u 2), where v ′ is a newly constructed virtual edge without information. The formula is expressed as: ( u 1, v , u 2)→{( u 1, v ′, u ′),( u ′, v ′, u 2)} Then a three-layer graph convolutional network is used for hierarchical feature aggregation, and its node-level embedding matrix Calculated by the following formula:

[0046] in is the adjacency matrix, is the node feature matrix of the fused edge attributes, N is the number of nodes, D is the embedding dimension, and GCN is the step of the graph convolutional network.

[0047] Finally, the graph-level embedding vector is generated:

[0048] In the formula represents the n-th node embedding vector, is the trainable parameter matrix, is the Sigmoid activation function. Finally, the embedding representation of the target subgraph G contains three types of multi-granularity features: the original subgraph embedding , implicit subgraph embedding set and extended subgraph embedding , providing differentiated feature expressions for subsequent similarity calculations.

[0049] S4. Perform weighted aggregation on multi-granularity embedded features, perform high-order interaction through a neural tensor network, and output event similarity scores through dimensionality reduction using a multi-layer perceptron.

[0050] In an optional embodiment of the present invention, step S4 is to achieve effective fusion of different embedding vectors and design a weighted aggregation mechanism to construct a comprehensive embedding vector of the event subgraph, which is formally defined as:

[0051] in: The original embedding vector representing the target subgraph; is the enhanced embedding of the extended neighborhood subgraph; k is the number of hidden subgraphs; The embedded representation corresponding to each latent subgraph; To satisfy The adjustable weight coefficient is verified by grid search. The model achieves optimal performance.

[0052] Next, for any event subgraph to be detected , which is integrated into and Input to a neural tensor network for modeling high-order interactions:

[0053] in , ,and are learnable weights, and is the graph-level embedding of two graphs, is an activation function, is a hyperparameter that controls the output dimension. By adopting the neural tensor network method, more stable interactive output between graphs can be obtained.

[0054] Then, the interaction feature vector is obtained through the above processing Finally, a multi-layer perceptron (MLP) is constructed to achieve feature space dimensionality reduction:

[0055] Finally, the mean square error is used as the objective function for end-to-end optimization:

[0056] Where D represents the set of event subgraph pairs in the training set, is the true similarity score. This loss function implements parameter updates through the back-propagation algorithm to ensure that the model prediction value is aligned with the true distribution.

[0057] Final output: Similarity score .

[0058] The present invention is verified and explained with specific examples below.

[0059] This paper experiments on the Yelp2018, Amazon-Book, and MovieLens-1M datasets. Yelp2018 is suitable for recommending commercial places, covering user reviews and business information in 2018; Amazon-Book is suitable for book recommendations, covering book metadata and user rating information; MovieLens-1M is suitable for movie recommendations, containing user ratings and basic attribute information of movies.

[0060] For these three datasets, all images included in the experiments were randomly divided into training, validation, and test sets in a 6:2:2 ratio. The experimental environment is based on the PyTorch framework and trained on an NVIDIA RTX 2080 GPU. In this section, the Adam optimizer was used to dynamically adjust the model's learning rate and optimize model parameters.

[0061] The present invention uses mean square error (MSE), Spearman rank correlation coefficient ( ), Kendall rank correlation coefficient ( )and The accuracy at ) to evaluate the accuracy and other performance indicators of various methods. The experimental results are shown in Tables 1, 2, and 3 below: Table 1 Experimental comparison on the Yelp2018 dataset

[0062] Table 2 Experimental comparison on the Amazon-Book dataset

[0063] Table 3 Experimental comparison on the MovieLens-1M dataset

[0064] In the task of similar event analysis, our model demonstrates comprehensive and significant performance advantages. Comparative experiments on three representative datasets—Yelp2018, Amazon-Book, and MovieLens-1M—show that our model surpasses existing models in multiple core metrics, particularly in event ranking relevance and high-precision prediction capabilities, demonstrating its robustness and generalization capabilities in complex scenarios.

