Enterprise big data-oriented adaptive multi-modal entity disambiguation method and system

Abstract

The invention discloses a self-adaptive multi-modal entity disambiguation method and system for enterprise big data, and relates to the technical field of enterprise big data analysis. The method comprises the following steps: constructing an influence graph by extracting stock right relation data and guarantee relation data, and screening core nodes by adopting a PageRank algorithm; adaptively adjusting the influence propagation weight based on the relation strength to obtain node score distribution; recognizing a fracture relationship in combination with historical name change, and repairing the atlas through multi-modal feature similarity; conflict intensity distribution is generated by comparing node score changes before and after restoration, and time sequence optimization is carried out on attenuation parameters in combination with enterprise business transformation events; and finally, enterprise entity identity determination is completed based on the optimized conflict intensity distribution. The entity disambiguation accuracy and stability in a complex enterprise network can be improved.

Description

Technical Field

[0001] This invention relates to the field of enterprise big data analytics, and in particular to an adaptive multimodal entity disambiguation method and system for enterprise big data. Background Technology

[0002] In the field of enterprise big data analytics, researching how to accurately identify and differentiate different enterprise entities is particularly important, as it directly relates to the accuracy of supply chain management, risk assessment, and business decision-making. With the continuous expansion of data scale and the increasing complexity of enterprise relationships, entity disambiguation technology has become a key support for ensuring data quality and business insights, especially when dealing with multi-source heterogeneous data, its importance is self-evident.

[0003] However, existing methods often struggle to adapt to complex corporate relationship networks and dynamically changing business environments when dealing with disambiguation of corporate entities. Many technologies lack a comprehensive consideration of multiple relationships between companies, such as equity or guarantees, leading to incomplete identification results. More importantly, existing solutions often overlook the impact of events such as company name changes or business transformations when handling changes over time, causing breaks in data connections and consequently affecting the accuracy of the analysis.

[0004] Focusing on core technological challenges, dynamic changes over time become a pressing issue. Business entities may undergo name changes or business direction adjustments throughout their history, such as shifting from manufacturing to services. These changes can cause breaks in the existing relationship network over time. Furthermore, this disruption not only affects the identification of individual entities but also impacts the assessment of connections between related entities throughout the entire network. Imagine a core supplier changing its name due to business transformation; its historical connections with downstream companies may be severed. If these pre- and post-change relationships cannot be reconnected, its true influence in the supply chain cannot be accurately assessed, leading to biased risk assessments.

[0005] Therefore, how to effectively track the dynamic changes of enterprise entities in the time dimension and reconnect the relationship network that has been broken due to name or business adjustments has become a key issue in the research of adaptive multimodal entity disambiguation in enterprise big data. Summary of the Invention

[0006] To address the above issues, this invention proposes an adaptive multimodal entity disambiguation method and system for enterprise big data.

[0007] The present invention achieves the above objectives through the following technical solutions:

[0008] An adaptive multimodal entity disambiguation method for enterprise big data, the method comprising:

[0009] Extract equity relationship data and guarantee relationship data, construct an influence graph with corporate entities as nodes and relationship strength as edges, and use the PageRank algorithm to calculate node centrality to filter the core node set;

[0010] Starting from the core node set, the influence score is propagated along the edges of the influence graph. During the propagation process, the edge weights are adaptively adjusted by calling the exponential decay function according to the relationship strength to obtain the score distribution of each node.

[0011] Obtain the time series of the company's historical name changes and overlay it into the influence graph to identify and mark the relationship edge breaks caused by the historical name changes, and generate a list of broken paths;

[0012] For the list of fracture paths, the similarity between the nodes on both sides of the fracture edge is calculated in at least one preset feature dimension. The preset feature dimension includes the visual features corresponding to the enterprise text description, registration information, geographical location, and business license image. When the similarity exceeds a set threshold, the corresponding nodes are reconnected in the influence graph to obtain the repaired influence graph.

[0013] Based on the repaired influence map, the influence score propagation process is re-executed. By comparing the updated score differences between core nodes and edge nodes, a conflict intensity distribution characterizing the conflict in the evolution of enterprise relationships is generated.

[0014] The time series graph algorithm is used to analyze the changing trend of the conflict intensity distribution on the time axis. Combined with the business transformation event sequence of the enterprise, the impact of business transformation events on the score decay rate is tracked. The parameters of the exponential decay function are adaptively adjusted, and an optimized conflict intensity distribution is generated iteratively based on the adjusted parameters.

[0015] Based on the optimized conflict intensity distribution, the identity of the target enterprise entity is determined, and the identity determination result is used as the final output of the enterprise entity disambiguation.

[0016] As a preferred embodiment of the present invention, the equity relationship data is extracted from the business registration database of the enterprise multi-source database, and includes at least the shareholding ratio; the guarantee relationship data is extracted from the credit information disclosure database of the enterprise multi-source database, and includes at least the guarantee amount.

[0017] The construction of the influence map includes:

[0018] The strength of the equity relationship is calculated based on the shareholding ratio, and the strength of the guarantee relationship is calculated based on the amount of the guarantee.

[0019] By treating corporate entities as nodes and using the strength of equity relationships and guarantee relationships as the weights of the edges connecting the corresponding corporate entity nodes, a weighted directed graph is constructed as the initial influence graph.

[0020] For each node in the initial influence map, multi-dimensional features of the node are collected, including equity level depth, number of external guarantees, historical violation records and industry classification codes;

[0021] Input multi-dimensional features into a pre-trained abnormal node detection model and output the anomaly confidence score for each node;

[0022] Remove nodes with abnormal confidence levels exceeding a preset abnormal threshold from the initial influence graph to generate a filtered influence graph.

[0023] As a preferred embodiment of the present invention, the step of using the PageRank algorithm to calculate node centrality to filter the core node set includes:

[0024] A personalized PageRank algorithm with individualized node restart probabilities is used to calculate the PageRank score of each node in the filtered influence graph, which serves as an indicator of node centrality. For any node in the filtered influence graph... Restart probability Based on the number of years the enterprise has been in operation at this node The calculation formula is as follows: ,in It is a constant. The smallest positive number to prevent division by zero errors;

[0025] The personalized PageRank algorithm is executed iteratively until the PageRank scores of each node converge, resulting in a set of PageRank scores.

[0026] For each node in the PageRank score set, calculate its degree centrality, where degree centrality is the total number of edges connected to that node. Calculate a correction coefficient based on the node's degree centrality. For any node in the PageRank score set... Correction coefficient The calculation formula is ,in For nodes The degree centrality normalized value, These are preset weighting factors;

[0027] Node PageRank score multiplied by correction factor Generate corrected node influence scores;

[0028] Based on the corrected node influence score, the top N nodes are selected as candidate core nodes, where N is the preset number of candidate nodes.

[0029] Calculate the local density of each candidate core node in the feature space formed by multi-dimensional features. The local density represents the degree of aggregation of other candidate core nodes centered on the candidate core node within a preset distance threshold or within a specified number of nearest neighbors.

[0030] The node influence score of the candidate core node is multiplied by the corresponding local density to obtain the comprehensive influence index. The candidate core nodes are sorted in descending order according to the comprehensive influence index, and the top M nodes are selected to form the core node set, where M is the preset number of core nodes and M≤N.

[0031] As a preferred embodiment of the present invention, obtaining the score distribution of each node specifically includes:

[0032] For nodes in the core node set, the node influence score is used as the initial propagation value, and the initial propagation value of nodes that do not belong to the core node set is set to 0.

