Social event detection method based on hierarchical multi-relation structure entropy minimization

Abstract

The present invention discloses a social event detection method based on hierarchical multi-relationship structural entropy minimization, the method comprising: integrating a BA graph generator with a large language model LLM to realize the construction and evolution of a single-relationship message graph to a multi-relationship message graph; guiding the fine adjustment and enhancement of the multi-relationship message graph structure based on the large language model LLM; decoding the basic structure of the multi-relationship message graph based on a hierarchical two-dimensional multi-relationship structural entropy minimization algorithm; obtaining social event detection based on hierarchical two-dimensional multi-relationship structural entropy minimization, and outputting multiple message graph partitions, each partition representing a social event. The method of the present invention introduces a hierarchical two-dimensional multi-relationship structural entropy minimization algorithm, which can detect communities in a multi-relationship message graph. When the multi-relationship structural entropy is used to minimize the graph entropy, a greater reduction can be observed, and the decoding of structural information is more effective, which improves the performance of social event detection.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing methods, and in particular to a social event detection method based on minimization of hierarchical multi-relationship structure entropy. Background Art

[0002] How to accurately and timely detect social events and conduct in-depth analysis and understanding of them has become a topic of great research value. Social events are complex and diverse. Different events may have different characteristics and manifestations, which requires detection algorithms to have strong adaptability and generalization capabilities. In addition, with the vigorous development of social media and the rapid spread of information, real-time detection and analysis of social events have become increasingly critical. Social event detection requires us to quickly and accurately identify, cluster, and extract interrelated clusters from massive amounts of real social information. Due to the strong dynamics of social information, social events often do not exist in isolation, but form a complex network structure containing multiple relationships. Therefore, traditional text mining methods are difficult to effectively deal with such complex situations.

[0003] In social event detection, graph neural networks classify and analyze nodes in social networks through learning and reasoning, which helps to accurately identify key nodes related to social events, and then deeply analyze the nature, scope of influence and development trend of events. However, one limitation of the graph neural network method is that it mainly relies on messages that share exactly the same attributes to connect, and fails to establish effective connections for messages with similar semantics but inconsistent attributes. In addition, manual labeling training and pre-determination of the number of predicted events are required, which to some extent limits its flexibility and efficiency in practical applications.

[0004] In this context, social event detection is solved by minimizing the graph structural entropy. While retaining the advantages of graph neural network-based methods, it does not require a priori given number of events and is unsupervised. In social event detection, structural entropy provides us with valuable perspectives and insights. Structural entropy has achieved remarkable results in a single type of node-to-node relationship graph, but in real-world networks, the situation is often much more complicated. Structural entropy assumes that there is only one type of relationship between nodes, but in fact, graphs are inherently multiple and contain heterogeneous relationship types. Many networks exhibit a multiplexed characteristic, that is, there may be multiple types of connections between nodes. This heterogeneity is crucial to revealing the structure of the graph, but it is ignored in traditional structural entropy metrics. Therefore, how to effectively process multi-relationship message graphs has become an urgent problem to be solved in current social event detection. Summary of the invention

[0005] The technical problem to be solved by the present invention is how to provide a social event detection method based on hierarchical multi-relational structural entropy minimization that can decode structural information more effectively and improve social event detection performance.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a social event detection method based on hierarchical multi-relation structure entropy minimization, comprising the following steps:

[0007] S1, integrates the BA graph generator and the large language model LLM to realize the construction and evolution of single-relation message graph to multi-relation message graph;

[0008] S2, based on the large language model LLM to guide the structural refinement and enhancement of multi-relational message graphs;

[0009] S3, based on the hierarchical two-dimensional multi-relation structural entropy minimization algorithm, decodes the basic structure of the multi-relation message graph;

[0010] S4, based on the decoding results, obtains social event detection based on hierarchical two-dimensional multi-relational structural entropy minimization, and outputs multiple message graph partitions, each partition representing a social event.