[0065] The present invention achieves significant improvements in the Spearman rank coefficient ρ and Kendall rank coefficient τ, indicators that measure the quality of similar event ranking. Compared to suboptimal models, the present invention's ranking correlation is optimized by an average of more than 3%-5%, and the improvement is particularly significant in complex event scenarios, such as the MovieLens-1M dataset, where user behavior is sparse. This demonstrates that the present invention can more accurately capture the potential correlations between events and overcome the limitations of traditional models that rely on local features or a single dimension, thereby generating similarity rankings that are more consistent with real-world scenarios.

[0066] In the key cutting-edge prediction metric P@10, our method consistently leads the market, achieving an accuracy improvement of 1.5%-2%. This advantage stems from the model's deep focus on high-value events. Through dynamic weight assignment and multi-level feature fusion, our method effectively distinguishes core events from noise, ensuring accurate ranking of high-priority events in the recommendation sequence. For example, in fine-grained similarity analysis tasks such as Amazon-Book, this feature significantly improves the efficiency of matching users' core interests.

[0067] The proposed method also performs well in terms of mean squared error, with an average error reduction of approximately 2%-3% compared to the comparison model. This result reveals the model's dual advantages: on the one hand, by introducing an implicit relationship mining strategy for event subgraphs, it suppresses the risk of overfitting in complex scenarios; on the other hand, its multi-granularity edge attribute graph embedding strategy enhances the model's ability to represent sparse data, thereby improving prediction stability while maintaining prediction accuracy.

[0068] Notably, the proposed method demonstrates consistent superiority across diverse datasets, including commercial scenarios (Yelp2018), product recommendations (Amazon-Book), and film preferences (MovieLens-1M). By integrating implicit event subgraph associations, multi-granularity graph embedding, and robustness optimization strategies, the proposed method performs well in similar event analysis tasks.

[0069] Experiments on three heterogeneous datasets, Yelp2018, Amazon-Book, and MovieLens-1M, show that the proposed method has the advantages of MSE, ranking correlation, and The MSE is significantly better than the mainstream baseline model in terms of indicators such as P@10. Specifically, its MSE is 2.1%-3.8% lower than the suboptimal model on average. and The performance of the proposed model improved by 3.5%-6.2%, respectively, and the P@10 optimization range reached 1.5%-2.0%. Ablation experiments further revealed the synergistic effect of each module: removing latent subgraph mining or subgraph neighborhood expansion resulted in a 12%-20% performance drop, respectively, while abandoning the neural tensor network increased prediction error by 20%-40%, confirming the necessity of model design. Weight coefficient sensitivity experiments showed that a configuration strategy that prioritizes the original subgraph, followed by latent subgraphs, and supplemented by extended subgraphs achieves a global optimal solution on the error surface by maintaining a balance between structural stability and noise suppression.

[0070] This paper breaks through the methodological limitations of traditional graph similarity calculations, which rely too heavily on explicit structures. It proposes a progressive analysis paradigm of "endogenous reconstruction, external expansion, and multi-source fusion," providing new insights for modeling complex events within knowledge graphs. In practice, the model employed in this paper demonstrates strong generalization capabilities in both sparse data scenarios and fine-grained semantic matching tasks. By mining implicit relationships and leveraging contextual awareness, it significantly improves the interpretability and accuracy of similar event analysis.

[0071] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0072] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0073] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0074] Specific embodiments are used in the present invention to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

[0075] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the present invention.

Claims

1. A similar event analysis method based on subgraph feature reconstruction and multimodal fusion, characterized in that: The following steps are involved: Generate an event knowledge graph based on event text data, calculate the extended Jaccard similarity of the event knowledge graph, filter related event node pairs through adaptive thresholds and add implicit edges to generate a reconstructed implicit event knowledge graph; Perform a global bidirectional breadth-first search on the event knowledge graph to obtain a set of candidate extended event nodes, dynamically filter extended event nodes by attention weight and semantic matching score, and merge the extended event knowledge graph based on path importance evaluation; Generate multi-granularity embedding features for event knowledge graph, implicit event knowledge graph and extended event knowledge graph through graph convolutional network; Multi-granularity embedded features are weightedly aggregated, high-order interactions are performed through neural tensor networks, and event similarity scores are output through dimensionality reduction using a multi-layer perceptron.

