Public health monitoring information analysis method and system based on graph data, and storage medium

Abstract

The application provides a public health monitoring information analysis method and system based on graph data, a storage medium, and the like. By obtaining public health monitoring correlation information, a heterogeneous correlation graph is established, topological structure analysis is performed, a correlation path set is generated, subgraph pattern mining is performed based thereon, a repeatedly occurring entity relationship combination structure is extracted, a propagation correlation pattern is generated, the propagation correlation pattern is input into a risk transmission network construction process, the influence weight of each propagation correlation pattern is determined, a directed weighted risk transmission network is constructed in combination with a relationship chain structure, path prediction analysis is performed, a key propagation path from an initial risk entity to a target susceptible entity is identified, and a public health event propagation path prediction result is generated in combination with the correlation strength and relationship type of the entities on the key propagation path. The application can improve the practicality and decision support value of public health monitoring information analysis.

Description

Technical Field

[0001] This invention relates to the field of data processing, and more specifically, to a method, system, and storage medium for analyzing public health monitoring information based on graph data. Background Technology

[0002] In the field of public health surveillance, integrating multi-source surveillance data for information analysis can reveal the correlations between medical behaviors, population activities, and environmental factors to support the prediction of transmission pathways. Currently, most methods rely on entity relationship extraction to construct association networks and identify potential transmission links through path traversal and pattern mining. However, when dealing with entity relationships, the indirect influence relationships between medical behavior entities, population activity entities, and environmental factor entities are often insufficiently explored, resulting in limited integrity of the association network structure. In the identification of transmission patterns, there is a lack of effective capture of the dynamic characteristics of patterns over time, making it difficult to reflect the evolutionary characteristics of transmission patterns over time. These factors collectively lead to insufficient accuracy and interpretability of public health event transmission pathway prediction results, making it difficult to meet the needs of refined prevention and control decision-making. Summary of the Invention

[0003] This invention provides a method, system, and storage medium for analyzing public health monitoring information based on graph data.

[0004] In a first aspect, embodiments of the present invention provide a method for analyzing public health monitoring information based on graph data. The method includes: acquiring public health monitoring association information; establishing a heterogeneous association graph containing entities of medical behavior, population activity, and environmental factors through entity relationship extraction; wherein the node set of the heterogeneous association graph contains unique identifiers and attribute features of entities, and the edge set contains direct interaction relationships and indirect influence relationships between entities; performing topological structure parsing on the heterogeneous association graph, traversing the connectivity relationships between entity nodes, and generating an association path set containing node sequences, relationship type sequences, and path length information; wherein each path in the association path set represents a potential propagation link between different entities; and performing subgraph pattern mining based on the association path set to extract recurring combinations of entity relationships. The system generates a propagation association pattern with time interval markers, which includes core entity pairs and relationship chain structures connecting them. This propagation association pattern is then input into a risk transmission network construction process. The influence weight of each propagation association pattern is determined through node importance assessment. A directed weighted risk transmission network is constructed based on the relationship chain structure, where directed edges represent propagation directions and edge weights represent influence intensity. Path prediction analysis is performed on the risk transmission network to identify key propagation paths from initial risk entities to target susceptible entities. The propagation path prediction results for public health events are generated by combining the association strength and relationship type of entities along these key propagation paths. These prediction results include a path entity sequence, a relationship type sequence, and a cumulative influence intensity.

[0005] Secondly, embodiments of the present invention provide a computer system, comprising: a memory storing a computer program; and a processor for loading the computer program to implement the public health monitoring information analysis method based on graph data as described above.

[0006] Thirdly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the methods described above.

[0007] This invention establishes a heterogeneous association graph containing entities related to medical behaviors, population activities, and environmental factors through entity relationship extraction. Topological structure analysis of the heterogeneous association graph generates a set of association paths. Based on this set, subgraph pattern mining is performed to generate propagation association patterns with time interval labels. These patterns are then input into a risk transmission network construction process to build a directed weighted risk transmission network. Path prediction analysis is performed on this network to generate public health event propagation path prediction results. This invention comprehensively captures the complex relationships between entities related to medical behaviors, population activities, and environmental factors, improving the comprehensiveness of public health monitoring information analysis. By extracting recurring entity relationship combinations to generate propagation association patterns with time interval labels, it achieves... Effectively mining the transmission patterns of public health events enhances the reliability of transmission pattern identification and provides more targeted analytical basis for subsequent risk assessment. By inputting transmission association patterns into the risk transmission network construction process and determining the influence weight of each transmission association pattern through node importance assessment, the directed weighted risk transmission network constructed in conjunction with the relationship chain structure can accurately reflect the differences in transmission direction and influence intensity between different entities, improve the accuracy of risk transmission path analysis, and finally generate public health event transmission path prediction results that include path entity sequences, relationship type sequences, and cumulative influence intensity. This can provide clear and intuitive transmission path guidance for public health event monitoring, early warning, and prevention and control decisions, effectively enhancing the practicality and decision support value of public health monitoring information analysis. Attached Figure Description

[0008] Figure 1 This is a flowchart of a public health monitoring information analysis method based on graph data provided in an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of the composition of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0010] Please see Figure 1 The flowchart below illustrates a public health monitoring information analysis method based on graph data, provided by an embodiment of the present invention. This method can be executed by a computer system and may include the following steps:

[0011] Step S100: Obtain public health monitoring related information, and establish a heterogeneous association graph containing medical behavior entities, population activity entities and environmental factor entities through entity relationship extraction. The node set of the heterogeneous association graph contains the unique identifier and attribute characteristics of the entities, and the edge set contains the direct interaction relationship and indirect influence relationship between entities.

[0012] Public health surveillance-related information refers to the collection of various data related to medical behaviors, population activities, and environmental factors in the field of public health surveillance. This information may come from multiple different data sources, such as medical records from healthcare institutions, population mobility statistics, and monitoring data from environmental sensors. Medical behavior entities refer to specific behaviors or operations related to healthcare, such as diagnosis and treatment. Each medical behavior entity has its unique attributes, such as treatment-related characteristics, service recipient characteristics, and compliance with operational standards. Population activity entities represent various activities of the population, including characteristics such as activity scale, duration, and spatial coverage, such as large gatherings and daily commutes. Environmental factor entities refer to environmental factors that affect public health, such as air quality, water quality, and temperature. Their attributes include monitoring value fluctuation characteristics, spatiotemporal distribution characteristics, and contact characteristics with related populations.

[0013] In one implementation, step S100 may include the following steps S110-S160:

[0014] Step S110: Perform multi-source data block processing on public health monitoring-related information, dividing it into medical record data blocks, population flow data blocks, and environmental sensor data blocks according to the data source type. Each data block contains a timestamp sequence and data field description information.

[0015] Multi-source data chunking is the process of classifying and organizing public health surveillance information from different data sources. Since public health surveillance information may come from multiple different channels, such as hospital information systems, population flow statistics from transportation departments, and environmental monitoring sensors, the formats and content of this data may differ. Therefore, it is necessary to divide this data according to its source type, forming different data blocks.

[0016] The medical record data block primarily contains various medical records from medical institutions, such as medical records, diagnostic reports, and treatment records. These data record patients' basic information, diagnoses, and treatment processes, which are crucial for analyzing medical behavior entities. The population flow data block contains information on population movement at different times and spaces, such as traffic flow and population migration. This data reflects the characteristics of population activity entities. The environmental sensor data block consists of various environmental data collected by environmental sensors, such as air quality monitoring data and water quality monitoring data, used to describe the attribute characteristics of environmental factor entities. Each data block contains a timestamp sequence and data field description information. The data field description information provides detailed explanations of each field in the data block, including the field name, meaning, and data type, facilitating subsequent data processing and analysis. For example, in the medical record data block, the timestamp sequence records the time of a patient's visit, and the data field description information explains the specific medical information represented by each field, such as "diagnosis result" and "treatment method."

[0017] Step S120: Perform entity boundary recognition operation in each data block, extract entity candidate segments by scanning text data through a sliding window, determine the entity type of the candidate segments based on the named entity recognition model, and output the preliminary recognition results of medical behavior entities, crowd activity entities, and environmental factor entities.

[0018] Entity boundary identification is the process of accurately identifying the start and end positions of entities within data blocks. Since text data within data blocks may contain a large amount of information, appropriate methods are needed to determine entity boundaries. Sliding window scanning is a commonly used text processing technique that extracts text fragments within a fixed-size window as candidate entity fragments by sliding the window across the text data.

[0019] Named entity recognition (NER) is a machine learning or deep learning-based model used to determine the entity type of extracted entity candidate fragments. After training, this model can identify different types of entities, such as medical behavior entities, crowd activity entities, and environmental factor entities. In this embodiment, the NER model can employ an architecture combining a bidirectional long short-term memory network (Bi-LSTM) and a conditional random field (CRF). Bi-LSTM can effectively capture contextual information in the text sequence, while CRF is used to globally optimize entity labels, improving the accuracy of entity type determination.

[0020] Specifically, the extracted entity candidate fragments are input into the named entity recognition model. The model classifies each candidate fragment based on its learned features and rules, determining its entity type. For example, if a candidate fragment contains keywords like "blood routine test," the model classifies it as a medical behavior entity; if it contains "large concert," it classifies it as a crowd activity entity; and if it contains "air quality index," it classifies it as an environmental factor entity. Finally, the model outputs preliminary identification results for medical behavior entities, crowd activity entities, and environmental factor entities.

[0021] Step S130: Perform entity disambiguation processing on the preliminary identification results, calculate the attribute similarity between candidate entities of the same type, merge entities whose attribute similarity reaches the set standard, assign a globally unique identifier to the merged entity, and establish an entity-identifier mapping table.

[0022] In one implementation, step S130 may specifically include the following steps S131-S136.

[0023] Step S131: Extract the static attribute set and dynamic attribute set of candidate entities of the same type from the preliminary identification results. The static attribute set includes entity type label, unique identifier prefix and core attribute field, and the dynamic attribute set includes feature parameter sequence and timestamp information that change over time.

[0024] When performing entity disambiguation, the static attribute set and dynamic attribute set of candidate entities of the same type are first extracted from the preliminary identification results. The static attribute set is a collection of attributes that do not change over time. Among them, the entity type label is used to clarify the type of entity, such as medical behavior entity, crowd activity entity, or environmental factor entity; the unique identifier prefix is ​​a pre-defined part of the identifier to distinguish different types of entities, which can help to quickly identify the general category of the entity; the core attribute field is the key attribute that describes the essential characteristics of the entity. For example, the core attribute fields of a medical behavior entity may include diagnosis and treatment items, treatment effects, etc.

[0025] The dynamic attribute set contains a sequence of feature parameters and timestamp information that change over time. The feature parameter sequence records the changes in the feature values ​​of an entity at different points in time. For example, the fluctuation characteristics of the monitored values ​​of environmental factor entities can be represented by the feature parameter sequence.

[0026] Step S132: Perform a structured comparison of the static attribute set, assign weight coefficients according to the importance level of the attribute fields. The importance level is determined based on the entity type and the attribute's contribution to the differentiation in historical disambiguation cases. Calculate the weighted attribute matching score, which is the sum of the products of the matching value of each field and the corresponding weight coefficient.

[0027] Structured comparison is a process of detailed comparative analysis of static attribute sets. Since the importance of each attribute field in the static attribute set in distinguishing entities may vary, it is necessary to assign a corresponding weight coefficient to each attribute field. The determination of importance levels is based on the entity type and the attribute's distinguishing contribution in historical disambiguation cases. For example, in a medical behavior entity, the attribute field "treatment items" may distinguish different medical behaviors more effectively than "doctor's name," therefore, "treatment items" may have a higher importance level and be assigned a larger weight coefficient.

[0028] After determining the weighting coefficients, the matching value for each attribute field needs to be calculated. The matching value represents the similarity between two candidate entities in that attribute field, and can be calculated using a string matching algorithm (such as the edit distance algorithm). Then, the matching value for each field is multiplied by its corresponding weighting coefficient, and finally all products are summed to obtain the weighted attribute matching score. This score comprehensively considers the importance and matching situation of each attribute field, and can more accurately reflect the similarity between two candidate entities in terms of static attributes.

[0029] Step S133: Perform time series similarity analysis on the dynamic attribute set, calculate the morphological similarity of the feature parameter sequence, extract the trend features, periodic features and mutation point distribution features of the sequence, and generate dynamic trend similarity through multi-feature fusion. The dynamic trend similarity comprehensively reflects the degree of similarity between the overall morphology and local features of the sequence.

[0030] Time series similarity analysis is the process of comparing feature parameter sequences in a dynamic set of attributes. Since these feature parameter sequences change over time, it is necessary to analyze their morphological similarity. Morphological similarity is an indicator that measures the degree of similarity in shape between two feature parameter sequences.