[0033] Perform multiple rounds of iterative propagation on the influence graph, with each iteration executing the following steps:

[0034] a. Nodes with a current influence score greater than 0 will be designated as the source nodes for propagation in this iteration;

[0035] b. Traverse each outgoing edge of each propagation source node and obtain the equity relationship strength or guarantee relationship strength of each outgoing edge as the basic edge weight. Choose the corresponding attenuation coefficient based on whether the outgoing edge is an equity relationship or a guarantee relationship. Based on the current iteration round According to the formula Calculate the adjusted edge weight in this iteration. Multiply the current influence score of the source node by the adjusted weight. The increment of the influence score transmitted along the edge to the target node is obtained;

[0036] c. Sum the incremental influence scores that each node receives from all incoming edges, and use this sum as the new influence score that the node receives in this iteration;

[0037] d. Calculate the variance of the newly added influence scores for all nodes in this round. If the variance is lower than the first preset threshold, increase the decay coefficient in the next iteration. The upper limit of the value range; if the variance is higher than the second preset threshold, the attenuation coefficient will be reduced in the next iteration. The lower limit of the range of values;

[0038] e. Add the newly gained influence points obtained in this round to the cumulative influence points of each node, and reset the current influence points of all propagation source nodes to 0;

[0039] Repeat steps a to e above until the cumulative influence score change of all nodes in two consecutive iterations is lower than the preset convergence threshold.

[0040] The cumulative influence scores of each node when the convergence condition is met are normalized to generate the score distribution of each node.

[0041] As a preferred embodiment of the present invention, the generation of the fracture path list specifically includes:

[0042] The historical name change records of the enterprise are extracted from the enterprise change record database of the enterprise multi-source database and sorted according to the effective time of the change to form the historical name change time series. Each record in the historical name change time series includes a historical name and the corresponding start and end time period.

[0043] Label each relationship edge in the influence graph with the start and end time periods of the relationship it represents, and iterate through each relationship edge in the influence graph to perform a breakage detection:

[0044] From the historical name change sequence of the nodes at both ends of the relationship edge, find the historical names that are valid within the start and end time period of the relationship edge. Compare the found historical names with the current names of the corresponding nodes in the influence graph. If the comparison results are inconsistent, or if it is determined from the enterprise unified identification code that the two nodes do not belong to the same entity after the name change, mark the relationship edge as a broken edge.

[0045] When there are multiple name records with overlapping time periods in the historical name change sequence of the same node, all relation edges related to that node within the overlapping time period are marked as broken edges.

[0046] In the influence map, starting from the marked fracture edge, the search proceeds along the direction of the edge, and the nodes connected through the fracture edge are connected in sequence to form a fracture path.

[0047] The priority score of each fracture path is calculated based on the cumulative influence score of each node in the fracture path, the number of fracture edges in the path, and the relationship type.

[0048] The fracture paths are sorted according to the priority scores to generate the fracture path list.

[0049] As a preferred embodiment of the present invention, obtaining the repaired influence map includes:

[0050] Perform cross-modal consistency verification on the feature vectors corresponding to each preset feature dimension of the nodes on both sides of the fracture edge. When the cross-modal consistency meets the preset consistency condition, continue to perform similarity calculation.

[0051] When multiple preset feature dimensions exist, the similarity of the feature vectors corresponding to each preset feature dimension is calculated, and the time decay function is applied to the similarity of each preset feature dimension based on the time difference of historical name change to generate time decay similarity. The relationship structure consistency factor is calculated based on the overlap of the adjacent node sets of the nodes on both sides of the fracture edge in the influence map. The time decay similarity and the relationship structure consistency factor are weighted and fused according to the preset fusion weight to generate the final comprehensive similarity.

[0052] When the final comprehensive similarity meets the set threshold, an edge connecting the nodes on both sides of the broken edge is established in the influence graph, and the original relationship strength corresponding to the broken edge is used as the edge weight of the connecting edge to update the connection structure of the influence graph to generate the repaired influence graph.

[0053] As a preferred embodiment of the present invention, the generation of the conflict intensity distribution characterizing the evolutionary conflict of enterprise relationships includes:

[0054] The influence score propagation process is re-executed on the repaired influence graph to obtain the repaired node score distribution.

[0055] The node score distribution before and after the repair is compared at the node level to obtain the difference in the updated score of each node.

[0056] Based on the comparison of the set of adjacent nodes of each node in the influence map before and after the repair, the structural perturbation factor characterizing the degree of change in the node's adjacency structure is calculated.

[0057] According to the preset weighting rules, the updated score difference of the nodes and the structural disturbance factor are weighted and combined to obtain the conflict intensity index of each node.

[0058] Based on the core node set, the nodes are divided into core nodes and edge nodes, and the conflict intensity index distribution of core nodes and edge nodes are statistically analyzed respectively.

[0059] A conflict intensity distribution is generated based on the difference in conflict intensity index distribution between core nodes and edge nodes.

[0060] As a preferred embodiment of the present invention, the generation of the optimized conflict intensity distribution includes:

[0061] A time series diagram of conflict intensity is constructed based on the conflict intensity distribution at different times, and the change in conflict intensity between adjacent time windows is calculated.

[0062] Extract business transformation event sequences from the enterprise's multi-source database and align the occurrence times of the business transformation events with the conflict intensity time sequence diagram;

[0063] Based on the type of business transformation event, determine the corresponding lag window for the event, track the change in conflict intensity within the lag window, and calculate the impact coefficient of the business transformation event on the score decay rate.

[0064] When multiple business transformation events occur within adjacent time windows, an event superposition factor is constructed based on the superposition relationship between the events, and the event superposition factor is combined with the impact coefficient of each business transformation event to form a comprehensive impact coefficient.

[0065] The decay parameters of the exponential decay function are adaptively adjusted based on the comprehensive influence coefficient, including adjusting the edge-level decay parameters separately for edges whose conflict intensity exceeds a preset threshold.

[0066] The conflict intensity distribution is regenerated based on the adjusted attenuation parameters, and the above process is repeated if the conflict intensity distribution does not reach the preset convergence condition in order to obtain an optimized conflict intensity distribution.

[0067] As a preferred embodiment of the present invention, the identification of the target enterprise entity based on the optimized conflict intensity distribution includes:

[0068] Construct a temporal distribution vector of the conflict intensity of the target enterprise entity across multiple time windows based on the optimized conflict intensity distribution;

[0069] The similarity of the temporal distribution vectors of conflict intensity of different target nodes that are suspected to belong to the same enterprise entity is calculated, and the consistency index of the direction of change of conflict intensity between the two is calculated.

[0070] Peak time identification is performed on the temporal distribution vector of conflict intensity of different target nodes, and the overlap of peak times is compared to obtain peak alignment.

[0071] The structural perturbation factors of different target nodes are compared based on the optimized influence map to obtain the differences in structural perturbation.

[0072] Based on the temporal similarity of conflict intensity, the consistency of the direction of conflict intensity change, the peak alignment, the difference of structural disturbance, and the similarity on the preset feature dimensions, a weighted fusion is performed to generate the identity determination score and the corresponding identity confidence of the target enterprise entity.

[0073] Based on the comparison result between the identity determination score or the identity confidence level and the preset threshold, it is determined whether the target enterprise entity belongs to the same entity, and the identity determination result is used as the final output of enterprise entity disambiguation.