[0011] A further technical solution is that the step S1 specifically includes:

[0012] S11, using the BA graph generator of the PyG library in PyTorch, creates a single-relation message graph of the same size for each relation type;

[0013] S12, generate text attributes for each node using a large language model LLM to ensure that the text attributes can reflect the role or characteristics of the node in the graph and strengthen the semantic features;

[0014] S13, uses a large language model LLM to convert node text attributes into embedded vectors to provide input features for subsequent graph neural networks;

[0015] S14, by identifying shared nodes in the single-relationship message graph and using them as connection points, different relationship graphs are merged along the relationship axis to form a multi-relationship message graph.

[0016] A further technical solution is that step S2 specifically includes:

[0017] S21, uses a large language model LLM to analyze the strength and relevance of relationships between nodes, and adjusts the edge weights in the multi-relationship message graph based on these analysis results to optimize the representation of the graph;

[0018] S22, using the node text similarity information provided by the large language model LLM, accurately add and delete edges in the multi-relation message graph to strengthen the correlation representation between nodes;

[0019] S23, obtains the response of the possible category of the node from the large language model LLM, and uses it as a pseudo label to guide the graph neural network to learn more appropriate edge weights and improve the graph structure;

[0020] S24, uses pseudo-label propagation as a regularization method to promote close connections between nodes of the same category and promote the effective separation of nodes of different categories.

[0021] A further technical solution is that the hierarchical two-dimensional multi-relational structural entropy minimization algorithm in step S3 includes the following steps:

[0022] Input: multi-relation message graph G'=(V, E', R), where V is a node set, E' is a heterogeneous edge set, and R is a relationship set; Output: an optimal coding tree T with a height of 2;

[0023] Initialization: For each node V in the multi-relational message graph G', add two nodes α and α to the encoding tree T - , where α is a leaf node of T, and T α Contains only nodes V, T α is a subtree of α, - is the parent node of α, and α - The height h(α - ) is 1;

[0024] Iterative process: select all nodes with height 1 from the optimal coding tree T to form a set P, and continuously remove the first min clusters from the set P to form a set {Pg}. Each cluster Pg contains a set of nodes. For each cluster Pg in {Pg}, min is n, the number of remaining parts in P;

[0025] Perform the following operations: enter an internal infinite loop to optimize the encoding inside the cluster Pg; initialize the multi-relational structural entropy MSE change ΔMSE to infinity; traverse each pair of nodes α in the cluster Pg i and α j , j>i, check if there is an edge between them, if there is an edge, calculate the merge α i and α j The change in MSE of the multi-relational structural entropy after ij ; If ΔMSE ij If it is smaller than the current ΔMSE, update ΔMSE and candidate merge node α o1 and α o2 ;

[0026] If ΔMSE is less than 0, the MERGE operation merges α o1 and α o2 ; If ΔMSE is not less than 0, or other exit conditions are met, then jump out of the inner loop; if a cluster Pg has only one node left after optimization, then jump out of the outer loop;

[0027] If the set P does not change in two consecutive iterations, the subgraph size n is doubled to consider larger clusters for optimization, the loop ends, and the optimized coding tree T is returned.

[0028] The beneficial effects of adopting the above technical scheme are: in the method described in the present invention, the multi-relationship structural entropy MSE expands the favorable characteristics of the structural entropy SE, solves the heterogeneous relationship types, and serves as an improved tool for measuring and interpreting complex multi-relationship message graph structures; a hierarchical two-dimensional multi-relationship structural entropy minimization algorithm is introduced, which can detect communities in multi-relationship message graphs. Compared with the structural entropy SE, the multi-relationship structural entropy MSE provides a deeper interpretation of the structure of the multi-relationship message graph. When the multi-relationship structural entropy is used to minimize the graph entropy, a greater reduction can be observed, the decoding of structural information is more effective, and the performance of social event detection is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0030] Figure 1 is a main flow chart of the method described in the embodiment of the present invention;

[0031] Figure 2 is a flowchart of step S1 in the method according to an embodiment of the present invention;

[0032] Figure 3 is a flowchart of step S2 in the method according to an embodiment of the present invention;

[0033] Figure 4 It is a schematic diagram of the basic structure of decoding a multi-relation message graph using hierarchical multi-relation structure entropy minimization in an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0036] like Figure 1 As shown, the embodiment of the present invention discloses a social event detection method based on hierarchical multi-relation structure entropy minimization, comprising the following steps:

[0037] S1, integrates the BA graph generator and the large language model LLM to realize the construction and evolution of single-relation message graph to multi-relation message graph;

[0038] S2, based on the large language model LLM to guide the structural refinement and enhancement of multi-relational message graphs;

[0039] S3, based on the hierarchical two-dimensional multi-relation structural entropy minimization algorithm, decodes the basic structure of the multi-relation message graph;

[0040] S4, based on the decoding results, obtains social event detection based on hierarchical two-dimensional multi-relational structural entropy minimization, and outputs multiple message graph partitions, each partition representing a social event.