[0031] Trend features reflect the overall trend of a sequence, such as whether it is an upward trend, a downward trend, or a stable trend; periodic features indicate whether the sequence has a periodic pattern of change; and the distribution features of abrupt change points record the location and magnitude of sudden changes in the sequence.

[0032] Multi-feature fusion is a process of comprehensively processing extracted trend features, periodic features, and abrupt change distribution features. By fusing these features, dynamic trend similarity can be generated. Dynamic trend similarity considers not only the overall pattern of the sequence but also the degree of similarity of local features. For example, air quality monitoring data sequences of two environmental entities may have similar overall trends, but abrupt changes may exist in certain time periods. If only the overall trend is considered, these local differences may be overlooked. However, dynamic trend similarity, through multi-feature fusion, can more accurately reflect the similarity between the two sequences.

[0033] Step S134: Construct a static-dynamic fusion model. Input the weighted attribute matching score and dynamic trend similarity into the fusion model. Map the two to the same feature space through nonlinear transformation to generate a standardized similarity score. The standardized similarity score eliminates the difference in dimensions and unifies them to a set numerical range.

[0034] The static-dynamic fusion model is a model that comprehensively considers the similarity of static and dynamic attributes. Since weighted attribute matching scores and dynamic trend similarity measure the similarity of candidate entities from different perspectives, their units of measurement and numerical ranges may differ. Therefore, these two metrics need to be input into the fusion model and mapped to the same feature space through a nonlinear transformation.

[0035] Nonlinear transformations can be implemented using neural network models, such as multilayer perceptrons (MLPs). A multilayer perceptron consists of an input layer, hidden layers, and an output layer. The input layer receives weighted attribute matching scores and dynamic trend similarity as inputs. The hidden layers perform nonlinear transformations on the inputs, and the output layer outputs standardized similarity scores.

[0036] Standardized similarity scores eliminate dimensional differences and unify them into a set numerical range, such as [0,1]. This facilitates subsequent comparisons and judgments. For example, if the standardized similarity scores of two candidate entities are close to 1, it means that they are very similar in both static and dynamic attributes; if the scores are close to 0, it means that their similarity is low.

[0037] Step S135: Introduce entity association context information, calculate the common neighbor ratio and shortest path distance of candidate entity pairs in the heterogeneous association graph. The common neighbor ratio is the ratio of the number of shared neighbor nodes to the total number of neighbor nodes, and the shortest path distance is the number of edges contained in the shortest path connecting two entities. Generate context association degree based on the common neighbor ratio and shortest path distance.

[0038] Entity association context information refers to the surrounding environment and relationships of a candidate entity in a heterogeneous association graph. In a heterogeneous association graph, each entity is connected to other entities through edges; these connected entities are the entity's neighbor nodes. The proportion of common neighbors and the shortest path distance are two important indicators for measuring the degree of association between candidate entity pairs.

[0039] The common neighbor ratio is obtained by calculating the ratio of the number of shared neighbor nodes between two candidate entities to their total number of neighbor nodes. If two candidate entities have a large number of shared neighbor nodes, it indicates that they have a strong correlation in the heterogeneous association graph.

[0040] The shorter the shortest path distance, the stronger the connection between two entities. For example, in a heterogeneous association graph, a short short path distance between two environmental factor entities may indicate a strong interaction in terms of their environmental impact. Contextual relevance can be generated based on the proportion of common neighbors and the shortest path distance. Contextual relevance comprehensively considers the local structure and connectivity of candidate entities in the heterogeneous association graph, providing a more complete reflection of the degree of association between them. Contextual relevance can be calculated using a linear combination, for example, Contextual Relevance = α * Proportion of Common Neighbors + β * (1 / Shortest Path Distance), where α and β are weighting coefficients that are adjusted according to the specific circumstances.

[0041] Step S136: Determine the final attribute similarity by combining the standardized similarity score and the contextual relevance. When the final attribute similarity reaches the set standard, merge the corresponding candidate entities, retain the record with the highest completeness in each attribute field as the attribute information of the merged entity, and update the node connection relationship of the entity-identifier mapping table and the heterogeneous association graph.

[0042] The standardized similarity score reflects the similarity of candidate entities in terms of static and dynamic attributes, while the contextual relevance reflects their association in the heterogeneous association graph. Combining these two indicators allows for a more accurate determination of whether two candidate entities should be merged. The final attribute similarity can be calculated using a weighted summation method, for example, final attribute similarity = γ * standardized similarity score + (1-γ) * contextual relevance, where γ is a weighting coefficient that can be adjusted according to the actual situation. When the final attribute similarity reaches a set standard, it indicates that the two candidate entities are very similar and can be merged into one entity. During the merging process, the record with the highest completeness in each attribute field needs to be retained as the attribute information of the merged entity. This is to ensure that the attribute information of the merged entity is more accurate and comprehensive. Simultaneously, the entity-identifier mapping table needs to be updated to associate the merged entity with a new globally unique identifier. Furthermore, the node connection relationships in the heterogeneous association graph need to be updated, connecting all edges originally connected to the two candidate entities to the merged entity to ensure the structural correctness of the heterogeneous association graph. For example, after merging two medical behavior entities, the patient nodes that were originally connected to the two entities are now all connected to the merged medical behavior entity.

[0043] Step S140: Extract the dynamic attribute features of each entity. The dynamic attribute features of the medical behavior entity include diagnosis and treatment related features, service object features, and operation standard compliance features. The dynamic attribute features of the population activity entity include activity scale features, duration features, and spatial coverage features. The dynamic attribute features of the environmental factor entity include monitoring value fluctuation features, spatiotemporal distribution features, and related population contact features.

[0044] Dynamic attribute features are characteristic information that describes how an entity changes over time. For different types of entities, their dynamic attribute features have different meanings and forms of expression.

[0045] The dynamic attributes of a medical practice entity mainly include diagnosis-related characteristics, service recipient characteristics, and operational compliance characteristics. Diagnosis-related characteristics reflect the specific content and effects of the medical practice, such as diagnostic accuracy and treatment success rate. Service recipient characteristics describe the characteristics of the population receiving medical services, such as the age distribution of patients and the severity of their conditions. Operational compliance characteristics are used to assess whether the medical practice conforms to relevant operational norms and standards, such as the compliance of surgical procedures and the rationality of drug use.

[0046] The dynamic attributes of crowd activity entities include activity scale, duration, and spatial coverage. Activity scale can be measured by the number of participants, reflecting the magnitude of the activity's impact. Duration records the length of time from start to finish, which is crucial for analyzing the temporal distribution and scope of influence of crowd activities. Spatial coverage describes the geographical area involved in the activity, such as the location of a large concert and its surrounding area of ​​influence.

[0047] The dynamic attributes of environmental factors include the fluctuation characteristics of monitored values, spatiotemporal distribution characteristics, and contact characteristics with related populations. The fluctuation characteristics of monitored values ​​reflect the changes in environmental factor monitoring data over time, such as the fluctuation range of the air quality index. The spatiotemporal distribution characteristics describe the distribution patterns of environmental factors in different times and spaces, such as the seasonal changes in water pollution in different regions. The contact characteristics with related populations indicate the frequency and extent of contact between environmental factors and the population, such as the impact of air pollution in a certain area on the health of surrounding residents.

[0048] Step S150: Identify the direct interaction relationships between entities, calculate the co-occurrence frequency of different types of entities within the same time window through entity co-occurrence analysis, determine the action subject and object relationship of entity pairs by combining semantic role labeling, and generate a set of direct interaction relationships.

[0049] Entity co-occurrence analysis (OCA) is a method that determines whether there is a relationship between different types of entities by statistically analyzing the frequency of their simultaneous occurrence within the same time window. The time window can be determined according to specific analytical needs, such as a day, a week, or a month. If two entities frequently co-occur within the same time window, there may be a direct interaction relationship between them. For example, in the medical field, if a doctor and a patient appear together multiple times in medical records within the same time period, it can be inferred that there is a direct interaction relationship of medical services between them. By semantically labeling text containing entities, the subject-object relationship of the action between entity pairs can be clarified. For example, in the sentence "The doctor treats the patient," "doctor" is the subject of the action, and "patient" is the object of the action. Combining the results of entity co-occurrence frequency and semantic role labeling, a set of direct interaction relationships can be generated. This set records all identified direct interaction relationships, including the subject, object, and type of the relationship. For example, the set of direct interaction relationships can record a relationship such as "doctor-patient-medical service," indicating that the doctor provides medical services to the patient.

[0050] Step S160: Mine indirect influence relationships between entities, construct an attribute association matrix based on entity attribute features, extract potential association dimensions through matrix singular value decomposition, mark entity pairs whose association dimension matching degree reaches the set standard as indirect influence relationships, and integrate direct and indirect influence relationships to construct a heterogeneous association graph.

[0051] Indirect influence relationships between entities refer to the influence between two entities through other intermediate entities or factors. Uncovering these indirect influence relationships requires starting with the attribute characteristics of the entities and constructing an attribute association matrix. An attribute association matrix is ​​a two-dimensional matrix where rows and columns represent different entities, and elements in the matrix represent the degree of association between the attribute characteristics of two entities. For example, the matrix elements can be filled by calculating the similarity between the attribute characteristics of two entities. If the attribute characteristics of two entities are highly similar, the corresponding element value in the matrix will also be larger. Latent association dimensions reflect hidden relationships between entities that may not be directly observable from attribute characteristics. Through singular value decomposition (SVD), the high-dimensional attribute association matrix can be reduced in dimensionality, extracting the most important association dimensions.

[0052] Association dimension matching degree is an indicator that measures the similarity between two entities in potential association dimensions. When the association dimension matching degree reaches a set standard, it indicates that there is an indirect influence relationship between the two entities. For example, when analyzing the relationship between environmental factor entities and medical behavior entities, it may be found that changes in certain environmental factors can indirectly affect the frequency of medical behaviors by influencing the health status of the population. Finally, the discovered indirect influence relationships are integrated with the previously identified direct relationships to construct a heterogeneous association graph. In the heterogeneous association graph, both direct and indirect influence relationships are represented by edges, and nodes represent different entities. In this way, the heterogeneous association graph comprehensively reflects the various relationships between entities, providing a foundation for subsequent analysis and applications.

[0053] Step S200: Perform topological structure parsing on the heterogeneous association graph, traverse the connectivity relationships between entity nodes, and generate an association path set containing node sequence, relationship type sequence, and path length information. Each path in the association path set represents a potential propagation link between different entities.

[0054] In one implementation, step S200 may specifically include the following steps S210-S260.

[0055] Step S210: Obtain the multi-level index structure of the pre-constructed heterogeneous association graph. The first level index is the entity type index, which divides the nodes into groups according to the entity type; the second level index is the relation type index, which divides the edges into groups according to the relation type; and the third level index is the attribute feature index, which divides the node subsets according to the range of entity attribute feature values.

[0056] A multi-level index structure is a pre-built data structure designed to improve the query and traversal efficiency of heterogeneous graphs. It categorizes and organizes nodes and edges in a heterogeneous graph according to different dimensions, facilitating quick location and access to the required information.

[0057] The first level of index is the entity type index, which groups nodes according to entity type, such as medical behavior entities, population activity entities, and environmental factor entities. This allows queries to directly locate nodes of a specific type, reducing unnecessary traversal. The second level of index is the relation type index, which groups edges according to relation type, such as "medical services" and "population flow." The relation type index allows for quick identification of edges with specific relation types, facilitating the analysis of specific relationships between entities. The third level of index is the attribute feature index, which divides nodes into different subsets based on the range of entity attribute feature values. For example, for environmental factor entities, nodes can be divided into different subsets based on the range of air quality index values. This allows queries to quickly filter nodes that meet the requirements based on attribute feature conditions.

[0058] A multi-level index structure can significantly improve the efficiency of querying and traversing heterogeneous relational graphs. For example, when searching for all paths related to a specific medical behavior entity, one can first locate the medical behavior entity node using the entity type index, and then further filter out edges and nodes that meet the criteria using the relation type index and attribute feature index, thereby quickly finding the required path information.

[0059] Step S220: Initialize the path exploration queue, add each entity node in the heterogeneous association graph as a starting node to the queue, and assign a path exploration context to each starting node. The context includes the set of visited nodes, the current path length, and a list of relationship type constraints.

[0060] A path exploration queue is a queue structure used to store paths to be explored. When performing topology analysis on a heterogeneous graph, path exploration needs to start from each entity node; therefore, each entity node in the heterogeneous graph is added to the path exploration queue as a starting node.