[0074] An adaptive multimodal entity disambiguation system for enterprise big data, the system comprising:

[0075] The data extraction and node filtering module is used to extract equity relationship data and guarantee relationship data. It constructs an influence graph with enterprise entities as nodes and relationship strength as edges, and uses the PageRank algorithm to calculate node centrality to filter the core node set.

[0076] The influence propagation module is used to propagate influence scores along the edges of the influence graph, starting from the core node set. During the propagation process, the edge weights are adaptively adjusted by calling an exponential decay function based on the relationship strength to obtain the score distribution of each node.

[0077] The name change processing module is used to obtain the historical name change time series of enterprises and overlay it into the influence map, identify and mark the relationship edge break caused by historical name changes, and generate a list of break paths.

[0078] The graph repair module is used to calculate the similarity of nodes on both sides of the fracture edge in at least one preset feature dimension for the fracture path list. When the similarity exceeds a set threshold, the corresponding nodes are reconnected in the influence graph to obtain the repaired influence graph.

[0079] The conflict intensity generation module is used to re-execute the influence score propagation process based on the repaired influence map, and generate a conflict intensity distribution that characterizes the conflict in the evolution of enterprise relationships by comparing the updated scores of core nodes and edge nodes.

[0080] The adaptive adjustment module is used to analyze the changing trend of the conflict intensity distribution on the time axis using a time series graph algorithm, and combined with the enterprise's business transformation event sequence, to track the impact of business transformation events on the score decay rate, adaptively adjust the parameters of the exponential decay function, and iteratively generate an optimized conflict intensity distribution based on the adjusted parameters.

[0081] The entity identity determination module is used to determine the identity of the target enterprise entity based on the optimized conflict intensity distribution, and the identity determination result is used as the final output of enterprise entity disambiguation.

[0082] The beneficial effects of this invention are as follows: By constructing an influence graph based on equity and guarantee relationships, and using the PageRank algorithm to select core nodes, it can highlight key influential enterprise entities in the enterprise network. An exponential decay function is introduced during the dynamic influence propagation process, allowing the propagation intensity to adaptively adjust according to relationship strength, thus more accurately depicting the true impact path of inter-enterprise relationships. By introducing historical name change sequences to label relationship break edges and using multimodal feature similarity to repair the break paths, the graph can maintain structural coherence even under conditions of frequent entity name changes. Furthermore, by comparing the differences in node scores before and after repair to construct a conflict intensity distribution, and combining it with business transformation event sequences to optimize the temporal propagation model, the conflict representation can more accurately reflect the true meaning of enterprise relationship evolution. The optimized conflict intensity distribution is used to determine the identity of enterprise entities. By combining temporal similarity, structural perturbation features, and multimodal feature similarity, a joint judgment from multiple dimensions—behavioral evolution, structural changes, and attribute information—is achieved, improving the stability and accuracy of entity disambiguation results in complex scenarios. Attached Figure Description

[0083] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the modular structure of the system in an embodiment of the present invention. Detailed Implementation

[0084] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0085] like Figure 1 As shown, this is an embodiment of the present invention. This embodiment provides an adaptive multimodal entity disambiguation method for enterprise big data. It is used to accurately determine whether two or more enterprise entities belong to the same entity when the enterprise entity has undergone historical name changes, complex equity relationships, guarantee relationships, or changes in relationship structure caused by business events. It is applicable to enterprise multi-source data scenarios such as business registration data, credit disclosure data, and large-scale commercial databases.

[0086] This method mainly includes the following steps:

[0087] S1: Extract equity relationship data and guarantee relationship data, construct an influence graph with enterprise entities as nodes and relationship strength as edges, and use the PageRank algorithm to calculate node centrality to filter the core node set.

[0088] In practical applications, capital and credit relationships between business entities are recorded by different types of data sources. To construct a network of relationships between enterprises, this embodiment first reads equity relationship data and guarantee relationship data from a multi-source enterprise database. Specifically, the equity relationship data extracted from the business registration database includes: investor identifier, invested entity identifier, and shareholding ratio. When reading this data, the system converts the shareholding ratio field into a standardized decimal form for subsequent use in constructing a unified form of relationship strength values. Simultaneously, the guarantee relationship data extracted from the credit information disclosure database includes: guarantor identifier, guaranteed entity identifier, and guarantee amount. Since guarantee amounts may vary by orders of magnitude between different enterprises, this embodiment uses logarithmic normalization to preprocess the guarantee amounts to enhance stability during graph structure calculations.

[0089] To ensure the comparability of weights for different types of relationships, this embodiment uses a unified strength calculation model for both types of relationships. For equity relationships, the shareholding ratio is first obtained and then calculated according to the formula... Calculating the strength of equity relationship ,in For equity ratio, The weighting coefficient for equity relationships is used to adjust the relative influence of equity edges in the overall graph; for guarantee relationships, the formula is used. Calculating the strength of the guarantee relationship ,in For the amount of the guarantee, This is the guarantee relationship weighting coefficient, used to control the effective contribution of the guarantee edge. In this embodiment, and All parameters can be set to fixed values ​​according to specific business needs, or dynamically adjusted based on historical statistical results. Those skilled in the art can freely set the calculation parameters according to the data source used, without affecting the effectiveness of this method.

[0090] After completing the above strength calculations, each company is treated as a node, and the equity or guarantee relationship between any two companies is added to the graph structure as a weighted directed edge. The direction of the equity relationship edge is "investor → investee," and the direction of the guarantee relationship edge is "guarantor → guaranteed entity." The weights of the edges are the strength values ​​calculated above. This embodiment can utilize Python's NetworkX, Spark GraphX, or Neo4j graph database to construct this weighted directed graph. The completed graph is the initial influence graph, representing the basic relationship structure between companies.

[0091] To identify potentially abnormal enterprise nodes in the data, this embodiment collects the following four types of features for each node in the initial influence graph:

[0092] Equity hierarchy depth: The depth of a company's hierarchy in the equity chain is calculated using graph-based depth-first traversal (DFS).

[0093] Number of external guarantees: This counts the number of guarantee records in which the company acts as the guarantor.

[0094] Historical violation record marking: Generate binary or multi-dimensional violation markings by reading public information such as corporate administrative penalties and operational abnormalities;

[0095] Industry classification coding: Based on the enterprise's business scope field, it is mapped to the standard industry classification system (such as GB / T4754).

[0096] All of the above multi-dimensional features can be obtained directly through database queries or graph algorithms.

[0097] To avoid erroneous data or extremely noisy nodes affecting the subsequent selection of center nodes, this embodiment inputs the above multi-dimensional features into a pre-trained anomaly detection model. This model can employ Isolation Forest, an AutoEncoder-based reconstruction error model, or a graph neural network-based anomaly detection model. The model outputs an anomaly confidence value for each node. For nodes with anomaly confidence values ​​exceeding a preset threshold, they are removed from the graph, and their associated edges are deleted. The filtered influence graph more accurately reflects the true relationship structure between enterprises, providing more reliable input data for subsequent PageRank calculations. In this embodiment, the threshold can be determined based on historical data statistics.

