A method for tracing enterprise tax risks based on multimodal temporal knowledge graphs

Abstract

This invention provides a method for tracing enterprise tax risk paths based on a multimodal temporal knowledge graph, comprising the following steps: Multimodal data acquisition step: collecting multimodal data from distributed data sources, including structured tax data, semi-structured text data, and unstructured audio / video data; Temporal graph construction step: constructing a multimodal temporal knowledge graph based on the multimodal data, wherein the entity relationship edges in the temporal knowledge graph carry timestamp attributes to characterize the dynamic characteristics of inter-enterprise relationships over time; Risk event identification step: analyzing real-time data using a multimodal fusion model to identify potential tax risk events and generate risk feature tensors. This invention can effectively track the dynamic transmission path of tax risks among related enterprises, improving the accuracy and efficiency of risk identification.

Description

Technical Field

[0001] This invention relates to tax risk analysis technology, and more specifically, to a method for tracing enterprise tax risk paths based on a multimodal temporal knowledge graph. Background Technology

[0002] In the contemporary business ecosystem, enterprises form a highly complex interactive system through supply chain networks, equity linkages, and lending channels, resulting in significant cross-enterprise transmission characteristics of tax risks. Risk events such as tight cash flow, abnormal invoice cancellations, or business closures often do not occur in isolation but spread gradually along related paths, forming multi-level chain reactions. Traditional tax audit methods mainly rely on static financial indicator threshold comparison mechanisms or isolated analysis of single enterprises, which can only capture instantaneous risk points and cannot effectively monitor the temporal evolution of risks in dynamic interconnected networks. Although existing technologies have attempted to introduce enterprise relationship graphs for risk modeling, such as building knowledge graphs based on multi-source financial and tax data to identify abnormal transaction nodes, or locating risk links through consistency comparison of the four flows of goods, contracts, funds, and invoices, these solutions still have systemic shortcomings. First, the knowledge graphs used are mostly static snapshots, lacking the ability to model over time. This makes it impossible to depict the dynamic transmission trajectory of key risk factors such as broken capital chains and invoice reversals between upstream and downstream enterprises over time. For example, it is impossible to distinguish the chronological order of the risk source enterprise and subsequent transmission nodes. Second, data processing is limited to structured tabular data, failing to integrate unstructured multimodal information sources such as enterprise announcements, public opinion monitoring data, production and operation video streams, and meeting recordings, resulting in incomplete coverage of risk signals. For example, key clues such as a sudden drop in vehicle entry and exit frequency reflected in video surveillance or negative emotions detected in audio analysis are ignored, leading to a significant lag in risk identification. Finally, at the path mining level, when facing large-scale enterprise networks, existing graph algorithms, such as distributed processing models based on the MapReduce framework, are computationally inefficient when dealing with long-chain transmission paths or circular structures such as fund repatriation. They struggle to automatically extract complete transmission links with temporal logic, and the generated path results lack interpretability, failing to provide auditors with intuitive risk tracing evidence. Summary of the Invention

[0003] To overcome the above problems, the purpose of this invention is to provide a method for tracking enterprise tax risk paths based on multimodal temporal knowledge graphs, which can effectively track the dynamic transmission path of tax risks among related enterprises and improve the accuracy and efficiency of risk identification.

[0004] This invention employs the following approach: a method for tracing enterprise tax risks based on multimodal temporal knowledge graphs, comprising the following steps:

[0005] Multimodal data acquisition steps: Collect multimodal data from enterprises from distributed data sources. The multimodal data includes structured tax data, semi-structured text data, and unstructured audio and video data. Temporal graph construction steps: Construct a multimodal temporal knowledge graph based on the multimodal data. The entity relationship edges in the temporal knowledge graph carry timestamp attributes to characterize the dynamic characteristics of inter-enterprise relationships over time. Risk event identification steps: Analyze real-time data using a multimodal fusion model to identify potential tax risk events and generate risk feature tensors. Transmission path tracing steps: Starting from the enterprise entity corresponding to the risk feature tensor, perform a reverse search based on temporal constraints in the multimodal temporal knowledge graph to mine and output a subgraph of the risk transmission path between related enterprises. Visualization steps: Visualize and render the transmission path subgraph and generate an audit report containing risk tracing information.

[0006] Furthermore, the multimodal data acquisition steps are further specified: accessing a distributed data source to collect multimodal heterogeneous data of the target enterprise group; the multimodal heterogeneous data includes structured tax data, semi-structured text data, and unstructured audio and video data; and using a pre-trained feature extraction model to extract temporal feature vectors from the unstructured audio and video data respectively.

[0007] Furthermore, the time-series graph construction steps are further specified: A knowledge graph containing enterprise entities, personnel entities, item entities, and event entities is constructed; transaction relationships in the structured tax data, dynamic events corresponding to the time-series feature vectors, and changes in entity attributes are stored as time-series edges in the graph database, with each time-series edge containing at least a valid time interval and a relationship type label, forming a multimodal time-series knowledge graph supporting time-slice queries; the multimodal time-series knowledge graph: Entity set Includes enterprise entity e, personnel entity p, item entity g, and event entity ev; time-series edge. Each edge is defined as a quadruple: edgev = (h, r, time, attrs), where Subject and object For relational types, The valid time interval is defined by attrs, which is a dynamic attribute.

[0008] Furthermore, the risk event identification step is further specified: based on real-time collected incremental data, risk signals are identified through a multimodal fusion model, and risk features of different modalities are aligned and fused to generate a standardized risk event feature tensor. This tensor includes the risk subject identifier, risk type, confidence level, and occurrence timestamp; the multimodal fusion model includes extracting multimodal features X= By performing feature alignment and fusion through a cross-attention layer, a risk feature tensor R is generated: R = MultiHeadAttention(X,X,X). Risk event features are generated by mapping the data to a risk category space using a multilayer perceptron (MLP): risk_event= =softmax(MLP(R)).

[0009] Furthermore, the transmission path tracing step is further specified as follows: taking the enterprise entity pointed to by the risk event feature tensor as the starting point, setting a time backtracking window and a transmission depth threshold, and executing a reverse time-series walk algorithm in the multimodal time-series knowledge graph; when traversing the path, the reverse time-series walk algorithm requires that the timestamps of all time-series edges on the path satisfy monotonically decreasing and fall within the time backtracking window, thereby mining the complete transmission link from the risk source to the current risk outbreak point, and outputting a risk transmission subgraph containing intermediate transmission nodes; the reverse time-series walk algorithm: taking the risk subject Starting from the time rewind window Inside, perform a reverse walk, with the walk path P= (in It must satisfy a strict time monotonically decreasing constraint: , and all timestamps .

[0010] Furthermore, the visualization step is further specified as follows: the risk transmission subgraph is visualized and rendered, and a structured audit task instruction set containing evidence collection guidance is automatically generated based on the path characteristics in the risk transmission subgraph; in the risk transmission subgraph, nodes and edges are differentiated according to the risk attenuation coefficient, and the location and time of key abnormal documents are marked; the structured audit task instruction set includes: an on-site verification list for enterprises marked as "risk sources", instructions for bank statement penetration query for "intermediate bridge" enterprises, and a cooperation notification template for "risk receiving end" enterprises.

[0011] Furthermore, the multimodal heterogeneous data acquisition and feature extraction steps specifically include: using a federated learning framework to deploy feature extraction models locally on each data holder's site; only exchanging encrypted model gradients or feature vectors to avoid the original tax data going out of domain, thereby achieving multi-source data feature alignment under privacy protection.

[0012] Furthermore, the multimodal temporal knowledge graph construction step further includes: for the unstructured audio and video data, the extracted temporal feature vector includes: the vehicle entry and exit frequency change curve and the start-up rate / shutdown event identifier of the enterprise's production and operation site obtained based on video stream analysis; the meeting minutes keywords and sentiment score obtained based on audio analysis; and attaching the vehicle entry and exit frequency change curve, shutdown event identifier, and sentiment score as dynamic attributes to the corresponding enterprise entity or event entity in the form of temporal edges.

[0013] Furthermore, the multimodal fusion model in the risk event identification step is a Transformer architecture based on an attention mechanism; this model performs weighted fusion of statistical outliers in structured data, negative sentiment scores in text data, and abnormal fluctuation features in audio and video data to generate a unified risk event representation vector, which is then mapped to a preset risk category label space.

[0014] Furthermore, in the time-constrained transmission path tracing step, the reverse time-series walk algorithm specifically includes: using a path scoring function based on a time-series graph neural network to perform probability evaluation on candidate transmission paths; when it is detected that the flow of funds, invoices, and goods form a closed-loop structure in the time series, the closed-loop path is marked as a "circular transmission" risk pattern and included in the output risk transmission subgraph; the reverse time-series walk algorithm introduces a Monte Carlo tree search mechanism during the walk process to balance the breadth and depth of path exploration, prioritizing the discovery of multi-level bridging transmission paths with a hidden depth exceeding a preset threshold; the Monte Carlo tree search mechanism, in order to solve the problem of balancing depth and breadth in graph search, introduces a path scoring function based on UCT (Upper Confidence Bounds for Trees) to guide the reverse walk, for the current node s, its candidate predecessor nodes The UCT value is calculated as follows: ;in: It is a path value function, representing the path from the current node to the destination. The cumulative confidence level traced back to the source of risk (which can be represented by the weighted sum of the risks at each node along the path); It is the total number of times node s was visited. It is the edge The number of times it is selected, C is the exploration constant, which controls the balance between exploration and utilization.