2. The similar event analysis method based on subgraph feature reconstruction and multimodal fusion according to claim 1 is characterized in that: The extended Jaccard similarity of the event knowledge graph is calculated as follows: in, is the extended Jaccard similarity of event nodes (u, v) of the same category that are not directly connected in the event knowledge graph, For event nodes of The set of neighbors, For event nodes of The set of neighbors, is the attenuation coefficient.

3. The similar event analysis method based on subgraph feature reconstruction and multimodal fusion according to claim 1 is characterized in that: The adaptive threshold is calculated as: in, is the adaptive threshold, is the global similarity mean, is the global similarity standard deviation, is the sensitivity coefficient.

4. The similar event analysis method based on subgraph feature reconstruction and multimodal fusion according to claim 1 is characterized in that: Perform a global bidirectional breadth-first search on the event knowledge graph to obtain a set of candidate extended event nodes, including: Perform a bidirectional breadth-first search on the event node set of the event knowledge graph to obtain its k-hop external neighbor event node set; The external neighbor event node set is detected to form a bridge structure with relation paths connected to multiple event nodes, and a candidate extended event node set is generated.

5. The similar event analysis method based on subgraph feature reconstruction and multimodal fusion according to claim 1 is characterized in that: The attention weight is calculated as: in, is the attention weight of the event node pair (u,v), is the embedding vector of event node u,v, is the learning parameter, T is the transpose, is the activation function.

6. The similar event analysis method based on subgraph feature reconstruction and multimodal fusion according to claim 1 is characterized in that: The semantic matching score is calculated as follows: in, is the semantic matching score of event node u, is the extended event node set, v is the event node, Embed function for text.

7. The similar event analysis method based on subgraph feature reconstruction and multimodal fusion according to claim 1 is characterized in that: Dynamically filter extended event nodes by attention weight and semantic matching score, including: The event node score is calculated based on the attention weight and semantic matching score, specifically: in, Score event node u, is the weighting coefficient, is the attention weight of the event node pair (u,v), is the semantic matching score of event node u; The extended event nodes whose event node u score is greater than the set threshold are retained, and the retained extended event nodes form a refined event extension set.

8. The similar event analysis method based on subgraph feature reconstruction and multimodal fusion according to claim 1 is characterized in that: The event knowledge graph is merged and extended based on path importance evaluation, including: Perform meta-path-guided subgraph stitching on the filtered extended event nodes, extract all relationship paths between the filtered extended event nodes and the event knowledge graph, and calculate the path weights based on the PathRank path importance evaluation model, specifically: in, For path weight, For relationship type The prior weight of is the path length attenuation factor, For path Number of relationships; Select the optimal path, merge the relay event nodes and edges involved into the event knowledge graph, and obtain the extended event knowledge graph.

9. The similar event analysis method based on subgraph feature reconstruction and multimodal fusion according to claim 1 is characterized in that: The weighted aggregation of multi-granularity embedding features is: in, is the weighted aggregation vector, is the original embedding vector of the event knowledge graph, To extend the event knowledge graph, is the weight coefficient, is the embedding vector of the implicit event knowledge graph n, and k is the number of implicit event knowledge graphs.

10. The similar event analysis method based on subgraph feature reconstruction and multimodal fusion according to claim 1, characterized in that: The high-order interactions via neural tensor networks are: in, is the interaction function, is the activation function, is the weight set, c is the dimension hyperparameter, is the comprehensive embedding vector of event knowledge graph j, is a node set, Event Knowledge Graph i The comprehensive embedding vector of To learn the weights.

Images