[0041] Further, such as Figure 2 As shown, step S1 specifically includes the following steps:

[0042] S11, using the BA graph generator of the PyG library in PyTorch, creates a single-relation message graph of the same size for each relation type;

[0043] S12, using a large language model (LLM) to generate text attributes such as title, summary or description for each node, ensuring that the text attributes can reflect the role or characteristics of the node in the graph and strengthen the semantic features;

[0044] S13, uses a large language model LLM to convert node text attributes into embedded vectors to provide input features for subsequent graph neural networks;

[0045] S14, by identifying shared elements (nodes) in a single-relationship message graph and using them as connection points, different relationship graphs are merged along the relationship axis to form a multi-relationship message graph.

[0046] Further, such as Figure 3 As shown, the step S2 guides the structural refinement and enhancement of the multi-relation message graph based on the large language model LLM, and specifically includes the following steps:

[0047] S21, uses the large language model LLM to analyze the strength and relevance of the relationships between nodes, and adjusts the edge weights in the multi-relationship message graph based on these analysis results to optimize the representation of the graph;

[0048] S22, using the node text similarity information provided by the large language model LLM, accurately add and delete edges in the multi-relation message graph to strengthen the correlation representation between nodes;

[0049] S23, obtains the response of the possible category of the node from the large language model LLM, and uses it as a pseudo label to guide the graph neural network to learn more appropriate edge weights and improve the graph structure;

[0050] S24, uses pseudo-label propagation as a regularization method to promote close connections between nodes of the same category and promote the effective separation of nodes of different categories.

[0051] As an optimization solution for the above example, according to Figure 4 As shown in Figure 2, the degree heterogeneity of the single-relation message graph G is similar to the stationary probability vector generated by random surfing on the single-relation message graph G. Based on random surfing on the multi-relation message graph, the multi-relation structural entropy MSE is established, including the following steps:

[0052] The multi-relation message graph is represented as G′=(V, E′, R), where V is the node set, E′ is the heterogeneous edge set, and R is the relationship set. The adjacency matrix of the multi-relation message graph G′ is Item A′ i,j,R is equal to the weight of the edge associated with the relation r∈R starting from j∈V and ending at i∈V. Two transfer matrices, denoted as V and R, are used to simulate the selection of neighbor nodes and relations. Assume that A′ is irreducible, that is, for any fixed r, a slice of A′, is irreducible, V and R are constructed as set up and are two probability distributions. After walking on G′ for an infinite time, x′ and y will converge to two distributions that satisfy V x′y =X′ and R x′x′ =Stationary distribution of y.

[0053] Given a multi-relational message graph G′ and a coding tree T, assuming that the steady-state distributions x′ and y of nodes and relations have been obtained from multi-relational random walks on the multi-relational message graph G′, which are calculated according to the node and relationship transfer matrices V and R. The multi-relational structural entropy (MSE) of G′ relative to T is in the following form:

[0054]

[0055] Among them, p′→α represents the conditional probability of transferring from a node i to a node j belonging to a subset of node α during the multi-relational random walk process, and p′ α is the steady-state distribution probability of node α, indicating the probability that the random surfer stays at node α in the long-term random walk, p′ α- is the steady-state distribution probability of the parent node of node α, α - Represents the parent node of node α.