[0015] The beneficial effects of this invention are as follows: By constructing a multimodal temporal knowledge graph and performing a reverse search with temporal constraints, this invention can uncover risk transmission paths, effectively track the dynamic transmission paths of tax risks among related enterprises, and improve the accuracy and efficiency of risk identification; it enhances the accuracy of risk transmission tracing: by introducing a time dimension, this invention can clearly outline the spatiotemporal diffusion trajectory of risks, distinguish between "risk sources" and "risk bearers," and avoid misjudgments; it enhances the comprehensiveness and timeliness of risk identification: by integrating multimodal data such as audio, video, and text, risk signals that were originally hidden in unstructured information (such as videos of enterprise shutdowns and negative news about actual controllers) can be included in the analysis in a timely manner, achieving early warning; and it improves the interpretability of the algorithm: the generated transmission path is not just a risk score, but a visualized relationship link, providing tax inspectors with clear case-handling clues and evidence chains, meeting the intelligent needs of "data-driven tax administration." Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Traditional tax auditing methods primarily rely on static financial indicator comparisons or isolated analyses of single enterprises, making it difficult to capture the dynamic transmission process of tax risks within complex interconnected networks. Existing enterprise relationship graphs are mostly static snapshots, lacking a temporal dimension and failing to depict the transmission sequence of risk factors between upstream and downstream enterprises over time. Furthermore, most methods are based solely on structured tabular data, failing to effectively integrate unstructured multimodal information such as public opinion texts and operational video surveillance, resulting in delayed and incomplete risk detection. In addition, when dealing with large-scale enterprise networks, existing graph mining algorithms are inefficient at handling non-preset structures and struggle to automatically generate interpretable risk transmission paths.

[0019] Please see Figure 1 As shown, in response to this, the present invention proposes a method for tracing enterprise tax risk paths based on a multimodal temporal knowledge graph, comprising the following steps:

[0020] Multimodal data acquisition steps: Collect multimodal data from distributed data sources, including structured tax data, semi-structured text data, and unstructured audio and video data;

[0021] Steps for constructing a time-series knowledge graph: Based on the multimodal data, a multimodal time-series knowledge graph is constructed. The entity relationship edges in the time-series knowledge graph carry timestamp attributes, which are used to characterize the dynamic characteristics of the relationship between enterprises over time.

[0022] Risk event identification steps: Analyze real-time data using a multimodal fusion model to identify potential tax risk events and generate risk feature tensors;

[0023] Transmission path tracing steps: Starting from the enterprise entity corresponding to the risk feature tensor, perform a reverse search based on temporal constraints in the multimodal temporal knowledge graph to mine and output the transmission path subgraph of risk among related enterprises;

[0024] Visualization steps: Visualize and render the transmission path sub-graph, and generate an audit report containing risk tracing information.

[0025] For ease of understanding, the following explains some key terms in this embodiment:

[0026] Multimodal data refers to a collection of data obtained from different data sources or in different formats, encompassing various information expression methods. In this embodiment, the multimodal data specifically includes structured tax data, semi-structured text data, and unstructured audio and video data. These data are used together to comprehensively characterize the company's business activities and external environment. A mathematical expression for multimodal feature extraction is defined, assuming the set of companies is E. For any company... Its multimodal data source can be formalized as:

[0027] Structured data: , representing the financial indicators at time t and its numerical value ;

[0028] Semi-structured text: Such as financial reports and news;

[0029] Unstructured audio and video: Such as surveillance video and meeting recordings.

[0030] Temporal feature vectors are extracted using a pre-trained model Φ. For video frame sequences... Extract dynamic features using 3D-CNN or video Transformer: For audio streams, after extracting the Mel spectrogram, emotion and keyword features are obtained through an audio model: .

[0031] A multimodal temporal knowledge graph is a knowledge representation model that can integrate data from multiple modalities and capture the dynamic changes in relationships between entities over time. In this graph, the edges between entities are assigned timestamp attributes, which record the time information of the relationship's occurrence, duration, or termination, thus characterizing the evolution of relationships between enterprises. The multimodal temporal knowledge graph is described below. Entity set Includes enterprise entity e, personnel entity p, item entity g, and event entity ev; time-series edge. Each edge is defined as a quadruple: edgev = (h, r, time, attrs), where Subject and object For relational types, For the valid time interval, attrs are dynamic attributes (such as transaction amount and confidence level). Especially for dynamic events extracted from audio and video (such as "work stoppage events"), they are modeled as event entities ev, and connected via temporal edges. Associated with enterprises.

[0032] Using a property graph model, the following node and edge types are defined:

[0033]

[0034] A risk feature tensor is a multidimensional array used to quantify and represent the various attributes of identified tax risk events. This tensor can contain information such as the risk subject identifier, risk type, confidence level, and occurrence timestamp, providing structured input for subsequent risk transmission path tracing.

[0035] A transmission path subgraph refers to a set of paths in a multimodal temporal knowledge graph that start from a risk event's originating enterprise entity, pass through a series of related enterprise entities and temporal relationship edges, and ultimately reach the risk outbreak point or related enterprises. This subgraph clearly illustrates the propagation chain and temporal characteristics of risk in the enterprise network.

[0036] This embodiment provides a method for tracing enterprise tax risk paths based on a multimodal temporal knowledge graph. Specifically, the method includes the following operations:

[0037] In the multimodal data acquisition step, multimodal data from the enterprise is collected from distributed data sources. This multimodal data includes structured tax data, semi-structured text data, and unstructured audio and video data. Data acquisition can be performed through API interfaces, file imports, or database connections. For example, structured tax data can be directly exported from the tax bureau system or the enterprise's ERP system; semi-structured text data can be crawled from the enterprise's internal document management system or public websites; and unstructured audio and video data can be uploaded manually or obtained through simple file transfer protocols. After collection, this data can be stored in a central data warehouse for further processing.

[0038] In the time-series graph construction step, a multimodal time-series knowledge graph is built based on this multimodal data. The entity relationship edges in this time-series graph carry timestamp attributes, used to characterize the dynamic features of inter-enterprise relationships changing over time. After acquiring the multimodal data, the data can first be preprocessed, such as performing word segmentation and entity recognition on text data, and preliminary content recognition on audio and video data. Subsequently, through manual annotation or rule-based methods, enterprise entities, personnel entities, event entities, etc., are identified from this data, and relationships between them are established. These relationships can be assigned a simple occurrence time point as a timestamp, such as the date of a transaction or the signing date of a contract. These entities and timestamped relationships can be stored in a relational database or a simple key-value store, forming a basic graph structure with time attributes.

[0039] In the risk event identification step, a multimodal fusion model is used to analyze real-time data, identify potential tax risk events, and generate a risk feature tensor. Risk event identification can employ threshold-based or rule-based methods. For example, for structured tax data, abnormal thresholds for certain financial indicators can be set; exceeding these thresholds marks a risk signal. For text data, negative information can be identified based on keyword matching. For audio and video data, manual viewing or listening can be used to determine if any anomalies exist. These identified risk signals can be easily summarized, and a list containing the risk type, involved companies, and occurrence time can be generated manually or through a preset template, serving as the preliminary form of the risk feature tensor. The multimodal fusion model includes extracting multimodal features X= By performing feature alignment and fusion through a cross-attention layer, a risk feature tensor R is generated: R = MultiHeadAttention(X,X,X). Risk event features are generated by mapping the data to a risk category space using a multilayer perceptron (MLP): risk_event= =softmax(MLP(R)).

[0040] An encoder structure based on Transformer is used. This is for statistical outliers in structured data. Text sentiment score Audio and video fluctuation characteristics A unified representation is generated through a weighted fusion layer: Attention weights It is calculated from the query vector q and the modal features: The final risk category probability is: .

[0041] In the transmission path tracing step, starting with the enterprise entity corresponding to the risk feature tensor, a reverse search is performed in the multimodal temporal knowledge graph based on temporal constraints to mine and output a subgraph of the risk transmission path among related enterprises. Once the risk event and its starting enterprise are identified, a reverse traversal can be performed manually or using a simple breadth-first search (BFS) or depth-first search (DFS) algorithm in the temporal knowledge graph. During the traversal, a basic time-reverse order principle can be set, i.e., searching only along the relationship edges with decreasing timestamps to ensure the temporal rationality of the path. For example, starting from the risk outbreak point, find enterprises that are temporally related to it, and then continue tracing backward from these enterprises until a preset search depth is reached or no more related enterprises that meet the time-reverse order principle can be found. Finally, all paths that meet the conditions are combined into a transmission path subgraph.

[0042] In the visualization step, the transmission path subgraph is rendered and a tax audit report containing risk tracing information is generated. The obtained transmission path subgraph can be rendered using a general graph visualization tool. Enterprise entities in the graph can be represented as nodes, and relationships as edges. Different enterprise types or relationship types can be distinguished using simple color coding or labels. Simultaneously, based on the visualization results, a basic tax audit report can be manually written or generated using a preset template. This report includes a description of the risk event, a list of involved enterprises, and a preliminary transmission path diagram, providing tax auditors with preliminary risk tracing information.

[0043] This embodiment integrates multimodal data and constructs a temporal knowledge graph to dynamically capture the evolution of relationships between enterprises, overcoming the limitations of traditional static analysis. By identifying risk events through a multimodal fusion model and tracing the transmission path in reverse based on temporal constraints, it achieves a comprehensive and timely disclosure of the propagation chain of tax risks in complex enterprise networks, providing interpretable risk tracing information and decision support for tax audits.