[0098] The core idea of ​​the PageRank algorithm is that the importance of a node depends on the importance of the nodes pointing to it. This embodiment uses Personalized PageRank to calculate node centrality. To make the node centrality calculation more closely reflect the real-world characteristics of enterprise data, this embodiment does not use the traditional PageRank with a uniform restart probability, but instead configures a personalized restart probability for each node based on the enterprise's operating years. Let the operating years of the enterprise corresponding to the node be 1. Restart probability According to the formula Calculation, where This is a constant used to control the overall level of the restart probability; The smallest positive number to prevent division by zero errors. Companies with longer operating histories tend to be more stable and therefore have a higher probability of restarting.

[0099] A personalized PageRank algorithm with an "operational stability bias" was constructed using the above methods. This algorithm gives large, long-term operating enterprises higher centrality, which better aligns with the actual business logic of enterprise relationship networks. Subsequently, the personalized PageRank operation is performed on the filtered influence graph, iteratively updating node centrality until convergence, resulting in a set of PageRank scores containing all nodes.

[0100] Since some nodes may have an inflated PageRank due to a large number of connections, this embodiment corrects the PageRank score for degree centrality. First, the degree centrality of each node is calculated based on the filtered influence graph, which can be achieved by counting the total number of connections to each node; then, according to the formula... Calculate the correction factor ,in For nodes The degree centrality normalized value, The preset weighting factor controls the degree of correction. The PageRank score of a node is multiplied by the corresponding correction coefficient to obtain the corrected node influence score. This correction process can reduce the structural bias of some highly connected nodes and improve the stability of the final selection results.

[0101] To avoid excessive computational overhead or interference from noisy nodes that might result from directly calculating subsequent density based on all nodes, this embodiment employs a first-level screening mechanism, prioritizing a subset of nodes with higher influence from the entire set of nodes as candidate core nodes. The specific method is as follows:

[0102] Sort all nodes by their correction influence scores from highest to lowest;

[0103] The top N nodes are selected to form a set of candidate core nodes, where N is the preset number of candidates set by the system.

[0104] In this embodiment, N can be determined based on the size of the influence graph. For example, if the graph has 10,000 nodes, N can be set to 200; if the graph has more than 100,000 nodes, N can be set proportionally to 1% to 5% of the total number of nodes. This setting ensures that the candidate set covers the main high-influence nodes in the enterprise network, while effectively eliminating long-tail noise nodes, providing a cleaner input space for subsequent density calculations. In actual implementation, N can be set through system configuration files, database parameters, or runtime input, and a fixed value is not required, making this method flexible enough to adapt to different application scenarios.

[0105] For the selected set of candidate core nodes, this embodiment further calculates the local density of nodes in the multi-dimensional feature space to reflect the degree of node clustering in the feature space. The implementation steps are as follows:

[0106] Construct a feature vector for each node in the candidate set. Features include, but are not limited to: equity level depth, number of external guarantees, historical violation record markings, industry classification codes, etc.

[0107] The k-nearest neighbors (kNN) method is used to find the k nearest neighbors of each node in the feature space;

[0108] The number of k neighbors or their average distance is used as the local density value of a node.

[0109] In practical scenarios, k can be set to an integer value between 5 and 20, depending on the size of the candidate set. Those skilled in the art can implement this density calculation step using sklearn's NearestNeighbors function.

[0110] By introducing local density, this embodiment can identify enterprise nodes that simultaneously possess high influence and high feature clustering, ensuring that the core node selection results not only conform to structural centrality but also feature consistency, thereby enhancing the interpretability and robustness of the model.

[0111] To simultaneously integrate the corrected node influence score with its local density in the candidate set, this embodiment constructs a comprehensive influence index. : ,in The node influence score for candidate core nodes. This represents the local density value of the node in the feature space.

[0112] Candidate core nodes are sorted in descending order based on comprehensive influence indicators, and the top M nodes are selected to form a core node set, where M is the preset number of core nodes and M ≤ N. For example, typical values ​​for M are as follows:

[0113] If N=200, we can set M =20~50;

[0114] If the map is large, M can be set within the range of 10% to 30% of N.

[0115] The final set of core nodes will serve as the set of source nodes in the subsequent process of influence propagation, providing the most representative structural support for the entire entity disambiguation system.

[0116] S2: Starting from the core node set, propagate the influence score along the edges in the influence graph. During the propagation process, the edge weights are adaptively adjusted by calling the exponential decay function according to the relationship strength to obtain the score distribution of each node.

[0117] In one specific embodiment, step S2 includes:

[0118] S21: For each node in the core node set, its initial influence score is set to the corrected node influence score; for nodes that do not belong to the core node set, the initial influence score is set to zero.

[0119] S22: Initiate a multi-round iterative propagation process on the influence graph, performing the following steps in each iteration:

[0120] a. In the current iteration, traverse all nodes with non-zero influence scores and regard these nodes as propagation source nodes. Only propagation source nodes will transmit influence outward.

[0121] b. For each source node, read all outgoing edges of that node and process each outgoing edge sequentially. For each outgoing edge, obtain the equity relationship strength or guarantee relationship strength of each outgoing edge as the basic edge weight. Choose the corresponding attenuation coefficient based on whether the outgoing edge is an equity relationship or a guarantee relationship. Let the current iteration round be... (Incrementing from 1), according to the formula Calculate the adjusted edge weight in this iteration. Multiply the current influence score of the source node by the adjusted weight. The increment of the influence score transmitted along the edge to the target node is obtained;

[0122] c. The increment of the influence score obtained by each node from all incoming edges is accumulated and used as the new influence score obtained by the node in this iteration. In this embodiment, the new influence score is also used as the initial amount of the current influence score of the target node in the next iteration, and is used to continue to propagate outward in the next iteration.

[0123] d. Calculate the variance of the newly added influence scores for all nodes in this round. If the variance is lower than the first preset threshold, increase the decay coefficient in the next iteration. The upper limit of the value range; if the variance is higher than the second preset threshold, the attenuation coefficient will be reduced in the next iteration. The lower limit of the value range; the first preset threshold and the second preset threshold can be determined by historical data statistics or offline experiments, and can be configured by those skilled in the art according to specific application scenarios;

[0124] e. Add the newly gained influence score obtained in this round to the cumulative influence score of each node, and reset the current influence score of all propagation source nodes to 0, so as to indicate that the influence of these nodes in this iteration has been completely propagated outward;

[0125] S23: Repeat steps a~e above until the cumulative influence score change of all nodes in two consecutive iterations is lower than the preset convergence threshold, or the number of iterations reaches the preset maximum number of iterations, then stop the propagation process; the convergence threshold and the maximum number of iterations can be preset according to factors such as the graph size and business requirements.

[0126] S24: After the propagation process ends, the system performs normalization processing on the cumulative influence score of each node. For example, the cumulative influence score of all nodes is normalized to a real number in the range of 0 to 1. The normalized value constitutes the score distribution of each node in the influence map, providing input for subsequent enterprise relationship evolution conflict calculation, conflict intensity time series analysis, and enterprise entity identity determination.

[0127] S3: Obtain the time series of the company's historical name changes and overlay it into the influence graph. Identify and mark the relationship edge breaks caused by the historical name changes and generate a list of broken paths.

[0128] In one specific implementation, the process of generating the fracture path list includes the following steps:

[0129] S31: Extract all historical name change records of the target company from the enterprise change record database of the enterprise multi-source database. Each change record includes the name before or after the change, the effective time of the change, and some records also include information such as the change announcement time and the reason for the change. Arrange all name records in order of change effective time, and use the time point of two adjacent change records as the boundary. Determine the effective time period corresponding to each name as the time from the effective time of the name to the moment before the effective time of the next name, thereby constructing a complete historical name change time sequence.