[0056] Furthermore, in step S3, the basic structure of the multi-relation message graph is decoded based on the hierarchical two-dimensional multi-relation structural entropy minimization algorithm, and the basic structure of the multi-relation message graph is revealed by minimizing the proposed multi-relation structural entropy MSE. The MERGE operation is defined to merge two non-root nodes α in the coding tree T. o1 and α o2 , thereby optimizing the tree structure and reducing the multi-relationship structural entropy. The steps of the MERGE operation include:

[0057] Delete node α from the encoding tree T o1 and α o2 ;

[0058] Add a new node α to T n ;

[0059] New node α n The child node is the original α o1 and α o2 The combination of child nodes originally belonged to α o1 and α o2 All child nodes of will now become α n The child nodes of

[0060] Setting α n The parent node of is λ;

[0061] The merging operation changes the structure of the coding tree T, which causes the relevant MSE values ​​to change as follows:

[0062]

[0063] Furthermore, in step S3, the basic structure of the multi-relational message graph is decoded based on a hierarchical two-dimensional multi-relational structural entropy minimization algorithm, which aims to reveal the community structure in the multi-relational message graph G′ by optimizing the coding tree T;

[0064] In the initial stage, assuming that the coding tree T has no predefined communities, each node v∈V is assigned to its respective category;

[0065] Use a greedy algorithm to repeatedly merge two tree nodes in T to minimize the two-dimensional multi-relational structural entropy MSE until the merge operation no longer leads to a decrease in the multi-relational structural entropy MSE (i.e., ΔMSE < 0);

[0066] The set P of tree nodes of height 1 forms a partition of the node set V.

[0067] The present invention proposes a method for minimizing the entropy of a two-dimensional multi-relation structure in a hierarchical manner, integrating hierarchical graph partitioning to speed up the algorithm to decode and segment the message graph. The algorithm takes O(n 3 ), n is a hyperparameter and can be set to <<|V|.

[0068] In general, using hierarchical graph partitioning to accelerate the minimization of two-dimensional multi-relational structural entropy includes the following steps;

[0069] Hierarchical processing: divide the multi-relation message graph G′ into multiple subgraphs of size n, and minimize the MSE for each subgraph separately;

[0070] Iterative merging: After minimizing the MSE values ​​of all subgraphs, continue the optimization by using the clusters formed in the last iteration as nodes to be merged in subsequent iterations;

[0071] If at some point it is found that there are no nodes in the current subgraph that can be merged, the parameter n is appropriately increased to include more nodes for the next iteration;

[0072] When the algorithm considers all nodes, the MSE minimization process for the entire graph G′ is completed, and the algorithm terminates.

[0073] Specifically, the hierarchical two-dimensional multi-relational structural entropy minimization algorithm in step S3 includes the following steps:

[0074] Input: multi-relation message graph G'=(V, E', R), where V is a node set, E' is a heterogeneous edge set, and R is a relationship set; Output: an optimal coding tree T with a height of 2;

[0075] Initialization: For each node V in the multi-relational message graph G', add two nodes α and α to the encoding tree T - , where α is a leaf node of the coding tree T, and T α (a subtree) contains only nodes V, α - is the parent node of α, and α - The height h(α - ) is 1;

[0076] Iterative process: select all nodes with height 1 from the coding tree T to form a set P, and continuously remove the first min(n, the number of remaining parts in P) clusters from P to form a set {Pg}, where each cluster Pg contains a set of nodes, for each cluster Pg in {Pg};

[0077] Perform the following operations: enter an internal infinite loop to optimize the encoding inside cluster Pg; initialize the MSE change ΔMSE to infinity; traverse each pair of nodes α in cluster Pg i and α j (j>i), check if there is an edge between them, if there is an edge, calculate the merge α i and α j The MSE change after ΔMSE ij ; If ΔMSE ij If it is smaller than the current ΔMSE, update ΔMSE and candidate merge node α o1 and α o2 ;

[0078] If ΔMSE is less than 0, the MERGE operation merges α o1 and α o2 ; If ΔMSE is not less than 0, or other exit conditions are met, jump out of the inner loop; if a cluster Pg has only one node left after optimization, jump out of the outer loop;

[0079] If the set P does not change in two consecutive iterations, the subgraph size n is doubled to consider larger clusters for optimization. The loop ends and the optimized encoding tree T is returned. The set P of tree nodes with a height of 1 forms a partition of V, each partition representing a social event.