[0044] To address this issue, this invention proposes a method for tracing corporate tax risks based on a multimodal temporal knowledge graph. This method achieves comprehensive monitoring and in-depth analysis of corporate tax risks through steps such as multimodal data collection, temporal graph construction, risk event identification, transmission path tracing, and visualization. However, in practical applications, the basic multimodal data collection steps may face technical challenges such as scattered data sources, complex and diverse data types, and effectively extracting key information from unstructured data. This may lead to incomplete data coverage or missing key dynamic features, thereby affecting the accuracy of subsequent risk identification and tracking.

[0045] In response, this invention further proposes a specific scheme for the above-mentioned multimodal data acquisition steps, including accessing a distributed data source to collect multimodal heterogeneous data of the target enterprise group; the multimodal heterogeneous data includes structured tax data, semi-structured text data, and unstructured audio and video data; and using a pre-trained feature extraction model to extract temporal feature vectors from the unstructured audio and video data respectively.

[0046] Specifically, accessing distributed data sources aims to acquire data from multiple independent and potentially geographically dispersed data storage or processing systems. These data sources can include internal corporate systems such as financial systems, ERP systems, CRM systems, and supply chain management systems, as well as external platforms such as business registration information platforms, tax filing systems, banking transaction systems, social media platforms, and news websites. Accessing distributed data sources ensures the breadth and depth of data collection, covering various information from different business processes and external environments for the target enterprise group. Collecting multimodal heterogeneous data from the target enterprise group means focusing not only on individual enterprises but also on a group of enterprises with related relationships (such as equity, business dealings, and employee part-time positions) as a whole for data collection, in order to capture complex interactions and potential risk transmission between enterprises. Heterogeneous data emphasizes the diversity of data in terms of format, structure, and semantics.

[0047] The multimodal heterogeneous data encompasses various data types to ensure a comprehensive understanding of a company's operational status. Structured tax data typically refers to data with a fixed format and defined fields, such as a company's tax returns, invoice details, bank statements, and financial statements. This data can be directly used for quantitative analysis and rule matching. Semi-structured text data includes contracts, agreements, meeting minutes, email correspondence, announcements, and news reports. While this data lacks a strict fixed structure, it usually contains tags or markers, allowing for parsing and information extraction using natural language processing techniques. Unstructured audio and video data, such as surveillance videos from a company's production and operation sites, telephone recordings, and meeting recordings, lacks a predefined structure but contains rich behavioral patterns and event information.

[0048] Extracting temporal feature vectors from unstructured audio and video data using pre-trained feature extraction models is a crucial step in processing unstructured data. Pre-trained models are those trained on large-scale datasets and possess general feature extraction capabilities, such as Convolutional Neural Networks (CNNs) or Transformers for video analysis, and Recurrent Neural Networks (RNNs) or Transformers for audio analysis. These models can automatically learn and recognize patterns in audio and video data, such as object movement, personnel activity, vehicle entry and exit, and equipment operation status changes in videos, as well as speech content, emotional tone, and keyword frequency in audio. The extracted temporal feature vectors are sequential representations of these features over time, preserving the temporal order and dynamic evolution of events, rather than merely static snapshot information. For example, a curve showing the frequency of goods entering and leaving a company's warehouse can be extracted from a video, and the frequency of specific keywords and emotional fluctuation trends can be extracted from meeting recordings; all of these are presented in the form of temporal feature vectors.

[0049] Through the above technical solution, this invention effectively addresses the challenges of acquiring multi-source heterogeneous data. By accessing distributed data sources and focusing on target enterprise groups, it ensures comprehensive data coverage and relevance, laying the foundation for building a complete enterprise relationship network. Simultaneously, by clearly distinguishing between structured, semi-structured, and unstructured data types, and specifically utilizing pre-trained feature extraction models to extract temporal feature vectors from unstructured audio and video data, it effectively captures previously difficult-to-use dynamic and implicit information. These temporal feature vectors accurately reflect the real-time dynamics and behavioral patterns of enterprise operations, greatly enriching the foundation for constructing multimodal temporal knowledge graphs. This provides more refined and timely evidence for subsequent risk event identification, thereby significantly improving the accuracy of tax risk identification and the depth of transmission path tracing.

[0050] In some embodiments of the present invention, a multimodal temporal knowledge graph based on multimodal data is proposed, and the timestamp attribute of entity relationship edges is used to characterize the dynamic features of relationships between enterprises. However, in its implementation, if the entity type definition of the knowledge graph is not comprehensive enough, or if dynamic events and changes in entity attributes are not effectively stored in a unified and queryable temporal edge format, the expressive power of the graph may be limited, making it difficult to accurately capture and trace complex tax risk transmission paths.

[0051] In response, this invention further specifies the steps for constructing the time-series graph as follows: constructing a knowledge graph containing enterprise entities, personnel entities, item entities, and event entities; storing the transaction relationships in the structured tax data, the dynamic events corresponding to the time-series feature vectors, and changes in entity attributes as time-series edges in the graph database, with each time-series edge containing at least a valid time interval and a relationship type label, forming a multimodal time-series knowledge graph that supports time-slice queries.

[0052] Specifically, when constructing the knowledge graph, four core entity types are clearly defined: enterprise entities, personnel entities, item entities, and event entities. Enterprise entities are the core objects of tax risk analysis, representing tax subjects such as various companies and individual businesses. Their attributes can cover business registration information, equity structure, legal representative, business scope, industry classification, etc., serving as the foundational nodes for risk assessment and transmission path tracing. Personnel entities refer to natural persons closely related to the enterprise's operations, including legal representatives, shareholders, senior executives, actual controllers, and financial officers. Through personnel entities, implicit connections between enterprises can be revealed; for example, identifying related-party transactions or shell companies by having the same person work for different companies. Item entities represent specific objects involved in the enterprise's operations, such as specific types of raw materials, finished products, services provided, and fixed assets. Through item entities, the authenticity and rationality of cash flow and goods flow can be tracked, and fraudulent transactions or price anomalies can be identified. Event entities refer to behaviors or state changes that occur at a specific point in time during a company's business operations, such as transactions, contract signing, equity changes, administrative penalties, invoice issuance / cancellation, and asset mortgages. Event entities are key to capturing dynamic changes and can transform static entity relationships into dynamic processes over time.

[0053] Based on this, dynamic information from multiple sources is uniformly modeled as temporal edges and stored in a graph database. Temporal edges are special relationships in a knowledge graph that connect entities and carry temporal attributes. They not only represent the association between entities but also record the time information of the occurrence, duration, or termination of this association, thus enabling the graph to have dynamic evolution capabilities in the time dimension. Specifically, transaction relationships in the structured tax data, such as purchase and sale invoice relationships between enterprises, fund transfer records, and contract signing information, inherently carry clear transaction times or effective dates. These data can be directly converted into temporal edges connecting related enterprise entities, with attributes such as transaction amount and product category. The dynamic events corresponding to the temporal feature vectors typically originate from the analysis of unstructured data. For example, by processing audio and video data, changes in vehicle entry and exit frequency at enterprise production and operation sites, start / stop events, meeting minutes keywords, and sentiment tendencies can be extracted. These dynamic events can be abstracted as event entities and associated with corresponding enterprise entities through temporal edges, with the edges recording the occurrence time or duration of the event. Changes in entity attributes, such as changes in a company's legal representative, increases or decreases in registered capital, adjustments to business scope, and personnel changes, are inherently time-sensitive events. These changes can be linked to the corresponding entities as time-series edges, recording the effective or change time of the attribute change. Graph databases are database systems specifically designed for storing and querying graph-structured data. Their core advantage lies in their ability to efficiently handle complex and diverse relationships between entities. Through graph databases, operations such as path finding, community discovery, and centrality analysis can be easily performed, providing an efficient data storage and query foundation for tracing the transmission paths of tax risks.

[0054] To ensure the validity and queryability of time-series edges, each edge contains at least a valid time interval and a relation type label. The valid time interval refers to the start and end times of the relationship or event represented by the edge. This can be a precise point in time (e.g., the time a transaction occurs) or a time period (e.g., the period of tenure). This attribute is core to time-slice queries, allowing the system to precisely filter and query graph states within a specific time window, ensuring the timeliness and accuracy of risk analysis. The relation type label clarifies the specific semantic relationship represented by the edge, such as "issues an invoice to," "is the legal representative," "occurs a change in shareholding," or "participates in a meeting." Relationship type labels allow for the classification and semantic understanding of relationships in the graph, facilitating path search and pattern matching based on specific relationship types, thereby identifying different types of risk transmission mechanisms. Ultimately, this forms a multimodal time-series knowledge graph that supports time-slice queries. Time-slice queries refer to the ability to query the relationship status between entities in a knowledge graph based on a specified point in time or time period. For example, querying which companies A has equity relationships with on a specific date, or which companies B had large transactions with during a certain quarter. This query capability is crucial for tax risk tracking because it allows analysts to trace back to a specific point in time before a risk event occurs, reconstruct the relationship network at that time, and thus accurately identify the source and transmission path of the risk.