[0130] S32: For each relationship edge in the influence graph, the system writes the effective time period of the relationship edge into the edge attribute based on the date the enterprise relationship was created and the date it ended, including:

[0131] The start date of the relationship, such as the effective date of the equity relationship or the effective date of the guarantee contract;

[0132] The end date of the relationship, such as the date of equity transfer or the date of release of guarantee liability;

[0133] This valid time period is used to verify whether the node name has changed during the period the relationship exists.

[0134] S33: Traverse each relation edge in the influence graph and perform breakage detection as follows:

[0135] For the enterprise nodes at both ends of the relationship edge, the system searches for the names that were valid within the effective time period of the relationship edge in their historical name change time series, and compares the found historical names with the current names. If the historical names and current names are inconsistent, the system further checks the unified social credit codes of the corresponding enterprises. When the unified social credit codes are inconsistent, it indicates that the node name change has caused the entity to change, and the system marks the relationship edge as a broken edge.

[0136] In some scenarios, a company's historical name change records may have overlapping valid time periods. If there is a name conflict (name overlap) or multiple name records with overlapping valid time periods for the same node, all relationship edges associated with that node during the overlapping period are marked as broken edges to ensure that the relationship graph structure reflects the actual entity evolution.

[0137] S34: After obtaining all broken edges, perform a graph traversal operation starting from the broken edges. Graph traversal can be implemented using either depth-first search or breadth-first search. The specific process is as follows:

[0138] Use the starting node of the broken edge as the starting node of the path;

[0139] Move along the direction of the fracture edge to the target node;

[0140] If the target node continues to point to other broken edges, then continue to extend the path along the broken edges;

[0141] The expansion terminates when a node that is no longer connected to the broken edge is reached, forming a complete broken path.

[0142] Perform path expansion on all fracture edges according to the above steps to obtain a set containing multiple fracture paths.

[0143] S35: In order to determine the fracture paths that need to be repaired or analyzed in subsequent processing, this embodiment calculates a priority score for each fracture path. The scoring indicators include: the statistical value of the cumulative influence score of each node in the path, the number of fracture edges in the path, and the relationship type corresponding to the fracture edge (such as the priority of equity edge can be higher than that of guarantee edge or external contact edge).

[0144] Based on the above scoring indicators, a weighted scoring formula is constructed to generate a priority score for each fracture path;

[0145] S36: Sort all fracture paths in descending order based on priority scores, and organize the sorting results into a fracture path list. This list is used to determine the priority paths in subsequent fracture repair steps and serves as an important input structure for the fusion of historical name changes and enterprise relationship graphs.

[0146] S4: For the list of broken paths, calculate the similarity between the nodes on both sides of the broken edge in at least one preset feature dimension; when the similarity exceeds the set threshold, reconnect the corresponding nodes in the influence graph to obtain the repaired influence graph.

[0147] After generating a list of fracture paths and identifying fracture relationships, this embodiment performs multimodal similarity calculations on the preset feature dimensions of the nodes on both sides of the fracture edge, and performs structural repair on the influence graph based on the similarity determination results. This repair process aims to solve the graph fracture problem caused by factors such as changes in enterprise historical names, registration migrations, enterprise splits or mergers, so that the influence graph can accurately reflect the real relationships between enterprise entities.

[0148] In one specific implementation, the process of repairing the influence map includes the following steps:

[0149] For each fracture edge in the fracture path list, data corresponding to the nodes on both sides of the fracture edge in a preset feature dimension is read from the enterprise's multi-source database, and this data is converted into a corresponding feature vector. The preset feature dimensions include, but are not limited to, text vectors corresponding to the enterprise's text description, structured feature vectors corresponding to registration information, geocoding vectors corresponding to geographical locations, and visual feature vectors corresponding to business license images.

[0150] Cross-modal consistency checks are performed on the feature vectors corresponding to the nodes on both sides of the break edge in each preset feature dimension. Cross-modal consistency checks are used to detect potential conflicts between different modalities. For example, when the text descriptions of two nodes are highly similar, but their registration information or geographical location features differ significantly, it can be determined that the node pair has a modal conflict. A difference index is calculated for each modal difference and compared with preset consistency conditions. When all modal differences are within the range allowed by the consistency conditions, the cross-modal consistency requirement is met, and the similarity calculation step is initiated; when any modal difference exceeds the consistency condition range, the subsequent reconnection process is not continued.

[0151] Provided that cross-modal consistency is met, feature similarity calculation is performed on the nodes on both sides of the fracture edge when multiple preset feature dimensions exist. Specifically, when multiple preset feature dimensions exist, the system calculates the similarity between text feature vectors, registration information feature vectors, geographic location feature vectors, and visual feature vectors respectively. Text vector similarity can be calculated based on cosine similarity; structured registration information vectors can use Euclidean distance or cosine similarity; geographic location feature vectors can be similarity transformed based on geographic coordinates using Haversine distance; and visual feature vectors can be calculated based on feature space similarity extracted by convolutional neural networks or visual models.

[0152] Based on the time difference of historical name changes of nodes on both sides of the break, a time decay function is applied to the similarity of each preset feature dimension to obtain a time decay similarity. The time decay function assigns lower weights to earlier historical features according to a preset decay coefficient, ensuring that features closer to the break point contribute more to the similarity calculation. Based on the influence graph structure, a relationship structure consistency factor is calculated for the nodes on both sides of the break. This factor is calculated based on the degree of overlap between adjacent node sets in the influence graph; a higher consistency factor indicates a similar relationship neighborhood in the upstream or downstream relationship structure, and vice versa. The time decay similarity and the relationship structure consistency factor are then weighted and fused according to preset fusion weights to generate the final comprehensive similarity.

[0153] The final overall similarity is compared with a pre-set similarity threshold. When the final overall similarity meets the threshold, the nodes on both sides of the broken edge are identified as node pairs that should have their association re-established. For node pairs that need to be reconnected, an edge connecting the node pairs is created in the influence graph. When creating the edge, the original relationship strength corresponding to the broken edge is used as the edge weight of the new connection edge, so that the newly added connection edge can accurately reflect the strength of the historical relationship. For directional relationships, the connection edge is set according to the direction of the broken edge; for undirected edges, it can be set as a bidirectional connection.

[0154] After establishing new connections, update the relevant data structures in the influence graph, including the node connection table, edge weight table, and path index structure, so that the repaired influence graph can be correctly used in subsequent steps such as influence score propagation and relationship evolution conflict analysis.

[0155] Through the above-mentioned map restoration steps, in the presence of multiple preset feature dimensions, text, registration information, geographical location and visual features can be fully utilized, and time factors and structural consistency can be combined to achieve accurate identification and restoration of broken relationships, so that the influence map maintains structural continuity and entity consistency.

[0156] S5: Based on the repaired influence map, the influence score propagation process is re-executed. By comparing the updated score differences between core nodes and peripheral nodes, a conflict intensity distribution characterizing the conflict in the evolution of enterprise relationships is generated. This conflict intensity distribution is used to identify whether changes in enterprise structure reflect potential entity evolution conflicts, providing a data foundation for subsequent time series analysis and identity determination.

[0157] In one specific embodiment, step S5 includes:

[0158] S51: Perform the influence score propagation process based on the repaired influence graph to obtain the repaired node influence scores. ,in The node is assigned a number, and its influence score before repair is recorded as follows: Store the two sets of values ​​above;

[0159] S52: Calculate the updated score difference using the node score difference formula. , among which when This indicates an increase in the node's influence. The time indicates a decrease in node influence; this score difference reflects the direct effect of repair operations on the node.