[0080] The present invention integrates the BA graph generator and the large language model LLM to realize the construction and evolution of single-relationship to multi-relationship message graphs. Based on the large language model LLM, the structural refinement and enhancement of the multi-relationship message graph are guided, and multiple relationships are effectively integrated into a unified network structure. The generated multi-relationship message graph is more accurate in semantic expression, with richer semantic information and deep relationship patterns.

[0081] As the first method for detecting social events in a multi-relationship message graph, the multi-relationship structural entropy MSE of the present invention fully considers heterogeneous relationships. The multi-relationship structural entropy MSE expands the traditional concept of structural entropy SE, and pays special attention to and solves the problem of heterogeneous relationship types inherent in the multi-relationship message graph. This enables MSE to more comprehensively reflect the complexity and diversity of the multi-relationship message graph.

[0082] In summary, the multi-relation structural entropy MSE not only inherits the favorable characteristics of structural entropy SE and solves the heterogeneous problem, but also provides a deeper and more accurate explanation of the structure of multi-relation message graphs by introducing a new measurement method. This helps to better understand and analyze complex social events.

[0083] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A social event detection method based on hierarchical multi-relational structure entropy minimization, characterized in that The steps include: S1, integrates the BA graph generator and the large language model LLM to realize the construction and evolution of single-relation message graph to multi-relation message graph; Step S1 specifically includes: S11, using the BA graph generator of the PyG library in PyTorch, creates a single-relation message graph of the same size for each relation type; S12, generating text attributes for each node using a large language model LLM, wherein the text attributes include a title, a summary, or a description, ensuring that the text attributes can reflect the role or characteristics of the node in the graph and strengthen the semantic features; S13, uses a large language model LLM to convert node text attributes into embedded vectors to provide input features for subsequent graph neural networks; S14, merging different relationship graphs along the relationship axis to form a multi-relationship message graph by identifying shared nodes in the single-relationship message graph and using them as connection points; S2, based on the large language model LLM to guide the structural refinement and enhancement of multi-relational message graphs; Step S2 specifically includes: S21, uses a large language model LLM to analyze the strength and relevance of relationships between nodes, and adjusts the edge weights in the multi-relationship message graph based on these analysis results to optimize the representation of the graph; S22, using the node text similarity information provided by the large language model LLM, accurately add and delete edges in the multi-relation message graph to strengthen the correlation representation between nodes; S23, obtains the response of the possible category of the node from the large language model LLM, and uses it as a pseudo label to guide the graph neural network to learn more appropriate edge weights and improve the graph structure; S24, using pseudo-label propagation as a regularization method to promote close connections between nodes of the same category and promote the effective separation of nodes of different categories; S3, based on the hierarchical two-dimensional multi-relation structural entropy minimization algorithm, decodes the basic structure of the multi-relation message graph; S4, based on the decoding results, obtain social event detection based on hierarchical two-dimensional multi-relation structure entropy minimization, and output multiple message graph partitions, each partition represents a social event; The method of establishing the multi-relation structural entropy MSE includes the following steps: The multi-relation message graph is represented as G′=(V, E′, R), where V is the node set, E′ is the heterogeneous edge set, and R is the relationship set. The adjacency matrix of the multi-relation message graph G′ is Item A i ' ,j,r is equal to the weight of the edge associated with the relation r∈R starting from j∈V and ending at i∈V. Two transfer matrices, denoted as V and R, are used to simulate the selection of neighbor nodes and relations. Assume that A′ is irreducible, that is, for any fixed r, a slice of A′, is irreducible, V and R are constructed as set up and are two probability distributions. After walking on G′ for an infinite time, x′ and y will converge to two distributions that satisfy V x′y =X′ and R x′x′ = stationary distribution of y; Given a multi-relational message graph G′ and a coding tree T, assume that the steady-state distributions x′ and y of nodes and relations have been obtained from the multi-relational random walk on the multi-relational message graph G′. These distributions are calculated according to the node and relationship transfer matrices V and R. Relative to the coding tree T, the multi-relational structural entropy MSE of the multi-relational message graph G′ is in the following form: Among them, p′→α represents the conditional probability of transferring from a node i to a node j belonging to a subset of node α during the multi-relational random walk process, and p′ α is the steady-state distribution probability of node α, indicating the probability that the random surfer stays at node α in the long-term random walk, is the steady-state distribution probability of the parent node of node α, α - Represents the parent node of node α; In step S3, hierarchical graph partitioning is used to accelerate the minimization of two-dimensional multi-relational structural entropy, and the steps include: Hierarchical processing: Divide the multi-relation message graph G′ into multiple subgraphs of size n, and minimize the multi-relation structural entropy MSE for each subgraph; Iterative merging: After minimizing the multi-relational structural entropy MSE values ​​of all subgraphs, continue the optimization by using the clusters formed in the last iteration as the nodes merged in the subsequent iterations; If at some point it is found that there are no nodes in the current subgraph that can be merged, the parameter n is appropriately increased to include more nodes for the next iteration; After all nodes are considered, the process of minimizing the multi-relation structural entropy MSE of the entire multi-relation message graph G′ is completed; The hierarchical two-dimensional multi-relation structural entropy minimization algorithm in step S3 includes the following steps: Input: multi-relation message graph G'=(V, E', R), where V is a node set, E' is a heterogeneous edge set, and R is a relationship set; Output: an optimal coding tree T with a height of 2; Initialization: For each node V in the multi-relational message graph G', add two nodes α and α to the encoding tree T - , where α is a leaf node of T, and T α Contains only nodes V, T α is a subtree of α, - is the parent node of α, and α - The height h(α - ) is 1; Iterative process: select all nodes with height 1 from the optimal coding tree T to form a set P, and continuously remove the first min clusters from the set P to form a set {Pg}. Each cluster Pg contains a set of nodes. For each cluster Pg in {Pg}, min is the minimum value of n and the number of remaining parts in P. Perform the following operations: Define the MERGE operation to merge two non-root nodes α in the encoding tree T o1 and α o2 , thereby optimizing the tree structure and reducing the multi-relational structural entropy MSE; Enter an internal infinite loop to optimize the encoding inside the cluster Pg; initialize the multi-relational structural entropy MSE change ΔMSE to infinity; traverse each pair of nodes α in the cluster Pg i and α j , j>i, check if there is an edge between them, if there is an edge, calculate the merge α i and α j The change in MSE of the multi-relational structural entropy after ij ; If ΔMSE ij If it is smaller than the current ΔMSE, update ΔMSE and candidate merge node α o1 and α o2 ; If ΔMSE is less than 0, the MERGE operation merges α o1 and α o2 ; If ΔMSE is not less than 0, or other exit conditions are met, then jump out of the inner loop; if a cluster Pg has only one node left after optimization, then jump out of the outer loop; If the set P does not change in two consecutive iterations, the subgraph size n is doubled to consider larger clusters for optimization, the loop ends, and the optimized coding tree T is returned.