[0055] Through the aforementioned technical solution, this invention significantly enriches the semantic expression capabilities of knowledge graphs by clearly defining core entity types such as enterprises, personnel, goods, and events, enabling them to more comprehensively depict the enterprise tax ecosystem. Simultaneously, it unifies the modeling of transaction relationships in structured tax data, dynamic events corresponding to time-series feature vectors extracted from unstructured data, and entity attribute changes into time-series edges with valid time intervals and relationship type labels, storing them in a graph database. This allows the knowledge graph to efficiently store and manage time-sensitive dynamic relationships. This design not only enhances the graph's ability to capture changes in the time dimension, but more importantly, it forms a multimodal time-series knowledge graph that supports time-slice queries. This allows for precise querying and analysis of dynamic relationships between entities within specific time windows during subsequent risk transmission path tracing, effectively capturing and tracing back complex tax risk transmission paths that depend on time evolution, significantly improving the accuracy and depth of risk tracking.

[0056] In some of the embodiments of the present invention, although it is proposed to use a multimodal fusion model to analyze real-time data to identify potential tax risk events and generate risk feature tensors, in practical applications, if the real-time data is not refined and the risk features of different modalities are not effectively aligned and fused, the accuracy and standardization of risk identification may be insufficient, affecting the efficiency and reliability of subsequent risk transmission path tracking.

[0057] In response, this invention further proposes a risk event identification step that specifically involves: identifying risk signals based on real-time collected incremental data through a multimodal fusion model, aligning and fusing risk features from different modalities to generate a standardized risk event feature tensor, which includes the risk subject identifier, risk type, confidence level, and occurrence timestamp.

[0058] Specifically, incremental data based on real-time acquisition refers to the system's ability to continuously acquire the latest data from distributed data sources, ensuring the timeliness of risk identification. Incremental data specifically refers to data that has been added or changed since the previous round of data processing, rather than the full dataset. By processing only incremental data, computational resource consumption can be significantly reduced, and the response speed of risk identification can be improved, enabling the system to respond promptly to sudden or rapidly evolving tax risks. For example, when a company submits new invoice data, bank statements change, or related negative reports appear in the public media, this new data will be immediately captured and sent to the risk identification process.

[0059] The role of the multimodal fusion model here is to comprehensively analyze information from multiple modalities, including structured tax data, semi-structured text data, and unstructured audio and video data, to identify abnormal patterns or potential risk indicators, i.e., "risk signals." This model can learn the correlations and complementarities between different modalities of data. For example, abnormal transaction amounts in structured data may corroborate negative public opinion in text data or abnormal business activities in audio and video data. Identifying risk signals is the foundation for generating risk event feature tensors, which transform raw, heterogeneous data into a unified intermediate representation with risk indication significance.

[0060] Because data from different modalities differ in format, semantics, and temporal granularity, direct fusion may lead to information distortion or conflicts. Therefore, these risk features need to be aligned and fused before fusion. Alignment typically includes temporal alignment (mapping data from different modalities onto a unified timeline, e.g., associating text events with related transaction records over time), semantic alignment (unifying the representation of features describing the same concept in different modalities), and spatial alignment (if geographic location information is available). Fusion refers to integrating the aligned multimodal features into a unified, more discriminative feature vector or tensor using specific algorithms (such as feature-level fusion, decision-level fusion, or hybrid fusion) to comprehensively capture risk information.

[0061] A standardized risk event feature tensor is a unified data structure that encapsulates identified risk information in a predefined, standardized format. This standardization ensures comparability between different risk events and provides clear and consistent input for subsequent risk transmission path tracing steps. As a multidimensional array, the tensor can effectively represent complex multimodal features and their interrelationships; for example, it can contain multiple dimensions to represent different risk indicators, time series features, etc. This tensor includes a risk subject identifier, risk type, confidence level, and occurrence timestamp. The risk subject identifier uniquely identifies the enterprise or individual entity experiencing the risk, such as a company's unified social credit code. The risk type specifies the particular category of the identified risk, such as the risk of issuing false invoices, tax evasion, or a broken capital chain, which helps in the classification and targeted handling of risks. The confidence level represents the model's certainty about the identified risk event, typically a probability value between 0 and 1; a higher confidence level means a greater likelihood of the risk occurring. The occurrence timestamp records the point in time when the risk event was identified or actually occurred, which is crucial for subsequent time-constrained risk transmission path tracing, ensuring the correctness of the temporal logic.

[0062] Through the above technical solution, this invention effectively addresses the problems of insufficient identification accuracy and standardization caused by diverse data sources, high real-time requirements, and inconsistent feature representations in the risk event identification process. Analysis based on real-time collected incremental data significantly improves the timeliness and response speed of risk identification, enabling the system to quickly capture the latest risk dynamics. Simultaneously, by finely aligning and fusing risk features from different modalities, the information gap caused by heterogeneous data is overcome, allowing the multimodal fusion model to more comprehensively and accurately identify potential risk signals. The resulting standardized risk event feature tensor not only unifies the representation format of risk information, ensuring the integrity and standardization of key information such as risk subject identification, risk type, confidence level, and occurrence timestamp, but also provides high-quality, directly usable input for subsequent risk transmission path tracing, greatly improving the accuracy and efficiency of risk tracking, thereby providing more accurate and timely decision support for tax audits.

[0063] In some embodiments of the present invention, a multimodal fusion model is proposed to analyze real-time data, identify potential tax risk events, and generate standardized risk event feature tensors. Starting from the enterprise entity pointed to by this tensor, transmission path tracing is performed in a multimodal temporal knowledge graph. However, in complex enterprise relationship networks, simply searching for paths starting from risk events may not accurately trace the true source of risk transmission. Especially when facing massive and dynamically changing relational data, how to effectively utilize temporal information to ensure that the traced paths have actual causal relationships and avoid interference from irrelevant paths is a problem that needs to be solved.

[0064] To address this, the present invention further proposes a specific implementation method for the transmission path tracing step, namely, taking the enterprise entity pointed to by the risk event feature tensor as the starting point, setting a time backtracking window and a transmission depth threshold, and executing a reverse time-series walk algorithm in the multimodal time-series knowledge graph; when traversing the path, the reverse time-series walk algorithm requires that the timestamps of all time-series edges on the path satisfy monotonically decreasing and fall within the time backtracking window, thereby mining the complete transmission link from the risk source to the current risk outbreak point, and outputting a risk transmission subgraph containing intermediate transmission nodes.

[0065] Specifically, the risk event feature tensor is a data structure that standardizes the representation of identified tax risk events after processing by a multimodal fusion model. It includes a risk entity identifier, which clearly indicates which corporate entity is the direct trigger point or related party of the current risk. Starting with this corporate entity means that subsequent path tracing will begin from this known risk point and explore its upstream or related parties to find the source of the risk.

[0066] During the tracing process, setting a time backtracking window refers to limiting the search scope within a specific timeframe when tracing risk transmission paths. It defines the maximum time span that can be traced back from the time the risk event occurred. For example, it can be set to "the most recent year," "the most recent three years," etc. The purpose of setting this window is to focus on recent historical activities that are highly correlated with the current risk event, excluding historical data that is too distant and may no longer have a direct causal relationship, thereby improving the efficiency and accuracy of tracing and avoiding waste of computational resources. Simultaneously, the transmission depth threshold refers to the maximum number of node hops or relationship levels allowed to be traversed when searching paths in a multimodal time-series knowledge graph. For example, setting it to "3 hops" or "5 hops" means that a maximum of 3 or 5 related enterprise entities can be traced back to the risk origin. The purpose of setting this threshold is to control the breadth and depth of the search, preventing the path tracing algorithm from getting stuck in an infinite loop or exploring distant entities with extremely low correlation to the risk, thereby ensuring that the output risk transmission subgraph has high relevance and interpretability.

[0067] Building upon this foundation, the reverse temporal walk algorithm is a graph traversal algorithm that explores the paths of relationships between entities in a multimodal temporal knowledge graph in reverse temporal order. Unlike traditional graph traversal algorithms, its core lies in strictly adhering to time constraints; that is, during the backtracking process, the timestamps of all temporal edges on the path must satisfy a specific temporal order. This algorithm aims to start from a known risk outbreak point and trace the order of events and their transmission chains in reverse to identify the initial source and intermediate transmission nodes of the risk. The key constraint of the reverse temporal walk algorithm is that the timestamps of all temporal edges on the path must be monotonically decreasing and fall within the time backtracking window. Monotonically decreasing timestamps mean that when traversing from the risk outbreak point back to the source, the timestamp of each temporal edge traversed (e.g., transaction occurrence, event occurrence) must be strictly earlier than or equal to the timestamp of the previous step, ensuring a causal order in time. Meanwhile, the requirement of "falling within the time backtracking window" further restricts these temporal edges to occur within a preset time backtracking window, ensuring that the traced path is temporally relevant to the current risk event and avoiding tracing back to irrelevant historical events. Through the above algorithm, the complete transmission chain from the risk source to the current risk outbreak point is ultimately uncovered. The complete transmission chain refers to the complete path from the initial source enterprise entity that triggered the tax risk, through a series of intermediate related enterprises and events, to the enterprise entity currently identified as having risk. This chain clearly demonstrates the occurrence, development, and transmission process of risk, providing crucial tracing information for tax audits. Finally, the system outputs a risk transmission subgraph containing intermediate transmission nodes. The risk transmission subgraph is a subset of the multimodal temporal knowledge graph, containing the identified risk source, the current risk outbreak point, and all intermediate transmission nodes between them, as well as the temporal correlation edges between these nodes. This subgraph graphically and intuitively displays the risk transmission path, making complex enterprise relationships and time series information readily apparent, facilitating understanding and analysis by tax personnel. Intermediate transmission nodes refer to corporate entities or event entities that act as bridges or intermediaries in the process of risk transmission from its source to its outbreak point.