[0061] The path exploration context is a set of information assigned to each starting node, used to record the current state of path exploration. The visited node set records nodes already visited during the path exploration process, avoiding duplicate visits. The current path length records the number of edges in the path from the starting node to the current node. The relation type constraint list is a pre-defined list of relation types used to restrict the types of relations that can be used during path exploration. For example, in public health transmission analysis, only relation types related to disease transmission, such as "infection transmission" and "contact infection," can be considered. By assigning a path exploration context to each starting node, the independence and accuracy of the path exploration process can be ensured. Each starting node has its own context information, allowing for judgments and decisions during exploration based on this context, avoiding interference between different paths.

[0062] Step S230: Perform a hierarchical path traversal operation, starting from the starting node, and traverse the adjacent nodes corresponding to the relationship types according to the preset priority order of public health transmission characteristics based on the relationship type index.

[0063] In one implementation, step S230 may specifically include the following steps S231-S236.

[0064] Step S231: Read the preset relationship priority sorting table from the relationship type index. The priority sorting table is determined according to the strength of the relationship in the public health transmission chain. Transmission media relationship takes precedence over exposure contact relationship, and exposure contact relationship takes precedence over service provision relationship.

[0065] The pre-defined relationship priority ranking table is a table pre-set based on the strength of the relationship in the public health transmission chain. This table records the priority order of different relationship types and is used to guide the selection of nodes during path traversal.

[0066] Transmission vector relationships refer to disease transmission through a medium (such as air, water, or food). Because transmission vectors can spread diseases over a wide area and have strong transmissibility, they are usually prioritized higher in the priority list. Exposure-contact relationships refer to disease transmission caused by direct or indirect contact between people; their transmission range is relatively small, and they have a lower priority. Service provision relationships refer to non-transmissionary relationships such as medical and educational services; their role in the public health transmission chain is relatively weak, and therefore they have a lower priority.

[0067] By reading a predefined relationship priority sorting table from the relationship type index, it can be ensured that adjacent nodes are visited in the correct priority order during path traversal. For example, when exploring disease transmission paths, adjacent nodes connected by transmission medium relationships are given priority, followed by nodes connected by exposure contact relationships and service provision relationships.

[0068] Step S232: Obtain the list of neighboring nodes of the current node. The list of neighboring nodes includes nodes connected by different relationship types and their corresponding relationship attributes. The relationship attributes include the timestamp of the relationship occurrence, the duration, and the scope of influence.

[0069] The neighboring node list refers to the list of all nodes directly connected to the current node. During path traversal, it is necessary to obtain the neighboring node list of the current node in order to further explore the path.

[0070] The list of neighboring nodes not only contains information about the neighboring nodes but also includes relationship attributes connecting the current node and its neighbors. Relationship attributes record important information about the relationship, such as the timestamp of the relationship's occurrence, duration, and scope of influence. The timestamp records the specific time the relationship occurred, which is crucial for analyzing the temporal sequence of disease transmission. The duration records the length of time the relationship lasts, such as the duration of a single contact infection. The scope of influence describes the area affected by the relationship; for example, the scope of an airborne transmission event might include a specific area of ​​a city.

[0071] Step S233: Sort the list of adjacent nodes according to the relationship priority sorting table, and put the adjacent nodes corresponding to the relationship types with higher priority at the top. Adjacent nodes with the same priority are sorted according to the duration of the relationship.

[0072] First, based on the relationship priority sorting table, adjacent nodes corresponding to relationships with higher priority are listed first. This ensures that adjacent nodes more likely to spread disease are visited first during traversal. For example, if the transmission vector relationship has a higher priority than the exposure-contact relationship, then adjacent nodes connected by the transmission vector relationship will be listed before those connected by the exposure-contact relationship. For adjacent nodes of the same priority, they are sorted according to the duration of the relationship. The longer the relationship duration, the greater the likelihood of disease transmission; therefore, adjacent nodes with longer relationship durations are listed first. For example, among adjacent nodes connected by the exposure-contact relationship, nodes with longer contact durations are listed first. By sorting the list of adjacent nodes, the efficiency and accuracy of path traversal are improved, ensuring that more valuable paths are explored first.

[0073] Step S234: Select the sorted adjacent nodes as the next hop nodes in sequence, check whether the next hop node is in the set of visited nodes, if not, add it to the set of visited nodes, and update the current path length and relationship type sequence.

[0074] After sorting the list of adjacent nodes, each of the sorted adjacent nodes needs to be selected as the next hop node for path exploration. When selecting the next hop node, it is necessary to check whether the node is already in the set of visited nodes.

[0075] The visited node set is a collection recorded in the path exploration context to avoid revisiting nodes that have already been visited. If the next-hop node is not in the visited node set, it means that the node is a new node and can be added to the visited node set. Simultaneously, the current path length and relation type sequence need to be updated. The current path length needs to be incremented by 1, indicating that an edge has been added to the path. The relation type sequence needs to record the relation types connecting the current node and the next-hop node for subsequent analysis of the path's relation type characteristics.

[0076] For example, in path exploration starting from a medical behavior entity node, if the next hop node is a crowd activity entity node and that node is not in the set of visited nodes, then the crowd activity entity node is added to the set of visited nodes, the current path length is updated to the original length plus 1, and the relationship type connecting the two nodes is recorded in the relationship type sequence, such as "medical service".

[0077] Step S235: If there are unvisited nodes in the list of adjacent nodes, continue the traversal operation; when all adjacent nodes have been visited or the path length limit has been reached, store the current path information in a temporary path buffer.

[0078] During path traversal, the decision to continue depends on the visited status of adjacent nodes and the path length. If there are unvisited nodes in the adjacent node list, it means there are still paths to explore, so the traversal continues, selecting the next unvisited adjacent node as the next hop. If all adjacent nodes have been visited, or the path length limit has been reached, the exploration of the current path has ended. The path length limit is a pre-set maximum path length used to control the scope of path exploration and avoid infinite loops. When this occurs, the current path information is stored in a temporary path buffer. The temporary path buffer is a buffer used to temporarily store path information for subsequent processing and analysis.

[0079] Step S236: Perform conflict detection on the paths in the temporary path buffer. If there are path conflicts with the same node sequence but different relationship type sequences, retain the path with higher relationship type continuity, remove the conflicting path, and add the valid path to the path exploration queue.

[0080] Conflict detection is the process of examining paths in a temporary path buffer to identify and resolve conflicting paths. A path conflict refers to a path that has the same node sequence but different relation type sequences.

[0081] Relation type continuity refers to the degree of semantic association between adjacent relation types in a path. A higher degree of semantic association between adjacent relation types indicates better relation type continuity in the path. When handling path conflicts, paths with higher relation type continuity are retained because such paths are more logical and consistent with reality.

[0082] After removing conflicting paths, valid paths are added to the path exploration queue. Valid paths are those that have passed conflict detection and are free of conflicts, exhibiting good relation type continuity. Adding valid paths to the path exploration queue allows for further exploration and analysis. For example, when processing paths in a temporary path buffer, if two paths have the same node sequence but different relation type sequences, and one path has higher relation type continuity, then that path is retained, the other is removed, and the retained path is added to the path exploration queue.

[0083] Step S240: Dynamically extract path features during the traversal process, and calculate the entity type diversity, relation type continuity and attribute feature consistency of the current path in real time. Entity type diversity is the proportion of different types of entities in the path, relation type continuity is the semantic correlation between adjacent relation types, and attribute feature consistency is the proportion of shared attributes of entities in the path.

[0084] Entity type diversity is the proportion of different types of entities in a path to the total number of entities in the path, reflecting the richness of entity types in the path. For example, if a path contains three different types of entities: medical behavior entities, crowd activity entities, and environmental factor entities, then the path has high entity type diversity.

[0085] Relationship type continuity refers to the semantic correlation between adjacent relationship types. If the semantic correlation between adjacent relationship types is high, it indicates that the relationship type continuity of the path is good. For example, in a disease transmission path, if the adjacent relationship types are "infection transmission" and "contact infection" respectively, their semantic correlation is high, indicating that the relationship type continuity of the path is good.

[0086] Attribute consistency refers to the proportion of shared attributes among entities in a path. If entities in a path share a large number of attributes, it indicates a strong correlation between them and high attribute consistency. For example, in a path related to medical services, if multiple entities have the same "treatment effect" attribute, then the path has high attribute consistency.

[0087] Step S250: Perform dynamic pruning on the path. When the path length exceeds a reasonable range, the diversity of entity types does not meet the expected standard, or the continuity of relationship types fluctuates abnormally, terminate the current path exploration and remove the corresponding path context from the queue.

[0088] In one implementation, step S250 may specifically include the following steps S251-S256.

[0089] Step S251: Obtain the current path length as monitored in real time. The path length is measured by the number of edges contained in the path. When the number of edges exceeds a reasonable range, the length pruning condition is triggered, and the current path exploration is terminated.

[0090] Real-time monitoring of the current path length involves continuously obtaining the number of edges contained in the path during path traversal. Path length is measured by the number of edges; the path length is incremented by 1 for each edge traversed. A reasonable range is a pre-defined upper and lower limit for the path length. When the real-time monitored current path length exceeds this reasonable range, it indicates that the path is too long, may contain too much irrelevant information, or has exceeded the meaningful exploration range. In this case, a length pruning condition is triggered, terminating the current path exploration. For example, if the reasonable range is 1-10 edges, and the real-time monitored current path length is 12 edges, exceeding the reasonable range, then the current path exploration is terminated, and traversal of the path is stopped.

[0091] Step S252: Calculate the entity type diversity of the current path. Entity type diversity is the ratio of the number of different entity types in the path to the total number of entities in the path. When the ratio does not meet the expected standard, trigger the diversity pruning condition and terminate the current path exploration.

[0092] Entity type diversity is calculated by comparing the number of different entity types in a path with the total number of entities in the path, and then determining their ratio. The total number of entities in a path refers to the total number of all entity nodes contained in the path, while the number of different entity types refers to the number of different types of entities appearing in the path.

[0093] The expected standard is a pre-set threshold for entity type diversity. If the calculated entity type diversity ratio does not meet this expected standard, it indicates that the path information is insufficient and may not accurately reflect the relationships and propagation links between entities. In this case, diversity pruning is triggered, terminating the current path exploration.

[0094] Step S253: Analyze the continuity of the relation type sequence, perform sliding window smoothing on the relation type sequence, calculate the standard deviation of the semantic correlation of relation types within the window, and when the standard deviation exceeds the normal fluctuation range, it is determined that the continuity of relation types has abnormal fluctuations, triggering the continuity pruning condition and terminating the current path exploration.

[0095] Continuity analysis of relation type sequences is a process of determining the stability of semantic association between adjacent relation types by processing and calculating the sequence. Sliding window smoothing smooths the semantic association between relation types within the sequence by sliding a fixed-size window across it. Semantic association refers to the degree of semantic similarity between adjacent relation types, which can be calculated using a predefined semantic similarity calculation method. For example, "infection transmission" and "contact infection" have a high semantic association, while "infection transmission" and "service provision" have a low semantic association.

[0096] Standard deviation is a statistical indicator that measures the degree of data fluctuation. By calculating the standard deviation of the semantic relevance of relation types within a sliding window, it is possible to determine whether there are abnormal fluctuations in the continuity of relation types. The normal fluctuation range is a pre-set threshold of standard deviation. If the calculated standard deviation exceeds the normal fluctuation range, it indicates that there are abnormal fluctuations in the continuity of relation types. At this point, the continuity pruning condition is triggered, terminating the current path exploration.

[0097] Step S254: When any pruning condition is triggered, remove the current path context from the path exploration queue and release the allocated computing resources.

[0098] When any of the length pruning, diversity pruning, or continuity pruning conditions is triggered, it indicates that the current path does not meet the requirements and path exploration needs to be terminated. At this time, the current path context is removed from the path exploration queue, and the allocated computing resources are released.

[0099] Path exploration context is a set of information assigned to each starting node during path traversal, including the set of visited nodes, the current path length, and a list of relation type constraints. Removing the path exploration context can avoid consuming unnecessary computing resources and improve system efficiency. For example, when a length pruning condition is triggered, the current path's context information is removed from the path exploration queue, releasing the memory and other computing resources occupied by that path so that they can be used for exploring other paths.

[0100] Step S255: Record the trigger type and path status of the pruning event. The pruning event includes length pruning, diversity pruning and continuous pruning. The path status includes the node sequence at the time of pruning, the current path length and the relationship type sequence.

[0101] Recording the trigger type and path status of pruning events is for subsequent analysis and debugging. The trigger types for pruning events include length pruning, diversity pruning, and continuity pruning, corresponding to situations where the path length exceeds a reasonable range, entity type diversity does not meet expected standards, and relationship type continuity exhibits abnormal fluctuations, respectively. The path status records the node sequence at the time of pruning, the current path length, and the relationship type sequence. The node sequence records all nodes traversed in the path, the current path length records the number of edges contained in the path at the time of pruning, and the relationship type sequence records the relationship types connecting nodes in the path.