2. The social event detection method based on hierarchical multi-relational structure entropy minimization as claimed in claim 1, characterized in that: The multi-relation message graph G′ used must have strong connectivity, and A′ is irreducible. If it is not satisfied, it needs to be adjusted, including the following steps: When A′ is reducible, the multi-relational message graph G′ needs to be adjusted to ensure that x′ and y converge, and V and R are randomly adjusted so that have To speed up convergence, primality adjustment is performed by replacing V with cV+(1-c)E′, where E′ is the transfer matrix; c is set to 0.85, and E′=1 / |V|*1, where 1 is an all-1 matrix of |V|×|V|×|R|.

3. The social event detection method based on hierarchical multi-relational structure entropy minimization as claimed in claim 1, characterized in that: The steps of the MERGE operation include: Delete node α from the encoding tree T o1 and α o2 ; Add a new node α to the encoding tree T n ; New node α n The child node is the original α o1 and α o2 The combination of child nodes originally belonged to α o1 and α o2 All child nodes of will now become α n The child nodes of Set new node α n The parent node of is λ; The merging operation changes the structure of the coding tree T, which causes the relevant MSE values ​​to change as follows:

4. The social event detection method based on hierarchical multi-relational structure entropy minimization as claimed in claim 1, characterized in that: Return the optimized coding tree T. The tree node set P with height 1 forms a partition of V, and each partition represents a social event.

Images