[0068] By employing the aforementioned technical solution, starting with the enterprise entity pointed to by the risk event feature tensor, and setting a clear time backtracking window and transmission depth threshold, this invention effectively limits the scope of risk tracking and avoids invalid searches. More importantly, by executing a reverse temporal walk algorithm and strictly requiring that the timestamps of all temporal edges on the path satisfy monotonically decreasing and fall within the time backtracking window, it ensures that the mined transmission path has strict causal logic in time, thereby accurately tracing back from the current risk outbreak point to the true source of the risk. This time-constrained reverse tracking mechanism effectively solves the problem of identifying the true risk transmission link in complex multimodal temporal knowledge graphs, avoids irrelevant path interference that may occur with traditional methods, and finally outputs a risk transmission subgraph containing intermediate transmission nodes, providing clear and credible evidence of risk tracing for tax audits, greatly improving the accuracy and efficiency of risk warning and handling.

[0069] In some embodiments of the present invention described above, a reverse time-series walk algorithm is proposed to mine and output a subgraph of the risk transmission path between related enterprises. However, after obtaining the risk transmission subgraph, how to efficiently transform this complex graph information into action plans with clear guidance that tax auditors can directly implement, and how to intuitively present the severity of the risk and key evidence points, remains a problem that needs to be solved.

[0070] To address this, the present invention further proposes to visualize and render the risk transmission subgraph, and automatically generate a structured audit task instruction set containing evidence collection guidelines based on the path characteristics in the risk transmission subgraph. Specifically, the system transforms the abstract graph structure data into an intuitive graphical interface. For example, it can use graph visualization tools or libraries such as D3.js and ECharts for rendering, selecting different shapes, colors, or line styles according to node types (such as enterprise entities, personnel entities, item entities, and event entities) and edge types (such as transaction relationships and dynamic events) to enhance readability. Furthermore, the system can differentiate the coloring of nodes and edges in the risk transmission subgraph based on the degree of risk attenuation along the transmission path. For example, the risk source enterprise may be marked in dark red, while the color of related nodes and edges gradually lightens as the risk propagates outward, thus intuitively reflecting the severity and spread of the risk. At the same time, the system will also mark the location and time of key abnormal documents. For example, abnormal transaction vouchers such as red-inked invoices and voided invoices will be highlighted in the form of icons or text labels on the corresponding enterprise nodes or transaction edges, and the time of their occurrence will be clearly stated, providing investigators with direct evidence clues.

[0071] Furthermore, the structured audit task instruction set includes: an on-site verification checklist for enterprises marked as "risk sources," bank statement lookup instructions for "bridging" enterprises, and cooperation notification templates for "risk recipient" enterprises. For example, for enterprises identified as risk sources, the system will automatically generate a detailed on-site verification checklist, clearly defining the key content and steps of the verification, such as verifying their accounting documents, business premises, and personnel interviews; for enterprises that act as "bridging" channels for funds or business in risk transmission, the system will generate specific bank statement lookup instructions to guide auditors in investigating fund flows to reveal their specific role in risk transmission; and for the recipient enterprises ultimately affected by the risk, the system will generate standardized cooperation notification templates to quickly initiate cross-departmental or cross-regional collaborative investigations. These instruction sets are structured, meaning they have a unified format and fields, facilitating subsequent system processing or manual execution, greatly improving the standardization and automation level of audit work.

[0072] Through the above technical solution, this invention, based on the identification of risk transmission path sub-graphs, further provides the ability to transform complex graph information into action plans with clear guidance that tax auditors can directly execute. By differentiating the risk transmission sub-graphs and annotating key abnormal documents, auditors can intuitively understand the severity of the risk, the direction of transmission, and key evidence points, thereby quickly identifying the focus of the investigation. Simultaneously, the automatically generated set of structured audit task instructions, such as on-site verification checklists for risk source enterprises, bank statement lookup instructions for intermediate bridging enterprises, and cooperation notification templates for risk receiving enterprises, greatly improves the efficiency and accuracy of tax audit work, reduces the time spent on manual analysis and instruction generation, ensures the timeliness and effectiveness of risk handling, and avoids audit blind spots or delays caused by improper information interpretation or ambiguous instructions.

[0073] In some embodiments of the present invention, a method is proposed to collect multimodal heterogeneous data from enterprises from distributed data sources and extract temporal feature vectors from unstructured audio and video data using a pre-trained feature extraction model. However, in practice, due to the high sensitivity of enterprise tax data, directly aggregating raw multimodal heterogeneous data from various data holders to a central platform for processing can easily lead to data privacy leaks and data security compliance risks, thereby hindering the effective integration and utilization of multi-source data and limiting the comprehensiveness of risk analysis.

[0074] In response, this invention further proposes that the multimodal heterogeneous data acquisition and feature extraction steps specifically include: using a federated learning framework to deploy feature extraction models locally on each data holder's site; only exchanging encrypted model gradients or feature vectors to avoid the original tax data going out of domain, thereby achieving multi-source data feature alignment under privacy protection.

[0075] Federated learning is a distributed machine learning paradigm that allows multiple participants to collaboratively train a shared model without directly sharing the original data. Each data holder trains its model locally using its private data and only sends model updates (e.g., gradients of model parameters) to a central aggregation server. The aggregation server integrates these updates to improve the global model and then distributes the updated global model back to the participants, iterating in this way. This mechanism ensures that the original data always remains local, effectively protecting data privacy. Deploying feature extraction models locally at each data holder means that the initial stage of data processing—extracting valuable features from the raw multimodal data—is completed within the data producer's or owner's own computing environment. For example, a company's tax system, monitoring system, or office system can act as a data holder, with its local servers or computing units running models specifically designed to extract features from its internal data. This deployment ensures that the original sensitive data never leaves its controlled environment, thus guaranteeing data security from the source. In federated learning, only encrypted model gradients or feature vectors are exchanged. After training the model and extracting features locally, the data holder does not directly upload the original data. Instead, they upload the gradient information generated during model training or the feature vectors processed by the feature extraction model to the aggregation server. To further enhance security, these gradients or feature vectors are encrypted before transmission, using techniques such as homomorphic encryption, differential privacy, or secure multi-party computation. This ensures that even if intercepted during transmission, the original data content cannot be deduced, effectively preventing data leakage. Preventing original tax data from leaving the domain refers to data leaving its original storage or processing environment and entering external systems or platforms. Through the aforementioned local deployment and encrypted interaction mechanism, original, unprocessed sensitive tax data (such as specific transaction records, invoice details, financial statements, etc.) remains within the data holder's premises and is not transmitted to external central servers or third-party platforms. This fundamentally eliminates the privacy leakage risks associated with centralized data storage and transmission, complying with stringent data protection regulations. Ultimately, privacy-preserving feature alignment of multi-source data is achieved. Although the original data remains within the domain, the feature vectors or model gradients extracted locally by each data holder can be integrated on the aggregation server through a federated learning mechanism. This integration is not a simple data splicing, but rather an algorithm that ensures that features from different sources and modalities are semantically and dimensionally matched and fused to form a unified, high-dimensional feature representation for subsequent risk analysis. Because the entire process is accompanied by strict privacy protection measures, feature alignment using multi-source heterogeneous data is effectively achieved while ensuring data privacy, laying the foundation for constructing a comprehensive multimodal temporal knowledge graph.

[0076] Through the above technical solution, this invention ensures that the original tax data of each data holder is used for feature extraction and model training locally, thereby avoiding the risk of sensitive data leaving the domain directly. Specifically, only encrypted model gradients or feature vectors are exchanged, rather than the original data, which greatly enhances the privacy, security, and compliance of data transmission and processing. This mechanism effectively solves the data privacy and security challenges faced in the process of collecting multi-source heterogeneous data, significantly increasing the willingness and trust of data holders to participate in data sharing. Ultimately, without sacrificing data privacy, it achieves effective alignment and fusion of data features from different sources and modalities, providing a more comprehensive, accurate, and compliant data foundation for subsequent risk event identification and transmission path tracing, thereby improving the accuracy and reliability of tax risk analysis.

[0077] In some embodiments of the present invention described above, a multimodal temporal knowledge graph based on multimodal data is proposed, using unstructured audio and video data as input. However, in the actual construction process, how to effectively extract key information with temporal characteristics that can reflect the dynamics of enterprise operations from this unstructured audio and video data, and integrate it into the knowledge graph to enhance the precision and real-time performance of risk identification, is a problem that needs further resolution.

[0078] To address this, the present invention further proposes that the multimodal temporal knowledge graph construction steps further include: for the unstructured audio and video data, the extracted temporal feature vectors include: the vehicle entry and exit frequency change curve and the start-up rate / shutdown event identifier of the enterprise's production and operation site obtained based on video stream analysis; the meeting minutes keywords and sentiment score obtained based on audio analysis; and attaching the vehicle entry and exit frequency change curve, shutdown event identifier, and sentiment score as dynamic attributes to the corresponding enterprise entity or event entity in the form of temporal edges.

[0079] Specifically, the unstructured audio and video data refers to unstructured data such as raw video surveillance streams and meeting recordings generated during the daily operations of an enterprise. This data contains rich information about enterprise activities, but it requires specialized processing to transform it into structured features suitable for knowledge graph construction.