[0102] By recording the trigger type and path status of pruning events, the causes and effects of pruning can be analyzed, and the settings of pruning conditions can be optimized. For example, if length pruning events are frequently triggered, the settings of a reasonable range may need to be adjusted; if there are many diverse pruning events, the settings of the expected criteria may need to be adjusted.

[0103] Step S256: Periodically analyze the distribution characteristics of pruning events, and dynamically adjust the reasonable range, expected standard and normal fluctuation range according to the proportion of pruning types to optimize the balance between path traversal efficiency and integrity.

[0104] Periodically analyzing the distribution characteristics of pruning events refers to the process of statistically analyzing the triggering of pruning events at regular time intervals. The distribution characteristics of pruning events include the proportion of different pruning types and the frequency of pruning events.

[0105] Dynamically adjusting the reasonable range, expected standard, and normal fluctuation range based on the proportion of pruning types is to optimize the efficiency and completeness of path traversal. If the proportion of a certain pruning type is too high, it indicates that the current pruning conditions may be too strict or too lenient, and adjustments are needed. For example, if the proportion of length pruning events is too high, it may indicate that the reasonable range is set too small, causing many meaningful paths to be pruned. In this case, the reasonable range can be appropriately increased.

[0106] By dynamically adjusting pruning conditions, the efficiency of path traversal can be improved while ensuring the integrity of the traversal. For example, more meaningful paths can be retained within a reasonable range, avoiding the loss of important information due to excessive pruning.

[0107] Step S260: Collect all complete paths processed by pruning, extract the node sequence, relation type sequence and path length information in the path. The path length information includes the number of nodes, the total number of edges and the cumulative relation weight. Remove redundant paths containing duplicate node sequences and relation type sequences to obtain the set of associated paths.

[0108] After completing path traversal and pruning, it is necessary to collect all complete paths that have undergone pruning. A complete path is a path that has not been pruned and meets requirements such as length, entity type diversity, and relation type continuity.

[0109] Extract the node sequence, relation type sequence, and path length information from the complete path. The node sequence records all nodes traversed in the path, the relation type sequence records the relation types connecting the nodes, and the path length information includes the number of nodes, the total number of edges, and the cumulative relation weights. The cumulative relation weights refer to the sum of the weights of all edges in the path, reflecting the importance and influence of the path.

[0110] Redundant paths containing duplicate node sequences and relation type sequences are removed to avoid repeatedly analyzing the same path information. By removing redundant paths, a concise set of associated paths is obtained. Each path in the set of associated paths represents a potential propagation link between different entities, providing a foundation for subsequent subgraph pattern mining and risk transmission network construction.

[0111] Step S300: Perform subgraph pattern mining based on the set of associated paths, extract the repeated entity relationship combination structure, and generate a propagation association pattern with time interval markers. The propagation association pattern includes core entity pairs and the relationship chain structure connecting the two.

[0112] In one implementation, step S300 may specifically include the following steps S310-S360.

[0113] Step S310: Divide the associated path set into time dimensions, and divide the path into multiple time segments according to the time characteristics of the starting node. Each time segment contains path data within the same time window. The size of the time window is set according to the typical transmission cycle of public health events.

[0114] Time-based segmentation is the process of classifying a set of related paths according to their temporal characteristics. Because the spread of public health events has a certain time-sensitivity, the spread may differ across different time periods. Therefore, it is necessary to divide the set of related paths into multiple time segments based on the temporal characteristics of their starting points.

[0115] The time window size is set based on the typical transmission cycle of a public health event. The typical transmission cycle refers to the length of time from the onset of a public health event to its peak or end. For example, for a certain infectious disease, the typical transmission cycle might be several weeks or several months. Setting the time window size based on the typical transmission cycle divides the set of associated paths into meaningful time segments. Each time segment contains path data within the same time window. This facilitates subsequent subgraph pattern mining within each time segment, analyzing transmission association patterns across different time periods. For example, the set of associated paths can be divided into multiple time segments based on a one-month time window, and then subgraph patterns can be mined within each monthly time segment to analyze the disease transmission patterns within that month.

[0116] Step S320: Perform multi-scale subgraph pattern mining within each time segment. Set three mining scales: basic scale, medium scale, and extended scale. The path hop count threshold is 1 for the basic scale, 2 for the medium scale, and 3 for the extended scale. Mine candidate subgraph patterns at different scales.

[0117] Basic, intermediate, and extended scales are three different mining scales, each corresponding to a different path hop count threshold. The path hop count threshold refers to the maximum number of edges allowed in a path. The basic scale corresponds to a path hop count threshold of 1, which only considers the relationship between two directly connected nodes; the intermediate scale corresponds to a path hop count threshold of 2, which considers paths containing two edges; and the extended scale corresponds to a path hop count threshold of 3, which considers paths containing three edges.

[0118] By mining candidate subgraph patterns at different scales, we can discover entity relationship combinations of varying scales. For example, at the basic scale, we may find some direct transmission relationships, such as an infection source directly infecting susceptible individuals; at the intermediate scale, we may find transmission relationships through an intermediate node, such as an infection source infecting susceptible individuals through a certain medium; at the extended scale, we may find more complex transmission relationships, such as an infection source spreading to susceptible individuals through multiple intermediate nodes and relationship types.

[0119] Step S330: Perform structural standardization on the candidate subgraph patterns at each scale, and convert the subgraph patterns into directed chain representations with core entity pairs as the starting and ending points. The core entity pairs are determined by calculating the (out-degree - in-degree) value of each node in the subgraph, and the nodes with the largest and smallest values ​​are selected as the starting and ending entities, respectively.

[0120] Structural standardization is used to represent candidate subgraph patterns at various scales in a unified standard form, facilitating subsequent analysis and comparison. For example, in a subgraph pattern, after calculating the (out-degree - in-degree) value of each node, it is found that node A has the largest (out-degree - in-degree) value and node B has the smallest (out-degree - in-degree) value. Therefore, node A is taken as the starting entity and node B as the ending entity, and the subgraph pattern is converted into a directed chain representation with node A as the starting point and node B as the ending point.

[0121] In one implementation, step S330 may specifically include the following steps S331-S336.

[0122] Step S331: Traverse all entity nodes in the candidate subgraph pattern, calculate the in-degree and out-degree of each node. The in-degree is the number of edges pointing to the node, and the out-degree is the number of edges originating from the node. The absolute value of the difference between the in-degree and out-degree is used as the centrality index of the node.

[0123] When performing structural normalization on candidate subgraph patterns, it is first necessary to traverse all entity nodes in the subgraph pattern and calculate the in-degree and out-degree of each node. The in-degree refers to the number of edges pointing to the node, which reflects the node's ability to receive information or influence. The out-degree refers to the number of edges originating from the node, which reflects the node's ability to propagate information or influence.

[0124] The absolute value of the difference between the in-degree and out-degree serves as a centrality indicator for a node. Centrality measures a node's importance and influence within a subgraph. For example, if a node's out-degree is significantly greater than its in-degree, it indicates that the node is a source of information or influence, and its centrality indicator is high; conversely, if a node's in-degree is significantly greater than its out-degree, it indicates that the node is a receiver of information or influence, and its centrality indicator may also be high.

[0125] Step S332: Select the two nodes with the largest centrality index as the core entity pair. The node with the larger centrality index is the starting entity, and the node with the second largest centrality index is the ending entity. If there are nodes with the same centrality index, the starting point and the ending point are determined by the order of the timestamps of the node attribute features.

[0126] After calculating the centrality index of each node, the two nodes with the largest centrality index are selected as the core entity pair. The node with the larger centrality index is used as the starting entity, which is usually the starting point of information or influence; the node with the second largest centrality index is used as the ending entity, which is usually the receiving point of information or influence.

[0127] If nodes with the same centrality metric exist, the start and end points need to be determined by the order of their timestamps. Timestamps record when related events occurred at a node; the node with the earlier timestamp is more likely to be the starting entity. For example, in a subgraph pattern of disease transmission, if two nodes have the same centrality metric, but one node was infected earlier, then that node is considered the starting entity.

[0128] Step S333: Extract paths from the subgraph pattern based on the core entity pairs. Starting from the starting entity, traverse to the ending entity, record all possible node sequences and relation type sequences, and obtain a preliminary directed chain.

[0129] Path extraction based on core entity pairs is the process of finding all possible paths from the starting entity to the ending entity in a subgraph pattern. Starting from the starting entity, the graph is traversed to the ending entity using a graph traversal algorithm (such as depth-first search or breadth-first search), recording all nodes passed through the path and the relationship types of the connecting nodes.

[0130] The recorded sequence of nodes and relation types constitutes the initial directed chain. The initial directed chain may contain multiple different paths because there can be various ways of propagation from the originating entity to the destination entity. For example, in a subgraph pattern of disease transmission, there may be multiple paths from the source of infection to susceptible individuals, such as direct transmission and transmission through intermediate media; all of these paths will be recorded in the initial directed chain.

[0131] Step S334: Filter redundant nodes in the initial directed chain, remove intermediate nodes that do not contribute to the transmission of relationships between core entities, and nodes with both in-degree and out-degree of 1 and whose attribute features are not significantly related to adjacent nodes, and retain the filtered simplified node sequence.

[0132] Redundant node filtering is a process of optimizing the initial directed chain, aiming to remove intermediate nodes that do not contribute to the transmission of relationships between core entities. A node with no contribution is defined as one node with both in-degree and out-degree of 1 and whose attributes have no significant correlation with those of its neighbors.

[0133] A node with both in-degree and out-degree of 1 indicates that it is merely an intermediate transitional node in the path, playing no additional role in the transmission of relationships between core entity pairs. Furthermore, its attributes and features have no significant correlation with adjacent nodes, meaning that this node lacks close connections with its preceding and following nodes in terms of attributes and does not affect the expression of relationships between core entity pairs. For example, in a disease transmission path, if an intermediate node is simply a regular transit point, and its related attributes (such as location and environment) have no substantial impact on the spread of the disease from the source of infection to susceptible populations, and its in-degree and out-degree are both 1, then this node can be considered a redundant node.

[0134] Step S335: Standardize the annotation method of relation type sequence, merge relation types with different expressions but the same semantics into standard relation types, and generate a standardized relation type sequence.

[0135] A standardized method for labeling relation type sequences aims to eliminate discrepancies in relation type representations, facilitating subsequent analysis and comparison. In real-world data, multiple different expressions may represent the same relation type; for example, "infection transmission," "contagion," and "disease spread" may all indicate the transmission of a disease. Merging these semantically identical relation types with different expressions into a standard relation type improves data consistency and understandability.

[0136] Relationship type merging can be achieved by establishing a relation type mapping table. This mapping table records the correspondence between all different relation types and standard relation types. When processing the relation type sequence, each relation type is replaced with its corresponding standard relation type according to the mapping table. For example, "contagion" and "disease spread" are both replaced with "infection transmission". After this processing, the generated standardized relation type sequence is more standardized and can more accurately reflect the relationships between entities.

[0137] Step S336: Combine the simplified node sequence with the standardized relation type sequence to obtain a directed chain representation with the core entity pair as the starting and ending points. The directed chain representation includes the starting entity identifier, the ending entity identifier, the intermediate node sequence, and the standardized relation type sequence.

[0138] The origin and destination entity identifiers uniquely identify the core entity pair, facilitating the location and identification of these two key entities in a heterogeneous association graph. The intermediate node sequence records all intermediate nodes along the path from the origin entity to the destination entity, reflecting the specific path of propagation. The normalized relation type sequence records the relation types connecting these nodes, reflecting the interaction patterns between entities. For example, in a directed chain representation of disease transmission, the origin entity identifier might be the unique identifier of the source of infection, the destination entity identifier the unique identifier of the susceptible population, the intermediate node sequence might contain intermediate mediator nodes involved in the propagation process, and the normalized relation type sequence records the transmission patterns of the disease between different nodes.

[0139] Step S340: Calculate the support and confidence of the standardized candidate subgraph patterns. The support is the ratio of the number of paths containing the subgraph pattern to the total number of paths. The confidence is the semantic consistency of the relation type sequence in the subgraph pattern. Candidate subgraph patterns whose support and confidence both reach the corresponding criteria are retained.

[0140] Support and confidence are two important metrics for evaluating the effectiveness and reliability of standardized candidate subgraph patterns. Support reflects the frequency of a subgraph pattern across all paths; it is the ratio of the number of paths containing that subgraph pattern to the total number of paths. Higher support indicates that the subgraph pattern is more prevalent in the data and may be more representative. For example, if a subgraph pattern appears in most paths, its support is high, potentially reflecting a common propagation pattern.