[0160] S53: To measure the degree of change in node adjacency relationships before and after repair, extract the nodes respectively. Let the set of adjacent nodes before and after the repair be denoted as the set of adjacent nodes before the repair. After repair, the adjacency set is Node structure disturbance factor The calculation formula is: ,in This is a symmetric difference operation, representing the change in scale of a set; a small constant is added to the denominator. To prevent division by zero, The larger the value, the more drastic the structural change;

[0161] S54: According to the preset weighting rules, adjust the updated scores of the nodes. With structural disturbance factor By combining these, a node-level conflict intensity index can be formed. : ,in , These are system parameters that can be set based on industry data and experience. Describe the changes in influence. Describe structural changes;

[0162] This weighting method can be flexibly adjusted based on the functional characteristics of nodes in the graph, and can effectively reflect the overall disturbance caused by the repair of the enterprise's graph structure.

[0163] S55: Based on the core node set, nodes are divided into core nodes. and edge nodes The conflict intensity of the two types of nodes is calculated separately:

[0164] Mean conflict intensity of core nodes: ;

[0165] Mean edge node conflict intensity: ;

[0166] Distribution indicators such as median, variance, and quantiles can also be calculated as needed, but are not limited to these.

[0167] S56: Generate the overall conflict intensity distribution based on the conflict intensity difference between core nodes and edge nodes. : This conflict intensity value reflects whether the repair operation has led to a significant inconsistency between the core structure and the peripheral structure, thus providing a quantifiable description of the conflict in the relationship evolution; it can also output a list of node-level conflict intensities. This results in a multi-level conflict intensity distribution consisting of "node level + structure level + classification level".

[0168] S6: The time series graph algorithm is used to analyze the changing trend of conflict intensity distribution on the time axis. Combined with the business transformation event sequence of the enterprise, the impact of business transformation events on the score decay rate is tracked. The parameters of the exponential decay function are adaptively adjusted, and the optimized conflict intensity distribution is generated iteratively based on the adjusted parameters. This can reflect the dynamic relationship between the enterprise's graph structure and business activities, making the conflict analysis results more consistent with the real evolution process of the enterprise's entity relationships.

[0169] In one embodiment, step S6 includes:

[0170] S61: According to time window The conflict intensity is processed by time series analysis, and the overall conflict intensity is calculated in each time window, for example... ,in For nodes In the time window The intensity of the conflict The number of nodes;

[0171] The rate of change of conflict intensity between adjacent time windows is expressed as ;

[0172] Using time windows as nodes and the rate of change of conflict intensity as edge weights, a time series graph of conflict intensity is constructed. This time series graph is used to observe the trend of conflict intensity over time.

[0173] S62: Extract business transformation event sequences from the enterprise's multi-source database, including adjustments to main business, changes in business location, changes in shareholding structure, major investments, or major asset disposals. Each event includes: the time of occurrence. Event Type Event intensity (such as investment amount, scope of business change, etc.);

[0174] Align events to their corresponding time windows based on when they occurred. And record the set of all events that occur within that time window. .

[0175] S63: Different types of events often have a lag effect on the enterprise relationship structure. Therefore, the system sets a lag window for each type of event. This indicates that the impact of the event may occur after the event has taken place. The changes in conflict intensity are gradually revealed within a time window; the changes are tracked within the lag window corresponding to the event. ,when When the value is large, it indicates that the event has a significant impact on the evolution of the relationship;

[0176] Calculate the event impact coefficient based on conflict variations and event intensity. : ,in As a regulating factor, To prevent the use of tiny constants with a denominator of 0;

[0177] S64: When multiple business transformation events occur within adjacent time windows, construct an event superposition factor based on the superposition relationship between the events. This indicates the amplification effect of interactions between events: ,in The correlation matrix is ​​set according to the event type (e.g., there is a strong correlation between adjustments to the main business and changes in management); if the event types are highly correlated, the amplification effect is more obvious.

[0178] The comprehensive impact coefficient is defined as: ,in For superimposed adjustment factors;

[0179] S65: Adaptively adjust the decay parameters of the exponential decay function based on the comprehensive influence coefficient, including adjusting the edge-level decay parameters individually for edges whose conflict intensity exceeds a preset threshold; edge-level adjustment allows specific relationships to decay faster or slower in the propagation of influence, which helps improve the accuracy of conflict evolution modeling.

[0180] S66: Regenerate the conflict intensity distribution based on the adjusted attenuation parameters, and repeat the above process if the conflict intensity distribution does not reach the preset convergence condition, so as to obtain an optimized conflict intensity distribution.

[0181] S7: Based on the optimized conflict intensity distribution, the identity determination of the target enterprise entity is performed, and the identity determination result is used as the final output of the enterprise entity disambiguation.

[0182] Furthermore, based on the optimized conflict intensity distribution generated in step S6, the conflict intensity index values ​​of the target enterprise entity nodes are extracted over multiple time windows, and the conflict intensity time-series distribution vector is defined as follows: ,in For the target enterprise entity node, For entities in the time window The intensity of conflict, this vector reflects the temporal behavioral characteristics of the evolution of firm relationships.

[0183] For two target nodes that are suspected to belong to the same entity and Calculate the temporal similarity of their conflict intensity, for example, using time-weighted cosine similarity. Time weight Used to enhance sensitivity to critical evolutionary moments.

[0184] Compare the directions of change in the conflict intensity of the two entities, based on the definition of direction:

[0185] ;

[0186] Directional consistency is defined as: ,in This indicates a consistency indicator function; high directional consistency indicates that the relationship between the two entities is evolving in a synchronized manner.

[0187] Peak detection is performed on the temporal distribution vector of conflict intensity, and the peak time sets of the two entities are extracted respectively, denoted as... and ;

[0188] Peak alignment is calculated as follows: ;

[0189] Peak alignment is used to measure whether the high-sensitivity period of conflict is synchronized, and it is an important behavioral characteristic of the same entity.

[0190] The structural perturbation difference between the two target nodes is calculated. ;

[0191] Multimodal feature similarity is calculated based on preset feature dimensions (text description, registration information, geographical location, and visual features of the business license). It is used to reflect the identity at the level of enterprise entity attributes;

[0192] The above indicators are weighted and combined according to preset fusion weights to form a comprehensive identity judgment score:

[0193] ;

[0194] Furthermore, an identity confidence level is constructed based on the judgment score. ,in This is a normalization function, such as the sigmoid function.

[0195] Compare the identity determination score or identity confidence level with a preset threshold: That is, to determine whether two entities belong to the same enterprise entity, the final determination result is the final output of enterprise entity disambiguation.

[0196] like Figure 2 As shown, another embodiment of the present invention provides an adaptive multimodal entity disambiguation system for enterprise big data, comprising:

[0197] The data extraction and node filtering module is used to extract equity relationship data and guarantee relationship data. It constructs an influence graph with enterprise entities as nodes and relationship strength as edges, and uses the PageRank algorithm to calculate node centrality to filter the core node set.

[0198] The influence propagation module is used to propagate influence scores along the edges of the influence graph, starting from the core node set. During the propagation process, the edge weights are adaptively adjusted by calling an exponential decay function based on the strength of the relationship, so as to obtain the score distribution of each node.