[0080] The vehicle entry and exit frequency curve obtained from video stream analysis of enterprise production and operation sites is generated by analyzing the monitoring video streams of the enterprise's production and operation sites in real time or offline, identifying and counting the number of vehicles entering and exiting, and recording the time of occurrence, thereby generating vehicle entry and exit frequency data that changes over time. This can be achieved by using computer vision technologies, such as object detection algorithms (e.g., YOLO, Faster R-CNN), to identify vehicles in video frames, and combining this with object tracking algorithms (e.g., DeepSORT, ByteTrack), to continuously track vehicles and distinguish between repeated counts and newly entering / leaving vehicles. By setting time windows (e.g., hourly, daily), the number of vehicle entries and exits within the window is counted and plotted as a time series curve. This curve can intuitively reflect the intensity of the enterprise's logistics activities and the level of production activity.

[0081] Operating rate / shutdown event identification uses video stream analysis to determine whether a company's production line, workshop, or overall operation is in a normal working (operational) or stopped (shutdown) state, generating corresponding event identifiers. This can be inferred by analyzing the operating status of key equipment in the video (e.g., machine rotation, lights on, personnel activity density, etc.) or by combining abnormal changes in vehicle entry and exit frequency curves. For example, extremely low vehicle entry and exit frequency over a long period, no personnel activity in the production area, and key equipment remaining stationary for extended periods can be identified as shutdown events. Conversely, it indicates an operational state. These identifiers are typically binary or multi-valued discrete events with explicit timestamps.

[0082] Meeting minutes keyword and sentiment scores derived from audio analysis involve processing audio data from internal or external meetings to extract key thematic words and assess speakers' sentiment (e.g., positive, negative, neutral). This can be achieved by using Automatic Speech Recognition (ASR) technology to convert meeting audio into text, followed by Natural Language Processing (NLP) techniques for keyword extraction (e.g., TF-IDF, TextRank) and sentiment analysis (e.g., based on dictionaries, machine learning models, or deep learning models). Keywords reveal the focus of the meeting discussion, while sentiment scores reflect the internal atmosphere and attitudes towards specific events. This information is also timestamped, corresponding to the time the meeting took place.

[0083] Using the vehicle entry / exit frequency variation curve, shutdown event identifier, and sentiment score as dynamic attributes, attached to the corresponding enterprise or event entities as temporal edges, means using the time-series information extracted from audio and video data as attributes of relationships between entities in the knowledge graph, or directly adding it as a new temporal relationship (edge) to the graph. For example, the "vehicle entry / exit frequency variation curve" can be used as a temporal edge connecting an "enterprise entity" and a "time entity," with its weight or attribute value being the frequency at that time point. Alternatively, the "shutdown event identifier" can be used as a temporal edge pointing from an "enterprise entity" to a "shutdown event entity," with an accompanying timestamp of the event. For "meeting minutes keywords and sentiment scores," they can be used as temporal edge attributes connecting "enterprise entities" and "meeting event entities," or a "enterprise-convened-meeting" temporal edge can be directly created, with keywords and sentiment scores as edge attributes. This approach allows the knowledge graph to not only contain static relationships but also dynamically reflect subtle changes in enterprise operations.

[0084] Through the above technical solution, this invention can extract more insightful temporal features from unstructured audio and video data, such as vehicle entry and exit frequency change curves, start-up / shutdown event identifiers, and meeting minutes keywords and sentiment scores. These dynamic attributes are precisely attached to the corresponding enterprise or event entities in the form of temporal edges, greatly enriching the dynamic information dimensions of the multimodal temporal knowledge graph. This allows the knowledge graph to not only reflect static relationships between enterprises but also capture subtle changes in enterprise operational status, internal atmosphere, and potential risks in real time. For example, an abnormal decrease in vehicle entry and exit frequency or a significant increase in negative sentiment in meetings can serve as an early warning signal, assisting the risk event identification process in discovering potential tax risks earlier and more accurately. At the same time, these refined temporal features also provide richer and more accurate dynamic contextual information for subsequent transmission path tracing, making the mining of risk transmission paths more comprehensive and in-depth, thereby improving the real-time performance, accuracy, and early warning capabilities of enterprise tax risk tracking.

[0085] In some embodiments of the present invention described above, incremental data collected in real time is analyzed using a multimodal fusion model to identify potential tax risk events and generate risk feature tensors. However, when processing heterogeneous data from multiple sources, such as structured, semi-structured, and unstructured data, effectively aligning and weighting the risk features of different modalities, and accurately extracting high-confidence risk signals from complex features, is a key challenge to ensure the accuracy and robustness of risk identification. Simple data splicing or linear fusion may not be able to fully capture the complex correlations between modalities and key risk indication information, thereby affecting the accuracy of risk event identification and the effectiveness of subsequent tracking.

[0086] To address this, this invention further proposes a Transformer architecture based on an attention mechanism as the multimodal fusion model for the risk event identification step. This Transformer architecture is a deep learning model whose core lies in its self-attention mechanism, which effectively captures the dependencies between different positions in the input sequence. When applied to multimodal fusion, the model can simultaneously process inputs from different modalities, such as structured tax data, semi-structured text data, and unstructured audio and video data. Through the attention mechanism, it learns the correlations within and between each modality, thereby achieving deeper feature fusion.

[0087] Specifically, the model performs a weighted fusion of statistical outliers from structured data, negative sentiment scores from text data, and anomalous fluctuations from audio and video data. Statistical outliers in structured data refer to data points that significantly deviate from normal patterns, detected by statistical methods (such as Z-score, IQR, or LOF), such as abnormal transaction amounts or significant discrepancies between reported and historical data. Negative sentiment scores in text data are quantified by sentiment analysis of text data (such as contracts, emails, and news reports) using Natural Language Processing (NLP) techniques; high negative sentiment scores may indicate poor business performance or potential risks. Anomalous fluctuations in audio and video data refer to abnormal patterns in unstructured data detected by audio and video analysis techniques, such as abnormal changes in the frequency of people entering and exiting video surveillance, unplanned production line shutdowns, or specific sensitive words and emotional fluctuations appearing in meetings. The Transformer model's attention mechanism adaptively assigns weights to these risk features across different modalities, achieving dynamic weighted fusion and allowing the model to focus on features based on their importance during the fusion process. In this way, the model can generate a unified risk event representation vector that integrates risk information from all modalities, comprehensively representing the currently identified risk events. Subsequently, this representation vector is mapped to a predefined risk category label space, such as "risk of issuing false VAT invoices," "risk of tax evasion," and "risk of broken capital chain," thereby clearly identifying the type of risk.

[0088] Through the above technical solution, this invention overcomes the limitations of traditional fusion methods in effectively capturing complex correlations and key risk indications when processing multi-source heterogeneous data. This model utilizes its self-attention mechanism to deeply learn and dynamically weight statistical outliers from structured data, negative sentiment scores from text data, and abnormal fluctuations from audio and video data. This intelligent weighted fusion method enables the model to more accurately identify features that play a decisive role in risk events across different modalities and effectively suppress noise interference. The resulting unified risk event representation vector not only integrates the advantages of each modality but also significantly improves the accuracy, robustness, and granularity of tax risk event identification by mapping it to a predefined risk category label space, providing a high-quality risk signal starting point for subsequent risk transmission path tracing.

[0089] In some of the above implementations, although a reverse time-series walk algorithm based on time-series constraints has been proposed to uncover risk transmission links, in practical applications, when faced with complex and ever-changing tax risk transmission patterns, especially some highly concealed risk behaviors involving multi-party collaboration, traditional walk algorithms may find it difficult to effectively assess the risk intensity of different paths, and also difficult to proactively discover advanced risk patterns such as fund loops. Furthermore, how to efficiently explore and discover deep, non-explicit transmission paths in a vast knowledge graph remains a challenge.

[0090] To address this, the present invention further proposes that, in the transmission path tracing step under the aforementioned temporal constraints, the reverse temporal walk algorithm specifically includes: employing a path scoring function based on a temporal graph neural network to perform probability evaluation on candidate transmission paths; when it is detected that the flow of funds, invoices, and goods constitutes a closed-loop structure in the time series, the closed-loop path is marked as a "circular transmission" risk pattern and included in the output risk transmission subgraph; the reverse temporal walk algorithm introduces a Monte Carlo tree search mechanism during the walk process to balance the breadth and depth of path exploration, prioritizing the discovery of multi-level bridging transmission paths with a hidden depth exceeding a preset threshold.

[0091] To address the balance between depth and breadth in graph search, the Monte Carlo tree search mechanism introduces a path scoring function based on UCT (Upper Confidence Bounds for Trees) to guide the reverse traversal. For the current node s, its candidate predecessor nodes... The UCT value is calculated as follows: ;in: It is a path value function, representing the path from the current node to the destination. The cumulative confidence level traced back to the source of risk (which can be represented by the weighted sum of the risks at each node along the path); It is the total number of times node s was visited. It is the edge The number of times it is selected, C is the exploration constant, which controls the balance between exploration and utilization.