[0141] Confidence measures the semantic consistency of the relation type sequence within a subgraph schema. Semantic consistency refers to whether the various relation types in a relation type sequence are semantically coherent and reasonable. For example, in a subgraph schema of disease transmission, if the relation type sequence is "infection transmission - contact infection - infection transmission," its semantic coherence is good, and its confidence may be high; if the relation type sequence is "infection transmission - service provision - infection transmission," its semantic coherence is poor, and its confidence may be low.

[0142] Step S350: Perform evolutionary analysis on subgraph patterns across time segments, track the structural changes of the same subgraph pattern in consecutive time segments, calculate pattern structural similarity, and determine the structural similarity by weighted sum of node type sequence matching degree and relation type sequence matching degree. Merge pattern sequences that meet the evolutionary standard into pattern evolution chains.

[0143] Pattern structural similarity is an indicator that measures the degree of structural similarity of the same subgraph patterns across different time segments. It is determined by a weighted sum of node type sequence matching degree and relation type sequence matching degree. Node type sequence matching degree refers to the similarity of node type sequences in two subgraph patterns, which can be calculated by comparing the ratio of the number of identical node types to the total number of nodes. Relation type sequence matching degree refers to the similarity of relation type sequences in two subgraph patterns, which can also be calculated by comparing the ratio of the number of identical relation types to the total number of relation types. The weighted sum is used to comprehensively consider the importance of node types and relation types, and different weights can be assigned to node type sequence matching degree and relation type sequence matching degree according to the actual situation.

[0144] Step S360: Add time interval markers to each pattern evolution chain. The start time of the time interval marker is the starting point of the time segment when the pattern first appears, and the end time is the ending point of the time segment when the pattern structural similarity does not meet the evolution standard. Extract the core entity pairs, relationship chain structure and time interval information in the pattern evolution chain to generate propagation association patterns.

[0145] Adding time interval markers to the pattern evolution chain is to clarify the time range of each propagating pattern. The start time of the time interval marker is the beginning of the time segment in which the pattern first appears, recording when the propagating pattern begins to appear. The end time is the end of the time segment in which the pattern structural similarity does not meet the evolution criteria, indicating that the structure of the propagating pattern has changed significantly after that time and no longer conforms to the original pattern.

[0146] By extracting the core entity pairs, relationship chain structure, and time interval information from the pattern evolution chain, a complete propagation association pattern can be generated. The core entity pairs clarify the start and end points of propagation, the relationship chain structure describes the propagation path and method between entities, and the time interval information provides a time dimension reference for the analysis of the propagation pattern.

[0147] Step S400: Input the propagation association pattern into the risk transmission network construction process, determine the influence weight of each propagation association pattern through node importance assessment, and construct a directed weighted risk transmission network in combination with the relationship chain structure. The directed edges of the risk transmission network represent the propagation direction, and the edge weights represent the influence intensity.

[0148] In one implementation, step S400 may specifically include the following steps S410-S460.

[0149] Step S410: Extract the core entity pairs and relationship chain structure in the propagation association pattern, assign a unique pattern identifier to each core entity pair, and record the relationship type sequence, number of nodes, and time interval information in the relationship chain structure.

[0150] Extracting the core entity pairs and relationship chains from the propagation association patterns is a fundamental step in constructing a risk transmission network. Core entity pairs are the key components of the propagation association patterns, defining the start and end points of propagation. Assigning a unique pattern identifier to each core entity pair facilitates the differentiation and management of different propagation association patterns in subsequent processing.

[0151] In a relational chain structure, the sequence of relation types records the types of connections between entities, the number of nodes reflects the complexity of the propagation path, and the time interval information provides the time range of the propagation pattern. For example, in a disease transmission association pattern, the sequence of relation types may include "infection transmission" or "contact infection," the number of nodes may represent the number of intermediate nodes involved in the propagation process, and the time interval information may record the time period of disease transmission.

[0152] Step S420: Perform multi-dimensional importance assessment on entity nodes in the propagation association pattern. Calculate the node importance index from three dimensions: entity activity, propagation influence, and network embedding value. Entity activity is the number of relational interactions an entity participates in per unit time. Propagation influence is the frequency of an entity's appearance in all association paths. Network embedding value is the vector magnitude after mapping the node to a low-dimensional space using a graph embedding algorithm.

[0153] In one implementation, step S420 may specifically include the following steps S421-S426.

[0154] Step S421: Count the total number of relationship interactions participated in by the entity within the preset time window, distinguish between the number of active interactions of the entity as the subject of the relationship and the number of passive interactions of the entity as the object of the relationship, calculate the weighted interaction number in combination with the time decay factor. The time decay factor decreases as the interval between the time of the interaction and the current time increases. The weighted interaction number is then normalized to obtain the entity activity index.

[0155] The total number of relationship interactions participated in by an entity within a preset time window is the basis for calculating entity activity. The preset time window can be determined based on specific analytical needs, such as one day, one week, or one month. When counting interactions, it is necessary to distinguish between the number of active interactions by the entity as the subject of the relationship and the number of passive interactions by the entity as the object of the relationship. Active interactions reflect the entity's ability to initiate relationships, while passive interactions reflect the entity's acceptance of relationships. The time decay factor is used to account for the timeliness of interactions. As the interval between the interaction's occurrence and the current time increases, the impact of the interaction may gradually decrease. Therefore, the time decay factor decreases as the interval between the interaction's occurrence and the current time increases. For example, recent interactions have a greater impact on entity activity, while interactions that occurred a long time ago have a relatively smaller impact. By combining the time decay factor to calculate weighted interaction counts, the activity level of an entity at the current time can be more accurately reflected. The weighted interaction counts are then normalized, mapping them to a fixed numerical range, such as [0,1], to obtain the entity activity index. Normalization eliminates the dimensional differences in interaction counts between different entities, facilitating comparison and analysis.

[0156] Step S422: Traverse all paths containing the entity in the associated path set, extract the cumulative influence intensity of each path, assign path contribution weights to the paths based on the cumulative influence intensity, statistically analyze the position distribution of the entity in different paths, dynamically adjust the position weight coefficients of the starting position, middle position and ending position of the path with the path contribution weights, calculate the sum of the products of the position weight coefficients and the path contribution weights and perform a second normalization process to obtain the entity propagation influence index.

[0157] Traversing all paths containing the entity in the set of associated paths is for a comprehensive assessment of the entity's propagation influence. The cumulative influence strength of each path reflects its importance in the propagation process and can be calculated based on factors such as the weights of the edges in the path. For example, in a disease transmission path, if the edge weights in the path are large, it indicates that the path has a strong propagation ability and a high cumulative influence strength.

[0158] Path contribution weights are assigned to paths based on their cumulative impact intensity; paths with higher cumulative impact intensity have greater contribution weights. The location distribution of entities within different paths also affects their propagation influence. An entity's role differs at the start, middle, and end points of a path, thus requiring different location weight coefficients. These location weight coefficients are dynamically adjusted based on the path contribution weights; for example, in paths with high cumulative impact intensity, the location weight coefficient for an entity at the start position may be relatively large.

[0159] The sum of the products of the location weight coefficient and the path contribution weight is calculated to obtain a comprehensive propagation influence score. This score is then subjected to a second normalization process, mapping it to the [0,1] interval, to obtain the entity's propagation influence index. In this way, the influence of an entity in the propagation process can be assessed more accurately.

[0160] Step S423: Sample the heterogeneous association graph, set the entity type bias sampling probability, the entity type related to propagation has a higher sampling probability than other types, generate the context sequence of the entity, map the entity to a low-dimensional vector through the word vector model, calculate the magnitude of the vector and perform a third normalization process to obtain the entity network embedding index.

[0161] Sampling heterogeneous association graphs reduces computational complexity while highlighting entities relevant to transmission. Entity type bias sampling probabilities are set so that entities related to transmission have a higher sampling probability than others. For example, in a heterogeneous association graph of disease transmission, entities directly related to transmission, such as infection sources and susceptible populations, have a higher sampling probability, while entities less related to transmission have a lower sampling probability.

[0162] Entity context sequences are generated through sampling. These sequences record information about other entities surrounding the entity, reflecting its local environment within the graph. Word vector models, such as the Word2Vec model, map text or entities to low-dimensional vectors. The entity's context sequence is input into the word vector model, mapping the entity to a low-dimensional vector. The vector's magnitude is calculated, reflecting the entity's position and features in the low-dimensional space. A third normalization process is applied to the vector magnitude, mapping it to the [0,1] interval, yielding the entity network embedding index. The entity network embedding index measures the structural importance and features of entities in a heterogeneous association graph.

[0163] Step S424: Construct a dimension weight learning model based on historical public health event transmission case data. Input the entity importance labeling results in the case data and the corresponding normalized entity activity index, entity transmission influence index and entity network embedding index. Adjust the weight coefficients of each dimension through the gradient descent algorithm to obtain a dynamically updated dimension weight allocation scheme. Calculate the comprehensive importance index of the entity by weighted summation.

[0164] The dimensional weight learning model is used to determine the weight coefficients of the three dimensions—entity activity, dissemination influence, and network embedding—when calculating the overall importance index. This model is built based on historical public health event dissemination case data, which includes labeled entity importance results and corresponding entity activity, dissemination influence, and network embedding indices.

[0165] These data are input into a dimension weight learning model, and the weight coefficients of each dimension are adjusted using a gradient descent algorithm. After multiple iterations, a dynamically updated dimension weight allocation scheme is obtained. Based on this scheme, the normalized entity activity index, entity propagation influence index, and entity network embedding index are weighted and summed to calculate the entity's comprehensive importance index. The comprehensive importance index takes into account information from three dimensions, more comprehensively reflecting the entity's importance in the propagation association pattern.

[0166] Step S425: Use box plot method to detect outliers in the comprehensive importance index, calculate the interquartile range of the index data, determine the upper and lower boundary values, mark entities that exceed the upper and lower boundary values ​​as entities to be adjusted, perform a second evaluation on the marked entities to be adjusted, extract the connection density and relationship type ratio of the entities in the local subgraph, and adjust the comprehensive importance index in combination with local features.

[0167] Box plots are a commonly used outlier detection method. By calculating the interquartile range (IMR) of the composite importance index data (the difference between the upper and lower quartiles), the distribution range of the data can be determined. Upper and lower boundary values ​​are then established based on the IMR; data points exceeding these boundaries are considered outliers.

[0168] Entities exceeding the upper and lower boundary values ​​are marked as entities requiring adjustment, as their overall importance index may be anomaly. A secondary evaluation is performed on these marked entities to extract their connection density and relation type percentages within the local subgraph. Connection density reflects the tightness of connections within the local subgraph, while relation type percentages reflect the distribution of different relation types within the local subgraph.

[0169] The overall importance index can be adjusted by incorporating local features. For example, if an entity to be adjusted has a high connectivity density in a local subgraph but a low overall importance index, its overall importance index may need to be appropriately increased. This approach allows for a more accurate assessment of entity importance, avoiding the influence of outliers on the overall evaluation results.

[0170] Step S426: Establish an entity importance index-time variation curve, record the change data of the entity's importance index in different time windows, calculate the trend slope of the curve, the trend slope is the cumulative sum of the index changes in continuous time windows, and predict the direction of change of entity importance through the trend slope.

[0171] The purpose of establishing an entity importance index-time variation curve is to observe how entity importance changes over time. Data on the changes in the entity's importance index across different time windows is recorded and plotted as a curve. The slope of the curve reflects the trend of the entity's importance index, which is the cumulative sum of the index changes over consecutive time windows.

[0172] If the trend slope is positive, it indicates that the entity's importance index is rising, and the entity's importance may increase in the future; if the trend slope is negative, it indicates that the entity's importance index is falling, and the entity's importance may decrease in the future. The trend slope can predict the direction of change in entity importance, providing a reference for the prevention and management of public health events. For example, during disease transmission, if the importance index of an infection source shows an upward trend, it may be necessary to strengthen the control of that infection source.

[0173] Step S430: Calculate the basic weight of the propagation association pattern based on the node importance index. The basic weight is the product of the importance indices of the starting entity and the ending entity in the core entity pair. The product result is normalized and mapped to a reasonable range.

[0174] The basic weight of the propagation association pattern reflects its fundamental importance in risk transmission and is calculated based on the importance index of the core entity to the starting and ending entities. The starting and ending entities play a crucial role in the propagation process; the higher their importance index, the greater the basic weight of the propagation association pattern is likely to be.