[0199] The name change processing module is used to obtain the time series of the company's historical name changes and overlay it into the influence graph to identify and mark the relationship edge breaks caused by historical name changes and generate a list of broken paths.

[0200] The graph repair module is used to calculate the similarity of nodes on both sides of a broken edge in at least one preset feature dimension for a list of broken paths. When the similarity exceeds a set threshold, the corresponding nodes are reconnected in the influence graph to obtain the repaired influence graph.

[0201] The conflict intensity generation module is used to re-execute the influence score propagation process based on the repaired influence map. By comparing the updated score differences between core nodes and edge nodes, it generates a conflict intensity distribution that characterizes the conflict in the evolution of enterprise relationships.

[0202] The adaptive adjustment module is used to analyze the changing trend of conflict intensity distribution on the time axis using a time series graph algorithm, and combined with the enterprise's business transformation event sequence, to track the impact of business transformation events on the score decay rate, adaptively adjust the parameters of the exponential decay function, and iteratively generate an optimized conflict intensity distribution based on the adjusted parameters.

[0203] The entity identity determination module is used to determine the identity of target enterprise entities based on the optimized conflict intensity distribution, and the identity determination result is used as the final output of enterprise entity disambiguation.

[0204] In summary, this invention can achieve more accurate and reliable entity disambiguation in enterprise big data environments, significantly improving the integrity of the graph structure, the ability to express temporal relationships, and the performance of identity determination.

[0205] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An adaptive multimodal entity disambiguation method for enterprise big data, characterized in that, The method includes: Extract equity relationship data and guarantee relationship data, construct an influence graph with corporate entities as nodes and relationship strength as edges, and use the PageRank algorithm to calculate node centrality to filter the core node set; Starting from the core node set, the influence score is propagated along the edges of the influence graph. During the propagation process, the edge weights are adaptively adjusted by calling the exponential decay function according to the relationship strength to obtain the score distribution of each node. Obtain the time series of the company's historical name changes and overlay it into the influence graph to identify and mark the relationship edge breaks caused by the historical name changes, and generate a list of broken paths; For the list of fracture paths, the similarity between the nodes on both sides of the fracture edge is calculated in at least one preset feature dimension. The preset feature dimension includes the visual features corresponding to the enterprise text description, registration information, geographical location, and business license image. When the similarity exceeds a set threshold, the corresponding nodes are reconnected in the influence graph to obtain the repaired influence graph. Based on the repaired influence map, the influence score propagation process is re-executed. By comparing the updated score differences between core nodes and edge nodes, a conflict intensity distribution characterizing the conflict in the evolution of enterprise relationships is generated. The time series graph algorithm is used to analyze the changing trend of the conflict intensity distribution on the time axis. Combined with the business transformation event sequence of the enterprise, the impact of business transformation events on the score decay rate is tracked. The parameters of the exponential decay function are adaptively adjusted, and an optimized conflict intensity distribution is generated iteratively based on the adjusted parameters. Based on the optimized conflict intensity distribution, the identity of the target enterprise entity is determined, and the identity determination result is used as the final output of the enterprise entity disambiguation.

2. The adaptive multimodal entity disambiguation method for enterprise big data according to claim 1, characterized in that, The equity relationship data is extracted from the business registration database of the enterprise multi-source database, and includes at least the shareholding ratio; The guarantee relationship data is extracted from the enterprise's multi-source database of credit information disclosure, and includes at least the guarantee amount; The construction of the influence map includes: The strength of the equity relationship is calculated based on the shareholding ratio, and the strength of the guarantee relationship is calculated based on the amount of the guarantee. By treating corporate entities as nodes and using the strength of equity relationships and guarantee relationships as the weights of the edges connecting the corresponding corporate entity nodes, a weighted directed graph is constructed as the initial influence graph. For each node in the initial influence map, multi-dimensional features of the node are collected, including equity level depth, number of external guarantees, historical violation records and industry classification codes; Input multi-dimensional features into a pre-trained abnormal node detection model and output the anomaly confidence score for each node; Remove nodes with abnormal confidence levels exceeding a preset abnormal threshold from the initial influence graph to generate a filtered influence graph.

3. The adaptive multimodal entity disambiguation method for enterprise big data according to claim 2, characterized in that, The step of using the PageRank algorithm to calculate node centrality to filter the core node set includes: A personalized PageRank algorithm with individualized node restart probabilities is used to calculate the PageRank score of each node in the filtered influence graph, which serves as an indicator of node centrality. For any node in the filtered influence graph... Restart probability Based on the number of years the enterprise has been in operation at this node The calculation formula is as follows: ,in It is a constant. The smallest positive number to prevent division by zero errors; The personalized PageRank algorithm is executed iteratively until the PageRank scores of each node converge, resulting in a set of PageRank scores. For each node in the PageRank score set, calculate its degree centrality, where degree centrality is the total number of edges connected to that node. Calculate a correction coefficient based on the node's degree centrality. For any node in the PageRank score set... Correction coefficient The calculation formula is ,in For nodes The degree centrality normalized value, These are preset weighting factors; Node PageRank score multiplied by correction factor Generate corrected node influence scores; Based on the corrected node influence score, the top N nodes are selected as candidate core nodes, where N is the preset number of candidate nodes. Calculate the local density of each candidate core node in the feature space formed by multi-dimensional features. The local density represents the degree of aggregation of other candidate core nodes centered on the candidate core node within a preset distance threshold or within a specified number of nearest neighbors. The node influence score of the candidate core node is multiplied by the corresponding local density to obtain the comprehensive influence index. The candidate core nodes are sorted in descending order according to the comprehensive influence index, and the top M nodes are selected to form the core node set, where M is the preset number of core nodes and M≤N.

4. The adaptive multimodal entity disambiguation method for enterprise big data according to claim 3, characterized in that, The obtained score distribution of each node specifically includes: For nodes in the core node set, the node influence score is used as the initial propagation value, and the initial propagation value of nodes that do not belong to the core node set is set to 0. Perform multiple rounds of iterative propagation on the influence graph, with each iteration executing the following steps: a. Nodes with a current influence score greater than 0 will be designated as the source nodes for propagation in this iteration; b. Traverse each outgoing edge of each propagation source node and obtain the equity relationship strength or guarantee relationship strength of each outgoing edge as the basic edge weight. Choose the corresponding attenuation coefficient based on whether the outgoing edge is an equity relationship or a guarantee relationship. Based on the current iteration round According to the formula Calculate the adjusted edge weight in this iteration. Multiply the current influence score of the source node by the adjusted weight. The increment of the influence score transmitted along the edge to the target node is obtained; c. Sum the incremental influence scores that each node receives from all incoming edges, and use this sum as the new influence score that the node receives in this iteration; d. Calculate the variance of the newly added influence scores for all nodes in this round. If the variance is lower than the first preset threshold, increase the decay coefficient in the next iteration. The upper limit of the value range; if the variance is higher than the second preset threshold, the attenuation coefficient will be reduced in the next iteration. The lower limit of the range of values; e. Add the newly gained influence points obtained in this round to the cumulative influence points of each node, and reset the current influence points of all propagation source nodes to 0; Repeat steps a to e above until the cumulative influence score change of all nodes in two consecutive iterations is lower than the preset convergence threshold. The cumulative influence scores of each node when the convergence condition is met are normalized to generate the score distribution of each node.