[0092] Specifically, the path scoring function based on a temporal graph neural network aims to quantitatively evaluate potential risk transmission paths discovered in a multimodal temporal knowledge graph. This function can be designed as a deep learning model, taking as input the entity sequence, relation sequence, and their corresponding timestamp information on the candidate path. Through layer-by-layer processing of a temporal graph neural network (TGNN), the model can capture the dynamic characteristics of information flow along the path and the temporal dependencies of interactions between entities. For example, variants of graph convolutional networks (GCN) or graph attention networks (GAT) can be used, combined with recurrent neural networks (RNN) or Transformer structures to process time-series data, thereby learning complex temporal patterns of the path. Finally, the function outputs a probability value between 0 and 1, representing the likelihood that the path constitutes a tax risk transmission link or its risk intensity. During training, known risk cases can be used as positive samples, and non-risk paths as negative samples. Supervised learning can be used to optimize the model parameters, enabling it to accurately distinguish between risky and non-risky paths.

[0093] Simultaneously, during the execution of the reverse time-series walk algorithm, the system is configured to detect whether the flow of funds, invoices, and goods forms a closed-loop structure in the time series. This closed-loop structure is typically a sophisticated and covert pattern in tax risks, such as inflating turnover through circular transactions, issuing false VAT invoices, or engaging in money laundering. The detection mechanism can check in real-time whether the current path backtracks to previously visited nodes, thus forming a loop, as the walk algorithm traverses the path. Once a loop is detected, the system further analyzes the edge types on the loop to determine whether it simultaneously contains the flow of funds, invoices, and goods, and whether these flows are logically coherent and closed in the time series. For example, funds flow from company A to company B, invoices are issued from company B to company C, and goods flow from company C to company A, and the timestamps of these events satisfy a reasonable causal order. Once such a closed loop is confirmed, the path is marked as a "circular transmission" risk pattern and included in the final output risk transmission subgraph for tax auditors to focus on.

[0094] Furthermore, to optimize the search efficiency and depth of the reverse temporal walk algorithm in large knowledge graphs, this invention introduces a Monte Carlo Tree Search (MCTS) mechanism. MCTS is a heuristic search algorithm that improves search efficiency by balancing exploration (discovering new paths) and utilization (deepening exploration of promising paths) during the search process. MCTS typically includes four phases: selection, expansion, simulation, and backpropagation. In the selection phase, the algorithm selects the most promising child nodes for exploration based on node statistics (such as visit count and success rate). In the expansion phase, new child nodes are generated for the selected nodes, i.e., new propagation path branches are explored. In the simulation phase, starting from the newly expanded nodes, a complete path simulation is completed randomly or using a heuristic strategy until a preset termination condition is reached. Finally, in the backpropagation phase, the simulation results (such as the path's risk score) are backpropagated from the end of the simulated path back to the root node, updating the statistical information of the nodes along the way. By iteratively executing these stages, MCTS can gradually build a search tree and guide subsequent search directions based on node statistics. This allows it to prioritize the discovery of hidden paths with high risk potential, especially multi-level bridging paths whose hidden depth exceeds a preset threshold, while balancing breadth and depth exploration.

[0095] Through the above technical solutions, this invention significantly improves the intelligence, accuracy, and comprehensiveness of tax risk transmission path tracking. The path scoring function based on a time-series graph neural network enables the system to more accurately assess the risk probability of candidate transmission paths and capture complex temporal dependencies, thereby improving the accuracy of risk identification. Simultaneously, by actively detecting the closed-loop structure formed by the flow of funds, invoices, and goods in the time series, this invention can identify the hidden and high-risk tax pattern of "circular transmission," revealing the cyclical flow of funds, invoices, and goods among related enterprises, greatly enhancing risk warning capabilities. Furthermore, the introduction of a Monte Carlo tree search mechanism optimizes the path exploration strategy, balancing the breadth and depth of the search, enabling the system to more efficiently discover deeply hidden and complex transmission paths involving multiple levels of bridging, avoiding the omission of key risk nodes. Overall, these improvements enable this method to more effectively reveal complex and hidden tax risk behaviors, providing more insightful support for tax audits, thereby improving the efficiency and effectiveness of tax risk management.

[0096] The following example will provide a more detailed explanation of the above technical solution:

[0097] In corporate tax risk management scenarios, traditional methods struggle to effectively identify and track the dynamic transmission of risks within complex corporate networks, especially when dealing with multimodal data and unpredictable structures (such as long chains or circular transmission). This method aims to address these issues.

[0098] Suppose the tax authorities notice unusual tax filing behavior from Company A recently, such as a sudden increase in the rate of voided VAT invoices. To analyze this potential tax risk in depth and trace its possible transmission path, this method will proceed according to the following process:

[0099] First, the multimodal data acquisition process is executed. The system connects to a distributed data source to collect multimodal heterogeneous data from the target enterprise group (including enterprise A and its related enterprises B, C, and D). This data includes:

[0100] 1. Structured tax data: such as VAT invoice details, tax returns, and enterprise registration information.

[0101] 2. Semi-structured text data: such as public news reports, social media sentiments, and internal announcements related to companies A, B, C, and D.

[0102] 3. Unstructured audio and video data: such as video surveillance recordings of Company A's business premises, audio recordings of company management meetings, etc.

[0103] To protect data privacy, the system employs a federated learning framework, deploying feature extraction models locally on each data holder's premises. For example, it extracts vehicle entry / exit frequency curves and operational / shutdown event identifiers from company A's video surveillance; and extracts meeting minutes keywords and sentiment scores from meeting audio. Only encrypted model gradients or feature vectors are interacted with, preventing the original tax data from leaving the domain and achieving multi-source data feature alignment while protecting privacy.

[0104] Next, the time-series graph construction step is performed. Based on the collected multimodal data, a multimodal time-series knowledge graph is constructed. This graph includes enterprise entities (enterprises A, B, C, and D), personnel entities (such as corporate legal representatives and senior executives), item entities (such as specific commodities and equipment), and event entities (such as large transactions and equity changes). Transaction relationships from structured tax data (such as enterprise A purchasing goods from enterprise B), dynamic events corresponding to time-series feature vectors extracted from unstructured audio and video data (such as a significant decrease in the frequency of vehicle entry and exit from enterprise A within a certain period, or negative sentiment in management meetings of enterprise C), and changes in entity attributes (such as a change in the legal representative of enterprise B) are stored as time-series edges in the graph database. Each time-series edge contains at least a valid time interval and a relationship type label, such as "enterprise A purchases goods from enterprise B, time interval: January 1, 2023 to January 31, 2023". These time-series edges represent the dynamic characteristics of inter-enterprise relationships changing over time, forming a multimodal time-series knowledge graph that supports time-slice queries.

[0105] Subsequently, the risk event identification step is executed. Based on real-time collected incremental data, risk signals are identified through a multimodal fusion model. For example, the system detects that Company A suddenly had a large number of red-inked and voided invoices in March 2023. This is combined with negative public opinion about Company A in text data and an abnormal decrease in activity at Company A's business premises reflected in video data. The multimodal fusion model (e.g., a Transformer architecture based on an attention mechanism) weights and fuses statistical outliers from structured data (such as invoice anomalies), negative sentiment scores from text data, and abnormal fluctuation features from audio and video data to generate a unified risk event representation vector. This vector is then mapped to a predefined risk category label space, identifying a potential tax risk event of "falsifying VAT invoices" for Company A. This model generates a standardized risk event feature tensor, which includes the risk subject identifier (Company A), risk type (falsifying VAT invoices), confidence level (e.g., 0.92), and occurrence timestamp (March 15, 2023).

[0106] Next, the transmission path tracing step is performed. Starting with the enterprise entity (Enterprise A) pointed to by the risk event feature tensor, a time backtracking window (e.g., 6 months) and a transmission depth threshold (e.g., 3 layers of association) are set, and a reverse time-series walk algorithm is executed in the multimodal temporal knowledge graph. When traversing the path, the algorithm requires that the timestamps of all temporal edges on the path satisfy monotonically decreasing and fall within the time backtracking window. For example, starting from the risk event of Enterprise A (March 15, 2023), the reverse tracing is performed: it is found that Enterprise A received a large amount of funds from Enterprise B in February 2023, Enterprise B signed an unusually large purchase contract with Enterprise C in January 2023, and the legal representative of Enterprise C had an equity relationship with Enterprise D in December 2022. The candidate transmission paths are probabilistically evaluated by a path scoring function based on a temporal graph neural network, and a Monte Carlo tree search mechanism is introduced to balance the breadth and depth of path exploration, prioritizing the discovery of multi-level bridging transmission paths with a hidden depth exceeding the preset threshold. If the flow of funds, invoices, and goods is detected to form a closed-loop structure in time series (e.g., Company D provided services to Company A in January 2023), this closed-loop path is marked as a "circular transmission" risk pattern. Ultimately, the complete transmission link from the risk source (e.g., Company D) to the current risk outbreak point (Company A) is uncovered, and a risk transmission subgraph containing intermediate transmission nodes (Company C, Company B) is output. Compared to traditional static graph analysis, this method can accurately capture the transmission order and correlation strength of risks in the time dimension, avoiding misjudgments or omissions due to a lack of time series information.