[0175] Multiplying the importance indices of the starting and ending entities yields a preliminary weight value. Since the ranges of importance indices may differ, the product needs to be normalized to a reasonable interval, such as [0,1], for ease of comparison and application. Normalization eliminates dimensional differences between the basic weights of different propagation association patterns, making them comparable. Calculating these basic weights provides a fundamental reference for subsequent analysis of the impact of propagation association patterns.

[0176] Step S440: Analyze the effect of the relation chain structure on the influence weight, calculate the relation chain length correction coefficient, relation type combination coefficient and time decay coefficient. The relation chain length correction coefficient is negatively correlated with the number of edges in the relation chain. The relation type combination coefficient is determined based on the synergistic effect of the relation type sequence. The time decay coefficient decreases as the distance between the time interval of the propagation association pattern and the current time increases.

[0177] In one implementation, step S440 may specifically include the following steps S441-S446.

[0178] Step S441: Count the number of edges in the relation chain as the relation chain length, set the baseline length, and when the actual length is less than the baseline length, the correction coefficient increases as the length decreases; when the actual length is greater than the baseline length, the correction coefficient decreases as the length increases, and the correction coefficient is kept within the preset adjustment range.

[0179] The number of edges in a relation chain is a direct method to determine its length. The number of edges reflects the complexity of the relation chain; more edges mean a longer chain and potentially more complex propagation processes. A baseline length is established as a standard for comparison and can be determined based on actual data and analytical needs.

[0180] When the actual length is less than the baseline length, it indicates that the relationship chain is relatively short, and the propagation process may be more direct and efficient. Therefore, the correction coefficient increases as the length decreases. When the actual length is greater than the baseline length, it indicates that the relationship chain is longer, and the uncertainty and loss in the propagation process may increase. Therefore, the correction coefficient decreases as the length increases. To ensure the rationality of the correction coefficient, it is kept within a preset adjustment range. The preset adjustment range can be set according to the actual situation, for example, [0.1, 1]. This avoids the correction coefficient being too large or too small, ensuring that its corrective effect on the influence weight is within a reasonable range.

[0181] Step S442: Construct a relationship type synergy matrix. The matrix elements represent the synergy effect values ​​of two adjacent relationship types. The synergy effect values ​​are determined based on the frequency of occurrence and propagation success rate of relationship type combinations in historical propagation cases. Traverse the adjacent relationship type combinations in the relationship chain and multiply the synergy effect values ​​to obtain the relationship type combination coefficient.

[0182] A relationship type synergy matrix is ​​a matrix used to describe the synergistic effect between different adjacent relationship types. The matrix elements represent the synergistic effect value of two adjacent relationship types, which is determined based on the frequency of occurrence and transmission success rate of relationship type combinations in historical transmission cases. For example, in historical disease transmission cases, if the relationship type combination "infection transmission - contact infection" occurs frequently and has a high transmission success rate, then its synergistic effect value is large; if the relationship type combination "infection transmission - service provision" occurs rarely and has a low transmission success rate, then its synergistic effect value is small.

[0183] The algorithm iterates through the combinations of adjacent relationship types in the relationship chain, multiplying the synergistic effect values ​​of each combination to obtain the relationship type combination coefficient. For example, in a relationship chain where the adjacent relationship type combinations are "infection transmission - contact infection" and "contact infection - infection transmission," multiplying the synergistic effect values ​​of these two combinations yields the relationship type combination coefficient for that relationship chain. This method allows for a comprehensive consideration of the synergistic effect of all adjacent relationship type combinations within the relationship chain, leading to a more accurate assessment of the chain's propagation capability.

[0184] Step S443: Extract the time difference between the midpoint of the time interval of the propagation correlation pattern and the current time. The time difference is measured in a preset time unit. The time decay coefficient decreases as the time difference increases. When the time difference is 0, the decay coefficient is 1.

[0185] Extracting the time difference between the midpoint of the propagation correlation pattern's time interval and the current time is a crucial step in calculating the time decay coefficient. The time difference is measured in a preset time unit, which can be selected based on actual conditions, such as days, weeks, or months.

[0186] The time decay coefficient decreases as the time difference increases because the influence of a transmission association pattern gradually weakens over time. When the time difference is 0, it means the midpoint of the time interval of the transmission association pattern is the current time, and the decay coefficient is 1, indicating that the influence of the pattern has not diminished. For example, for a currently occurring disease transmission pattern, its time decay coefficient is 1; while for a transmission pattern that occurred a long time ago, the time difference is larger, and the time decay coefficient will be smaller. The time decay coefficient allows for a temporal adjustment of the influence weights of transmission association patterns, making them more consistent with actual transmission patterns.

[0187] Step S444: Normalize the relationship chain length correction coefficient, relationship type combination coefficient, and time decay coefficient respectively, and map each coefficient to the preset adjustment range to ensure that the adjustment range of each coefficient on the comprehensive influence weight is balanced.

[0188] Normalizing the relationship chain length correction coefficient, relationship type combination coefficient, and time decay coefficient is to make them comparable and balanced. Different correction coefficients may have different value ranges. Through normalization, they can be mapped to a preset adjustment range, such as [0,1].

[0189] Normalization can eliminate the dimensional differences between the coefficients and ensure that their adjustment magnitudes on the overall influence weights are balanced.

[0190] Step S445: Collect actual transmission data of public health events according to a preset cycle, calculate the deviation between the predicted impact weight and the actual impact intensity, and fine-tune the calculation parameters of the relationship chain length correction coefficient, relationship type combination coefficient and time decay coefficient based on the deviation value to optimize the accuracy of impact weight assessment.

[0191] Collecting actual transmission data of public health events at predetermined intervals is to verify the accuracy of predicted impact weights. The predetermined interval can be determined based on actual circumstances, such as weekly, monthly, or quarterly. By collecting actual transmission data, we can understand the actual impact of transmission association patterns, i.e., the strength of the actual impact.

[0192] The deviation between the predicted impact weight and the actual impact intensity is calculated; the deviation value reflects the difference between the prediction and the actual situation. The calculation parameters of the relationship chain length correction coefficient, relationship type combination coefficient, and time decay coefficient are fine-tuned based on the deviation value. For example, if the predicted impact weight is greater than the actual impact intensity, it may be necessary to appropriately decrease the relationship chain length correction coefficient, relationship type combination coefficient, or time decay coefficient; if the predicted impact weight is less than the actual impact intensity, it may be necessary to appropriately increase these coefficients. By continuously fine-tuning the calculation parameters, the accuracy of the impact weight assessment can be optimized, making the prediction results closer to the actual situation and providing a more reliable basis for the prevention and management of public health events.

[0193] Step S446: Output a set of correction coefficients including relation chain length correction coefficient, relation type combination coefficient, and time decay coefficient.

[0194] The output correction coefficient set is designed to integrate the calculated relation chain length correction coefficient, relation type combination coefficient, and time decay coefficient for convenient subsequent use. This set contains key information for correcting the impact weights of propagation association patterns, allowing for a more accurate assessment of the actual influence of these patterns.

[0195] Step S450: Calculate the comprehensive impact weight of the propagation association pattern by combining the basic weight and various correction coefficients. The comprehensive impact weight is calculated by multiplying the basic weight by the relationship chain length correction coefficient, the relationship type combination coefficient, and the time decay coefficient. The propagation association pattern that reaches the minimum standard of comprehensive impact weight is retained.

[0196] The comprehensive base weight and various correction coefficients are designed to fully consider the characteristics of the propagation association pattern in different aspects, thereby more accurately calculating its overall impact weight. The base weight reflects the importance of the core entity pair, the relationship chain length correction coefficient considers the complexity of the relationship chain, the relationship type combination coefficient reflects the synergistic effect of relationship types, and the time decay coefficient considers the impact of time on the influence of the propagation pattern.

[0197] The comprehensive influence weight of the propagation association pattern is obtained by multiplying the base weight by the relationship chain length correction coefficient, the relationship type combination coefficient, and the time decay coefficient. This calculation method can integrate the effects of various factors, making the comprehensive influence weight more reflective of the influence of the propagation association pattern in the actual propagation process.

[0198] Propagation association patterns that meet the minimum overall impact weight should be retained. The minimum weight can be set according to actual needs and data characteristics. Propagation association patterns that do not meet the minimum overall impact weight can be discarded to reduce the complexity of subsequent analysis and focus on studying more influential propagation patterns.

[0199] Step S460: Based on the relational chain structure of the propagation association pattern, determine the direction and connection node of each directed edge it contains; use the comprehensive influence weight of the propagation association pattern as the strength contribution value of each directed edge contained in the pattern; for the same directed edge contributed by multiple propagation association patterns, merge all its strength contribution values ​​to obtain the final weight value of the directed edge; perform network consistency verification on the final weight value of the directed edge to ensure that the weight ratio of edges with different directions between the same entity pairs conforms to the law of public health transmission, and integrate all the verified directed edges to construct a directed weighted risk transmission network.

[0200] Determining the direction of directed edges and connecting nodes based on the relational chain structure of the propagation association pattern is a fundamental step in constructing a directed weighted risk transmission network. The relational chain structure clarifies the connections between entities and the direction of propagation, allowing us to determine the start and end points of each directed edge, as well as its direction.

[0201] The overall influence weight of the propagation association pattern is used as the strength contribution value of each directed edge contained in the pattern. Because the overall influence weight reflects the overall influence of the propagation association pattern, allocating it to each directed edge can reflect the degree of influence of the pattern on the edge.

[0202] For the same directed edge contributed by multiple propagation association patterns, it is necessary to merge all its strength contribution values. The merging method can be chosen according to the specific situation, such as using a weighted summation method. By merging the strength contribution values ​​of multiple propagation association patterns, the final weight value of the directed edge can be obtained, which more accurately reflects the actual influence of the edge in the propagation process.

[0203] Network consistency checks on the final weights of directed edges are performed to ensure the rationality of the network structure. The weight ratio of edges in different directions between the same entity should conform to the laws of public health transmission. For example, in a disease transmission network, the edge weight from the source of infection to the susceptible population should generally be greater than the edge weight from the susceptible population to the source of infection. Through network consistency checks, weight values ​​that do not conform to the laws can be identified and corrected, ensuring the reliability of the risk transmission network.

[0204] By integrating all validated directed edges, a directed weighted risk transmission network is constructed. This network can visually demonstrate the transmission path and impact of public health events, providing important reference for public health decision-making.

[0205] Step S500: Perform path prediction analysis on the risk transmission network, identify key transmission paths from the initial risk entity to the target susceptible entity, and generate public health event transmission path prediction results by combining the association strength and relationship type of entities on the key transmission path. The public health event transmission path prediction results include the path entity sequence, relationship type sequence, and cumulative impact strength.

[0206] In one implementation, step S500 may include the following steps S510-S5120.

[0207] Step S510: Locate the initial risk entity and the target susceptible entity in the risk transmission network. The initial risk entity is the environmental factor entity with abnormal indicators in the monitoring data or the medical behavior entity with clustered case reports. The target susceptible entity is the population activity entity with population density and the proportion of underlying diseases reaching the set conditions.

[0208] Locating the initial risk entities and target susceptible entities within the risk transmission network is the first step in path prediction analysis. Initial risk entities are potential sources of public health events. Environmental factors exhibiting abnormal indicators in monitoring data may trigger public health problems; for example, severely excessive air quality may lead to the spread of respiratory diseases. Medical activity entities with clustered case reports may be disease transmission centers, such as a hospital experiencing a large number of cases with the same symptoms. Target susceptible entities are those easily affected by public health events; population activity entities meeting certain criteria in terms of population density and underlying disease prevalence fit this description.

[0209] Step S520: Construct a set of spatiotemporal constraints. The time constraint is that the timestamp of the relationship in the path must be within the preset monitoring period. The spatial constraint is that the overlap rate of the spatial coverage of entities in the path reaches the spatial standard. Transform the spatiotemporal constraints into filtering rules for path search.

[0210] Constructing a set of spatiotemporal constraints aims to narrow down the scope of path searches and improve the accuracy of path prediction. The temporal constraints require that the timestamps of relationships occurring within the path fall within a preset monitoring period. This preset monitoring period can be determined based on the characteristics of the public health event and the analysis needs, such as a month or a quarter. This ensures that the searched paths occurred within a specific timeframe, excluding outdated or future transmission paths.

[0211] Spatial constraints require that the spatial overlap rate of entities within the path meets a spatial standard. This spatial standard can be set according to specific circumstances, for example, requiring an overlap rate of over 50%. This ensures that the searched paths have a certain degree of spatial relevance, avoiding the discovery of irrelevant long-distance propagation paths.

[0212] By transforming spatiotemporal constraints into filtering rules for path search, only paths that satisfy the spatiotemporal constraints are considered during the path search process. This reduces unnecessary computation and improves the efficiency of path search.