5. The adaptive multimodal entity disambiguation method for enterprise big data according to claim 1, characterized in that, The generated fracture path list specifically includes: The historical name change records of the enterprise are extracted from the enterprise change record database of the enterprise multi-source database and sorted according to the effective time of the change to form the historical name change time series. Each record in the historical name change time series includes a historical name and the corresponding start and end time period. Label each relationship edge in the influence graph with the start and end time periods of the relationship it represents, and iterate through each relationship edge in the influence graph to perform a breakage detection: From the historical name change sequence of the nodes at both ends of the relationship edge, find the historical names that are valid within the start and end time period of the relationship edge. Compare the found historical names with the current names of the corresponding nodes in the influence graph. If the comparison results are inconsistent, or if it is determined from the enterprise unified identification code that the two nodes do not belong to the same entity after the name change, mark the relationship edge as a broken edge. When there are multiple name records with overlapping time periods in the historical name change sequence of the same node, all relation edges related to that node within the overlapping time period are marked as broken edges. In the influence map, starting from the marked fracture edge, the search proceeds along the direction of the edge, and the nodes connected through the fracture edge are connected in sequence to form a fracture path. The priority score of each fracture path is calculated based on the cumulative influence score of each node in the fracture path, the number of fracture edges in the path, and the relationship type. The fracture paths are sorted according to the priority scores to generate the fracture path list.

6. The adaptive multimodal entity disambiguation method for enterprise big data according to claim 1, characterized in that, The repaired influence map includes: Perform cross-modal consistency verification on the feature vectors corresponding to each preset feature dimension of the nodes on both sides of the fracture edge. When the cross-modal consistency meets the preset consistency condition, continue to perform similarity calculation. When multiple preset feature dimensions exist, the similarity of the feature vectors corresponding to each preset feature dimension is calculated, and the time decay function is applied to the similarity of each preset feature dimension based on the time difference of historical name change to generate time decay similarity. The relationship structure consistency factor is calculated based on the overlap of the adjacent node sets of the nodes on both sides of the fracture edge in the influence map. The time decay similarity and the relationship structure consistency factor are weighted and fused according to the preset fusion weight to generate the final comprehensive similarity. When the final comprehensive similarity meets the set threshold, an edge connecting the nodes on both sides of the broken edge is established in the influence graph, and the original relationship strength corresponding to the broken edge is used as the edge weight of the connecting edge to update the connection structure of the influence graph to generate the repaired influence graph.

7. The adaptive multimodal entity disambiguation method for enterprise big data according to claim 1, characterized in that, The generation of the conflict intensity distribution characterizing the evolution of enterprise relationship conflicts includes: The influence score propagation process is re-executed on the repaired influence graph to obtain the repaired node score distribution. The node score distribution before and after the repair is compared at the node level to obtain the difference in the updated score of each node. Based on the comparison of the set of adjacent nodes of each node in the influence map before and after the repair, the structural perturbation factor characterizing the degree of change in the node's adjacency structure is calculated. According to the preset weighting rules, the updated score difference of the nodes and the structural disturbance factor are weighted and combined to obtain the conflict intensity index of each node. Based on the core node set, the nodes are divided into core nodes and edge nodes, and the conflict intensity index distribution of core nodes and edge nodes are statistically analyzed respectively. A conflict intensity distribution is generated based on the difference in conflict intensity index distribution between core nodes and edge nodes.

8. The adaptive multimodal entity disambiguation method for enterprise big data according to claim 1, characterized in that, The generation of the optimized conflict intensity distribution includes: A time series diagram of conflict intensity is constructed based on the conflict intensity distribution at different times, and the change in conflict intensity between adjacent time windows is calculated. Extract business transformation event sequences from the enterprise's multi-source database and align the occurrence times of the business transformation events with the conflict intensity time sequence diagram; Based on the type of business transformation event, determine the corresponding lag window for the event, track the change in conflict intensity within the lag window, and calculate the impact coefficient of the business transformation event on the score decay rate. When multiple business transformation events occur within adjacent time windows, an event superposition factor is constructed based on the superposition relationship between the events, and the event superposition factor is combined with the impact coefficient of each business transformation event to form a comprehensive impact coefficient. The decay parameters of the exponential decay function are adaptively adjusted based on the comprehensive influence coefficient, including adjusting the edge-level decay parameters separately for edges whose conflict intensity exceeds a preset threshold. The conflict intensity distribution is regenerated based on the adjusted attenuation parameters, and the above process is repeated if the conflict intensity distribution does not reach the preset convergence condition in order to obtain an optimized conflict intensity distribution.

9. The adaptive multimodal entity disambiguation method for enterprise big data according to claim 1, characterized in that, The identification of target enterprise entities based on the optimized conflict intensity distribution includes: Construct a temporal distribution vector of the conflict intensity of the target enterprise entity across multiple time windows based on the optimized conflict intensity distribution; The similarity of the temporal distribution vectors of conflict intensity of different target nodes that are suspected to belong to the same enterprise entity is calculated, and the consistency index of the direction of change of conflict intensity between the two is calculated. Peak time identification is performed on the temporal distribution vector of conflict intensity of different target nodes, and the overlap of peak times is compared to obtain peak alignment. The structural perturbation factors of different target nodes are compared based on the optimized influence map to obtain the differences in structural perturbation. Based on the temporal similarity of conflict intensity, the consistency of the direction of conflict intensity change, the peak alignment, the difference of structural disturbance, and the similarity on the preset feature dimensions, a weighted fusion is performed to generate the identity determination score and the corresponding identity confidence of the target enterprise entity. Based on the comparison result between the identity determination score or the identity confidence level and the preset threshold, it is determined whether the target enterprise entity belongs to the same entity, and the identity determination result is used as the final output of enterprise entity disambiguation.

10. An adaptive multimodal entity disambiguation system for enterprise big data, based on the adaptive multimodal entity disambiguation method for enterprise big data as described in any one of claims 1-9, characterized in that, The system includes: The data extraction and node filtering module is used to extract equity relationship data and guarantee relationship data. It constructs an influence graph with enterprise entities as nodes and relationship strength as edges, and uses the PageRank algorithm to calculate node centrality to filter the core node set. The influence propagation module is used to propagate influence scores along the edges of the influence graph, starting from the core node set. During the propagation process, the edge weights are adaptively adjusted by calling an exponential decay function based on the relationship strength to obtain the score distribution of each node. The name change processing module is used to obtain the historical name change time series of enterprises and overlay it into the influence map, identify and mark the relationship edge break caused by historical name changes, and generate a list of break paths. The graph repair module is used to calculate the similarity of nodes on both sides of the fracture edge in at least one preset feature dimension for the fracture path list. When the similarity exceeds a set threshold, the corresponding nodes are reconnected in the influence graph to obtain the repaired influence graph. The conflict intensity generation module is used to re-execute the influence score propagation process based on the repaired influence map, and generate a conflict intensity distribution that characterizes the conflict in the evolution of enterprise relationships by comparing the updated scores of core nodes and edge nodes. The adaptive adjustment module is used to analyze the changing trend of the conflict intensity distribution on the time axis using a time series graph algorithm, and combined with the enterprise's business transformation event sequence, to track the impact of business transformation events on the score decay rate, adaptively adjust the parameters of the exponential decay function, and iteratively generate an optimized conflict intensity distribution based on the adjusted parameters. The entity identity determination module is used to determine the identity of the target enterprise entity based on the optimized conflict intensity distribution, and the identity determination result is used as the final output of enterprise entity disambiguation.

Images