[0213] Step S530: Perform multi-path collaborative search, searching for all possible paths from the initial risk entity node to the target susceptible entity node in the risk transmission network.

[0214] Multi-path collaborative search aims to comprehensively identify all possible paths from the initial at-risk entity node to the target susceptible entity node. Since public health events can spread through multiple pathways, searching only a single path may miss some important transmission information.

[0215] In risk transmission networks, graph search algorithms, such as breadth-first search or depth-first search, are used to gradually expand the search scope from the initial risk entity node until all paths to the target susceptible entity node are found. Through multi-path collaborative search, a more comprehensive understanding of the potential spread of public health events can be obtained, providing richer information for subsequent analysis.

[0216] Step S540: Calculate the cumulative influence strength of each possible path by multiplying the weight values ​​of all edges in the path. The larger the weight value, the greater the cumulative influence strength.

[0217] Calculating the cumulative impact strength of each possible path is to assess the importance of each path in the spread of the public health event. The cumulative impact strength is calculated by multiplying the weights of all edges in the path. The weight of an edge reflects its influence in the spread process; the larger the weight, the greater the contribution of the edge to the spread.

[0218] Step S550: Sort all possible paths according to their cumulative impact strength, and extract a set number of the top-ranked paths as candidate critical propagation paths.

[0219] Ranking all possible paths by cumulative impact strength is to filter out the most influential paths. Paths ranked higher usually have higher cumulative impact strength and are more likely to be key propagation paths.

[0220] Extract a predetermined number of paths from the top-ranked paths as candidate key propagation paths. The number can be determined based on actual needs and data characteristics. For example, paths ranking in the top 10% by cumulative influence can be selected as candidate key propagation paths. This approach allows focus on the most influential propagation paths, reducing the complexity of subsequent analysis.

[0221] Step S560: Analyze the path entity sequence and relation type sequence of candidate critical propagation paths, remove paths containing redundant entities or abnormal relation types, and retain paths with reasonable path structure and high cumulative influence as critical propagation paths.

[0222] Analyzing the entity sequences and relation type sequences of candidate critical transmission paths is to further filter out the truly critical transmission paths. Redundant entities are those that have no substantial impact on the transmission process; for example, in a disease transmission path, a location entity unrelated to disease transmission can be considered redundant. Abnormal relation types are those that do not conform to the laws of public health transmission; for example, the appearance of a relation type like "service provision" in a disease transmission path, which is unrelated to transmission.

[0223] Removing paths containing redundant entities or anomalous relationship types can improve the quality of critical transmission paths. Paths with reasonable structures and high cumulative impact are retained as critical transmission paths; these paths more accurately reflect the spread of public health events and provide a reliable basis for public health decision-making.

[0224] Step S570: Traverse each entity node on the critical propagation path and extract the attribute features of the entity node. The attribute features of the entity node include entity type, time features and spatial features. The time features are the time intervals in which the entity is active, and the spatial features are the geographical range covered by the entity.

[0225] Traversing each entity node along the critical propagation path is to gain a comprehensive understanding of the characteristics of the entities in the path. The attribute characteristics of entity nodes include entity type, temporal characteristics, and spatial characteristics, which can provide detailed information about the entity.

[0226] Entity type clarifies the category of an entity, such as a medical behavior entity, a population activity entity, or an environmental factor entity. Different types of entities may play different roles in the spread of public health events. Temporal characteristics record the time interval of an entity's activity, such as the infection time interval of an infection source, which is crucial for analyzing the temporal patterns of transmission. Spatial characteristics describe the geographical extent of an entity's coverage, such as the geographical location and coverage area of ​​a population activity site, which helps to understand the spatial scope of transmission.

[0227] Step S580: Evaluate the attribute correlation between adjacent entity nodes. The attribute correlation is calculated based on the shared temporal or spatial features between entity nodes. The more shared features, the higher the attribute correlation. Combine the weight value of the connecting edge with the attribute correlation to evaluate the tightness of the connection relationship.

[0228] Evaluating the attribute correlation between adjacent entity nodes is to understand the relationships between entities. Attribute correlation is calculated based on the shared temporal or spatial characteristics between entity nodes. If two adjacent entity nodes have overlapping active periods in time or overlapping coverage areas in space, then their attribute correlation is relatively high.

[0229] Step S590: Identify the sequence of relation types on the critical propagation path, count the frequency of occurrence of relation types in the path and the sum of the corresponding edge weights, mark the relation types with the highest frequency and the sum of their weights exceeding a certain proportion of the total weight of the path as critical relation types, and record the positional distribution of critical relation types in the path.

[0230] Identifying the sequence of relation types on critical propagation paths aims to identify the relation types that play a key role in the propagation process. The frequency of each relation type in the path and the sum of its corresponding edge weights are statistically analyzed; relation types with high frequency and large sums of edge weights are likely to contribute more to the propagation.

[0231] Relationship types that rank first and second in frequency and whose sum of weights exceeds a certain percentage of the total path weight are marked as key relationship types. This percentage can be set according to the actual situation, for example, requiring the sum of weights to exceed 50% of the total path weight. Record the positional distribution of key relationship types in the path, which can help analyze the key links and turning points in the propagation process.

[0232] Step S5100: Analyze the bottleneck nodes of the critical propagation path and calculate the betweenness centrality of each entity in the path. The betweenness centrality is the ratio of the number of sub-paths passing through the entity to the total number of sub-paths. Entities whose betweenness centrality reaches the set ratio are marked as bottleneck nodes. Bottleneck nodes are the key control points in the propagation path.

[0233] Analyzing bottleneck nodes in critical propagation paths aims to identify key control points within those paths. Betweenness centrality, a metric for measuring a node's importance within a path, is the ratio of the number of sub-paths passing through that node to the total number of sub-paths. Higher betweenness centrality indicates a stronger bridging role for the node during propagation, suggesting it may be a critical control point.

[0234] The set ratio can be adjusted according to the actual situation, for example, requiring betweenness centrality to reach 0.5 or higher. Entities with betweenness centrality reaching the set ratio are marked as bottleneck nodes. In the spread of public health events, bottleneck nodes may be the key targets for prevention and control.

[0235] Step S5110: Calculate the cumulative influence strength of the critical propagation path. The cumulative influence strength is a measure of the combined effect of all edge weights along the path.

[0236] Calculating the cumulative impact strength of a critical transmission path is to assess its overall influence on the spread of a public health event. The cumulative impact strength is a measure of the combined effect of all edge weights along the path, taking into account the influence of each edge.

[0237] Step S5120: Integrate the path entity sequence, association strength, key relationship types, bottleneck nodes, and cumulative impact strength to generate public health event transmission path prediction results. The results include a path topology map, key node annotations, a heat map of relationship type strength, and a cumulative impact trend curve. The path topology map uses directed arrows to indicate the transmission direction. The node size is proportional to the entity importance index, and the arrow thickness is proportional to the edge weight value.

[0238] Integrating the path entity sequence, association strength, key relationship types, bottleneck nodes, and cumulative impact intensity is intended to generate comprehensive predictions of public health event transmission paths. The path entity sequence records all entities traversed in the path; association strength reflects the tightness of connections between adjacent entities; key relationship types indicate the types of relationships that play a crucial role in the transmission process; bottleneck nodes are key control points in the transmission path; and cumulative impact intensity measures the overall influence of the path.

[0239] The generated public health event transmission path prediction results include a path topology map, key node annotations, a heatmap of relationship type strength, and a cumulative impact trend curve. The path topology map uses directed arrows to indicate the direction of transmission; node size is proportional to the entity importance index, and arrow thickness is proportional to the edge weight value, thus visually demonstrating the structure of the transmission path and the importance of each element. Key node annotations clearly identify the location and role of bottleneck nodes and other critical nodes. The relationship type strength heatmap uses color intensity to represent the strength of different relationship types, facilitating comparison and analysis. The cumulative impact trend curve shows how the cumulative impact intensity changes with the path, helping to understand the dynamic process of transmission.

[0240] Please see Figure 2 , Figure 2 This is a schematic diagram of a computer system provided in an embodiment of the present invention. The computer system includes at least one processor 210, a memory 250, at least one network interface 220, and an external interface 230. The various components in the computer system 200 are coupled together via a bus system 240. It is understood that the bus system 240 is used to implement communication between these components. In addition to a data bus, the bus system 240 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 2 The general labeled all buses as Bus System 240.

[0241] Processor 210 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0242] External interface 230 may include, for example, one or more speakers and / or one or more visual displays. External interface 230 may also include one or more input devices 432, such as a keyboard, mouse, microphone, touch screen display, camera, etc.

[0243] The memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 250 may optionally include one or more storage devices physically located away from the processor 210.

[0244] The memory 250 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 250 described in this embodiment is intended to include any suitable type of memory.

[0245] In some embodiments, memory 250 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0246] Operating system 251 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks;

[0247] The network communication module 252 is used to reach external devices, such as data sources, via one or more (wired or wireless) network interfaces 220. Exemplary network interfaces 220 include Bluetooth, WiFi, and Universal Serial Bus (USB), etc.

[0248] Presentation module 253 is configured to enable the display of information (e.g., external interface for operating peripheral devices and displaying content and information) via one or more output devices 231 (e.g., display screen, speaker, etc.) associated with external interface 230;

[0249] The input processing module 254 is used to detect and translate one or more user inputs or interactions from one or more input devices 232.

[0250] In one embodiment, the processor 210 executes the public health monitoring information analysis method based on graph data provided in the above embodiments of the present invention by running a computer program in the memory 250.

[0251] This invention also provides a computer-readable storage medium storing computer-executable instructions or a computer program. When the computer-executable instructions or the computer program are executed by a processor, the processor will execute the public health monitoring information analysis method based on graph data provided in this invention. For example, ... Figure 1 The method for analyzing public health surveillance information based on graph data is shown.

Claims

1. A method for analyzing public health monitoring information based on graph data, characterized in that, The method includes: Obtain public health monitoring related information, and establish a heterogeneous association graph containing medical behavior entities, population activity entities and environmental factor entities through entity relationship extraction. The node set of the heterogeneous association graph contains the unique identifier and attribute characteristics of the entities, and the edge set contains the direct interaction relationship and indirect influence relationship between entities. The heterogeneous association graph is subjected to topological structure parsing, and the connectivity between entity nodes is traversed to generate an association path set containing node sequence, relationship type sequence and path length information. Each path in the association path set represents a potential propagation link between different entities. Based on the set of associated paths, subgraph pattern mining is performed to extract the repeated entity relationship combination structure and generate a propagation association pattern with time interval markers. The propagation association pattern includes core entity pairs and the relationship chain structure connecting the two. The propagation association patterns are input into the risk transmission network construction process. The influence weight of each propagation association pattern is determined by node importance assessment. A directed weighted risk transmission network is constructed by combining the relationship chain structure. The directed edges of the risk transmission network represent the propagation direction, and the edge weights represent the influence intensity. The risk transmission network is subjected to path prediction analysis to identify key transmission paths from the initial risk entity to the target susceptible entity. The transmission path prediction results of the public health event are generated by combining the association strength and relationship type of the entities on the key transmission path. The transmission path prediction results of the public health event include the path entity sequence, the relationship type sequence, and the cumulative impact strength.

2. The method according to claim 1, characterized in that, The acquisition of public health monitoring-related information, through entity relationship extraction, establishes a heterogeneous association graph containing entities of medical behavior, population activity, and environmental factors, including: The public health monitoring-related information is processed by multi-source data block segmentation, and divided into medical record data blocks, population flow data blocks and environmental sensor data blocks according to the data source type. Each data block contains a timestamp sequence and data field description information. Entity boundary recognition is performed in each data block. Entity candidate segments are extracted by scanning text data through a sliding window. The entity type of the candidate segments is determined based on the named entity recognition model, and the preliminary recognition results of medical behavior entities, population activity entities, and environmental factor entities are output. The preliminary identification results are disambiguated by calculating the attribute similarity between candidate entities of the same type, merging entities whose attribute similarity reaches the set standard, assigning a globally unique identifier to the merged entity, and establishing an entity-identifier mapping table. The dynamic attribute features of each entity are extracted. The dynamic attribute features of the medical behavior entity include diagnosis and treatment related features, service object features, and operation standard compliance features. The dynamic attribute features of the population activity entity include activity scale features, duration features, and spatial coverage features. The dynamic attribute features of the environmental factor entity include monitoring value fluctuation features, spatiotemporal distribution features, and related population contact features. Identify the direct interaction relationships between entities, calculate the co-occurrence frequency of different types of entities within the same time window through entity co-occurrence analysis, determine the action subject and object relationship of entity pairs by combining semantic role labeling, and generate a set of direct interaction relationships; The method explores indirect influence relationships between entities, constructs an attribute association matrix based on entity attribute features, extracts potential association dimensions through matrix singular value decomposition, marks entity pairs whose association dimension matching degree reaches a set standard as indirect influence relationships, and integrates direct and indirect influence relationships to construct a heterogeneous association graph.

3. The method according to claim 2, characterized in that, The step of performing entity disambiguation processing on the preliminary identification results, calculating the attribute similarity between candidate entities of the same type, and merging entities whose attribute similarity reaches a set standard includes: From the preliminary identification results, static attribute sets and dynamic attribute sets of candidate entities of the same type are extracted. The static attribute set includes entity type label, unique identifier prefix and core attribute fields, while the dynamic attribute set includes feature parameter sequence and timestamp information that change over time. A structured comparison is performed on the static attribute set, and weight coefficients are assigned according to the importance level of the attribute fields. The importance level is determined based on the entity type and the attribute's contribution to the differentiation in historical disambiguation cases. The weighted attribute matching score is calculated as the sum of the products of the matching value of each field and the corresponding weight coefficient. Time series similarity analysis is performed on dynamic attribute sets to calculate the morphological similarity of feature parameter sequences, extract trend features, periodic features and mutation point distribution features of the sequences, and generate dynamic trend similarity through multi-feature fusion. Dynamic trend similarity comprehensively reflects the degree of similarity between the overall morphology and local features of the sequences. A static-dynamic fusion model is constructed. The weighted attribute matching score and the dynamic trend similarity are input into the fusion model. The two are mapped to the same feature space through nonlinear transformation to generate a standardized similarity score. The standardized similarity score eliminates the difference in units and unifies them to a set numerical range. Introducing entity association context information, the common neighbor ratio and shortest path distance of candidate entity pairs in the heterogeneous association graph are calculated. The common neighbor ratio is the ratio of the number of shared neighbor nodes to the total number of neighbor nodes, and the shortest path distance is the number of edges contained in the shortest path connecting two entities. The context association degree is generated based on the common neighbor ratio and the shortest path distance. The final attribute similarity is determined by combining the standardized similarity score and the contextual relevance. When the final attribute similarity reaches the set standard, the corresponding candidate entities are merged. The record with the highest completeness in each attribute field is retained as the attribute information of the merged entity. The node connection relationship of the entity-identifier mapping table and the heterogeneous association graph is updated.

4. The method according to claim 1, characterized in that, The process of performing topological structure parsing on the heterogeneous association graph, traversing the connectivity relationships between entity nodes, and generating an association path set containing node sequences, relationship type sequences, and path length information includes: Obtain the multi-level index structure of the pre-constructed heterogeneous association graph. The first level index is the entity type index, which groups nodes according to entity type; the second level index is the relation type index, which groups edges according to relation type; and the third level index is the attribute feature index, which divides node subsets according to the range of entity attribute feature values. Initialize the path exploration queue, add each entity node in the heterogeneous association graph as a starting node to the queue, and assign a path exploration context to each starting node. The context contains the set of visited nodes, the current path length, and a list of relationship type constraints. Perform a hierarchical path traversal operation, starting from the starting node, and traversing adjacent nodes corresponding to the relationship types according to the preset priority order of public health transmission characteristics based on the relationship type index. During the traversal, path features are dynamically extracted, and the entity type diversity, relation type continuity, and attribute feature consistency of the current path are calculated in real time. Entity type diversity is the proportion of different types of entities in the path, relation type continuity is the semantic association degree of adjacent relation types, and attribute feature consistency is the proportion of shared attributes of entities in the path. The path is dynamically pruned. When the path length exceeds a reasonable range, the diversity of entity types does not meet the expected standard, or the continuity of relationship types fluctuates abnormally, the current path exploration is terminated and the corresponding path context is removed from the queue. Collect all complete paths processed by pruning, extract the node sequence, relation type sequence and path length information in the path. The path length information includes the number of nodes, the total number of edges and the cumulative relation weight. Remove redundant paths containing duplicate node sequences and relation type sequences to obtain the set of associated paths.

5. The method according to claim 4, characterized in that, The hierarchical path traversal operation, starting from the starting node, traverses adjacent nodes corresponding to relation types according to a preset priority order based on the relation type index, according to the public health transmission characteristics, including: The preset relationship priority ranking table is read from the relationship type index. The priority ranking table is determined according to the strength of the relationship in the public health transmission chain. Transmission media relationships take precedence over exposure-contact relationships, and exposure-contact relationships take precedence over service provision relationships. Obtain the list of neighboring nodes of the current node. The list of neighboring nodes includes nodes connected by different relationship types and their corresponding relationship attributes. The relationship attributes include the timestamp of the relationship occurrence, the duration, and the scope of influence. Sort the list of adjacent nodes according to the relationship priority sorting table, and put the adjacent nodes corresponding to the relationship types with higher priority at the top. Adjacent nodes with the same priority are sorted according to the duration of the relationship. Select the sorted adjacent nodes as the next hop nodes in turn, check whether the next hop node is in the set of visited nodes, and if not, add it to the set of visited nodes, and update the current path length and relationship type sequence. If there are unvisited nodes in the list of adjacent nodes, continue the traversal operation; when all adjacent nodes have been visited or the path length limit has been reached, store the current path information in a temporary path buffer. Perform conflict detection on the paths in the temporary path buffer. If there are path conflicts with the same node sequence but different relationship type sequences, retain the path with higher relationship type continuity, remove the conflicting path, and add the valid path to the path exploration queue. The dynamic pruning of the path involves terminating the current path exploration when the path length exceeds a reasonable range, the diversity of entity types fails to meet the expected standard, or the continuity of relationship types exhibits abnormal fluctuations. This includes: Obtain the current path length as monitored in real time. The path length is measured by the number of edges contained in the path. When the number of edges exceeds a reasonable range, trigger the length pruning condition and terminate the current path exploration. Calculate the entity type diversity of the current path. The entity type diversity is the ratio of the number of different entity types in the path to the total number of entities in the path. When the ratio does not meet the expected standard, the diversity pruning condition is triggered, and the current path exploration is terminated. Analyze the continuity of the relation type sequence, perform sliding window smoothing on the relation type sequence, calculate the standard deviation of the semantic correlation of relation types within the window, and when the standard deviation exceeds the normal fluctuation range, it is determined that the continuity of relation types has abnormal fluctuation, triggering the continuity pruning condition and terminating the current path exploration; When any pruning condition is triggered, the current path context is removed from the path exploration queue, and the allocated computing resources are released. Record the trigger type and path status of pruning events. Pruning events include length pruning, diversity pruning, and continuous pruning. Path status includes the node sequence at the time of pruning, the current path length, and the sequence of relationship types. Regularly analyze the distribution characteristics of pruning events, and dynamically adjust the reasonable range, expected standard, and normal fluctuation range according to the proportion of pruning types to optimize the balance between path traversal efficiency and integrity.

6. The method according to claim 1, characterized in that, The step of performing subgraph pattern mining based on the set of associated paths, extracting recurring entity relationship combinations, and generating propagation association patterns with time interval markers includes: The associated path set is divided into time dimensions, and the path is divided into multiple time segments according to the time characteristics of the starting node. Each time segment contains path data within the same time window, and the size of the time window is set according to the typical transmission cycle of public health events. Multi-scale subgraph pattern mining is performed within each time segment. Three mining scales are set: basic scale, medium scale, and extended scale. The path hop threshold is 1 for the basic scale, 2 for the medium scale, and 3 for the extended scale. Candidate subgraph patterns are mined at different scales. The candidate subgraph patterns at each scale are structurally standardized and converted into directed chain representations with core entity pairs as the starting and ending points. The core entity pairs are determined by calculating the difference between the out-degree and in-degree of each node in the subgraph, and the nodes with the largest and smallest values ​​are selected as the starting and ending entities, respectively. Calculate the support and confidence of standardized candidate subgraph patterns. The support is the ratio of the number of paths containing the subgraph pattern to the total number of paths, and the confidence is the semantic consistency of the relation type sequence in the subgraph pattern. Candidate subgraph patterns whose support and confidence both meet the corresponding criteria are retained. Evolutionary analysis is performed on subgraph patterns across time segments to track the structural changes of the same subgraph pattern in consecutive time segments. Pattern structural similarity is calculated, and structural similarity is determined by a weighted sum of node type sequence matching degree and relation type sequence matching degree. Pattern sequences that meet the evolutionary standard are merged into a pattern evolution chain. Add time interval markers to each pattern evolution chain. The start time of the time interval marker is the starting point of the time segment when the pattern first appears, and the end time is the ending point of the time segment when the pattern structural similarity does not meet the evolution standard. Extract the core entity pairs, relationship chain structure and time interval information in the pattern evolution chain to generate propagation association patterns.

7. The method according to claim 6, characterized in that, The structural normalization process for candidate subgraph patterns at each scale, converting the subgraph patterns into directed chain representations with core entity pairs as start and end points, includes: Traverse all entity nodes in the candidate subgraph pattern, calculate the in-degree and out-degree of each node. The in-degree is the number of edges pointing to the node, and the out-degree is the number of edges originating from the node. The absolute value of the difference between the in-degree and out-degree is used as the centrality index of the node. The two nodes with the largest centrality index are selected as the core entity pair. The node with the larger centrality index is the starting entity, and the node with the second largest centrality index is the ending entity. If there are nodes with the same centrality index, the starting point and the ending point are determined by the order of the timestamps of the node attribute features. Based on the core entity pairs, path extraction is performed on the subgraph pattern. Starting from the starting entity, the path is traversed to the ending entity, and all possible node sequences and relation type sequences are recorded to obtain a preliminary directed chain. Redundant nodes are filtered in the initial directed chain, and intermediate nodes that do not contribute to the transmission of relationships between core entities are removed. Nodes that do not contribute refer to nodes with an in-degree and out-degree of 1 and whose attribute features are not significantly related to those of their neighboring nodes. The filtered and simplified node sequence is retained. A unified annotation method for relation type sequences is used to merge relation types with different expressions but the same semantics into a standard relation type, generating a standardized relation type sequence; By combining the simplified node sequence with the standardized relation type sequence, a directed chain representation with the core entity pair as the starting and ending points is obtained. The directed chain representation includes the starting entity identifier, the ending entity identifier, the intermediate node sequence, and the standardized relation type sequence.

8. The method according to claim 1, characterized in that, The process of inputting the propagation association patterns into the risk transmission network construction process, determining the influence weight of each propagation association pattern through node importance assessment, and constructing a directed weighted risk transmission network in conjunction with the relationship chain structure includes: Extract the core entity pairs and relationship chain structure from the propagation association pattern, assign a unique pattern identifier to each core entity pair, and record the relationship type sequence, number of nodes, and time interval information in the relationship chain structure. The importance of entity nodes in the propagation association pattern is evaluated in multiple dimensions. The node importance index is calculated from three dimensions: entity activity, propagation influence and network embedding value. Entity activity is the number of relational interactions that an entity participates in per unit time. Propagation influence is the frequency of occurrence of an entity in all association paths. Network embedding value is the vector magnitude after mapping the node to a low-dimensional space through a graph embedding algorithm. The basic weight of the propagation association pattern is calculated based on the node importance index. The basic weight is the product of the importance indices of the starting entity and the ending entity in the core entity pair. The product result is normalized and mapped to a reasonable range. The effect of relation chain structure on the influence weight is analyzed. The relation chain length correction coefficient, relation type combination coefficient and time decay coefficient are calculated. The relation chain length correction coefficient is negatively correlated with the number of edges in the relation chain. The relation type combination coefficient is determined based on the synergistic effect of the relation type sequence. The time decay coefficient decreases as the distance between the time interval of the propagation association pattern and the current time increases. The comprehensive impact weight of the propagation association pattern is calculated by combining the basic weight and various correction coefficients. The comprehensive impact weight is calculated by multiplying the basic weight by the relationship chain length correction coefficient, the relationship type combination coefficient, and the time decay coefficient. Propagation association patterns that achieve the minimum comprehensive impact weight are retained. Based on the relational chain structure of the propagation association pattern, the direction and connection node of each directed edge contained therein are determined; the comprehensive influence weight of the propagation association pattern is used as the strength contribution value of each directed edge contained in the pattern; for the same directed edge contributed by multiple propagation association patterns, all its strength contribution values ​​are merged to obtain the final weight value of the directed edge; the final weight value of the directed edge is verified for network consistency to ensure that the weight ratio of edges with different directions between the same entity pairs conforms to the laws of public health transmission, and all directed edges that pass the verification are integrated to construct a directed weighted risk transmission network.

9. A computer system, characterized in that, include: A memory, wherein a computer program is stored; A processor is configured to load the computer program to implement the public health monitoring information analysis method based on graph data as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 8.

Images