[0107] Finally, a visualization step is performed. The risk transmission subgraph is rendered visually, intuitively displaying the source of risk, intermediate transmission nodes, and the current outbreak point. In the risk transmission subgraph, nodes and edges are differentiated by risk attenuation coefficients; for example, the risk source company D is marked in dark red, intermediate bridging companies C and B in orange, and the risk receiver company A in yellow. The location and time of key abnormal documents (such as company A's red-inked invoices and company C's abnormal purchase contracts) are also marked. Simultaneously, based on the path characteristics in the risk transmission subgraph, a structured audit task instruction set containing evidence collection guidance is automatically generated. This instruction set includes: an on-site verification checklist for company D, marked as the "risk source," focusing on verifying its operational authenticity; bank statement lookup instructions for "intermediate bridging" companies C and B to track fund flows; and a cooperation notification template for company A, the "risk receiver," requiring it to provide more detailed transaction vouchers. This visualized and automated audit instruction generation significantly improves the efficiency and accuracy of tax audits, solving the problem that traditional methods struggle to automatically generate explainable risk transmission paths.

[0108] The following are specific examples:

[0109] Background: The tax bureau of a certain city detected an anomaly in Company A's system—a significant mismatch between its input invoices and production capacity—triggering a high-risk warning. This method is now needed to trace the source and transmission path of the risk.

[0110] Example steps:

[0111] Step 1: Multimodal Data Acquisition and Feature Extraction

[0112] Data source access: The system accesses structured data from the Golden Tax Phase III system (invoices and tax returns of Company A), internet public opinion data (news about layoffs at Company A), and factory surveillance video (vehicle entry and exit logs of Company A and its affiliated company B).

[0113] Feature extraction:

[0114] Using a video analytics model, we extracted the vehicle entry and exit frequency curves for Company A over the past three months, finding a continuous downward trend. The average value for the past week was 70% lower than the previous month, generating a time-series feature vector. .

[0115] Sentiment analysis was performed on news items related to "Company A is facing operational difficulties," resulting in a score of -0.8 (extremely negative). Text features were generated. .

[0116] The VAT burden rate was extracted from Company A's tax return, and its statistical outlier score compared to the industry average was 0.95 (extremely high).

[0117] Step 2: Construction of Multimodal Temporal Knowledge Graph

[0118] The system constructs a knowledge graph encompassing three companies: A, B, and C. Specifically:

[0119] Timeline edge 1: (Company A, Purchasing, Company B), time interval [2024-01-01, 2024-05-31], attributes: purchase amount 5 million, invoice code xxx...).

[0120] Timeline edge 2: (Company B, actual controller is Wang), time interval [2018-01-01, present].

[0121] Timing edge 3: (Company A, monitored, 2024-06-05), associated event entity "sudden drop in production capacity", attribute attached to vehicle frequency feature vector. With a confidence level of 0.9.

[0122] Step 3: Risk Event Identification

[0123] Real-time fusion analysis: For Company A, the multimodal fusion model (Transformer architecture) aligns and weights the features of the three modalities (outlier 0.95, sentiment -0.8, and production fluctuation -70%). The attention weights are: structured 0.5, text 0.3, and audio / video 0.2.

[0124] Generate risk tensor: fused feature vector After MLP mapping, the probability distribution in the risk label space is as follows: false invoicing probability 0.8, tax evasion probability 0.1, and normal probability 0.1.

[0125] Output risk event: The system generates a risk event feature tensor: (Risk subject = "Company A", Risk type = "Suspected of issuing false invoices", Confidence = 0.85, Occurrence timestamp = 2024-06-07 15:30:00).

[0126] Step 4: Propagation path tracing under timing constraints

[0127] Setting parameters: Starting from Company A, set the time backtracking window. Heaven, conduction depth threshold .

[0128] Reverse timing walk (combined with MCTS):

[0129] First jump: Starting from company A, search for its upstream companies within the time window. It is found that company B (a supplier) issued a large invoice to company A on t=2024−05−20, and the legal representative of company B is Wang. According to the UCT formula, this path has a high Q-value (risk accumulation). Because they are fewer in number, their UCT values ​​are higher, making them a preferred choice.

[0130] The second step: Continuing the reverse search from Company B, it was discovered that Company B's input invoices mainly came from Company C (characteristics of a shell company), and the actual controller of Company C was also Wang. At this point, it was found that the cash flow (A->B->C) and invoice flow (C->B->A) formed a closed-loop structure. The system calculates the circular risk score. The risk level is abnormally high and is marked as a "circular transmission" risk pattern.

[0131] Further investigation revealed that the funds ultimately flowed to Wang's overseas account (depth exceeding 3). This was due to MCTS's depth reward factor. This hidden, deep path has been successfully uncovered.

[0132] Output subgraph: Output a complete risk transmission subgraph containing the sequence "Wang (overseas account) <- Company C (shell company) <- Company B (bridging loan) <- Company A (risk outbreak point)".

[0133] Step 5: Visualization and Audit Report Generation

[0134] Differentiated rendering: In the transmission subgraph, the source "Wang's overseas account" is colored dark red (high risk). (High), Companies C and B, according to the risk decay function They were dyed orange and yellow respectively, and Company A was dyed red.

[0135] Automatically generate audit instruction sets:

[0136] Source of risk: Generating a "funds look-through query instruction for Wang and his overseas accounts".

[0137] Intermediate bridge: Generate an instruction to "verify each bank statement of Company B from January to May 2024, with a focus on verifying the authenticity of transactions with Company C".

[0138] Risk receiving end: Generates "A notice of assistance will be sent to the competent tax authority in the location of Company A, requesting verification of the contradictory reasons for the decline in production capacity and the surge in invoices".

[0139] Key evidence: Mark “Company C -> Company B invoice (2024-05-20)” and “Company A surveillance video production capacity inflection point (2024-05-25)” on the sub-graph as key abnormal documents and time points.

[0140] In a pilot application at a municipal tax bureau, this embodiment achieved the following results:

[0141]

[0142] Based on the table above, the system successfully identified a "ring-shaped" invoice fraud chain involving 23 companies. This chain, through six intermediaries, ultimately resulted in the issuance of 120 million yuan in fraudulent input invoices to the core company, A. Traditional methods can only identify direct upstream and downstream entities, while this method, through reverse tracing under time constraints, successfully traced back to the source cluster of shell companies.

[0143] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0144] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made within the scope of the claims of the present invention should be included in the scope of the present invention.

Claims

1. A method for tracing enterprise tax risk paths based on multimodal temporal knowledge graphs, characterized in that: Includes the following steps: Multimodal data acquisition steps: Collect multimodal data from distributed data sources, including structured tax data, semi-structured text data, and unstructured audio and video data; The steps for constructing a time-series knowledge graph are as follows: Based on the multimodal data, a multimodal time-series knowledge graph is constructed. The entity relationship edges in the time-series knowledge graph carry timestamp attributes, which are used to characterize the dynamic characteristics of the relationship between enterprises over time. Risk event identification steps: Analyze real-time data using a multimodal fusion model to identify potential tax risk events and generate risk feature tensors; Transmission path tracing steps: Starting from the enterprise entity corresponding to the risk feature tensor, perform a reverse search based on temporal constraints in the multimodal temporal knowledge graph to mine and output the transmission path subgraph of risk among related enterprises; Visualization steps: Visualize and render the transmission path sub-graph, and generate an audit report containing risk tracing information; The time-series graph construction steps are further specified: A knowledge graph containing enterprise entities, personnel entities, item entities, and event entities is constructed; transaction relationships, dynamic events corresponding to time-series feature vectors, and entity attribute changes in the structured tax data are stored as time-series edges in the graph database, with each time-series edge containing at least a valid time interval and a relationship type label, forming a multimodal time-series knowledge graph that supports time-slice queries; the multimodal time-series knowledge graph includes: Entity set Includes enterprise entity e, personnel entity p, item entity g, and event entity ev; time-series edge. Each edge is defined as a quadruple: edgev = (h, r, time, attrs), where Subject and object For relational types, The effective time interval is defined, and attrs are dynamic attributes. The risk event identification steps are further specified: based on real-time collected incremental data, risk signals are identified through a multimodal fusion model, and risk features from different modalities are aligned and fused to generate a standardized risk event feature tensor. This tensor includes the risk subject identifier, risk type, confidence level, and occurrence timestamp. The multimodal fusion model includes extracting multimodal features X= By performing feature alignment and fusion through a cross-attention layer, a risk feature tensor R is generated: R = MultiHeadAttention(X,X,X). Risk event features are generated by mapping the data to a risk category space using a multilayer perceptron (MLP): risk_event= =softmax(MLP(R)); The transmission path tracing step is further specified as follows: taking the enterprise entity pointed to by the risk event feature tensor as the starting point, setting the time backtracking window ΔT and the transmission depth threshold, and executing the reverse time-series walk algorithm in the multimodal time-series knowledge graph; when traversing the path, the reverse time-series walk algorithm requires that the timestamps of all time-series edges on the path satisfy monotonically decreasing and fall within the time backtracking window, thereby mining the complete transmission link from the risk source to the current risk outbreak point, and outputting a risk transmission subgraph containing intermediate transmission nodes; The reverse time-series walk algorithm: based on the risk subject Starting from the time rewind window Inside, perform a reverse walk, with the walk path P= ,in It must satisfy a strict time monotonically decreasing constraint: , and all timestamps .

2. The enterprise tax risk path tracing method based on multimodal temporal knowledge graph as described in claim 1, characterized in that: The multimodal data acquisition steps are further specified as follows: accessing a distributed data source to collect multimodal heterogeneous data of the target enterprise group; the multimodal heterogeneous data includes structured tax data, semi-structured text data, and unstructured audio and video data; using a pre-trained feature extraction model, extracting temporal feature vectors from the unstructured audio and video data respectively.

3. The enterprise tax risk path tracing method based on multimodal temporal knowledge graph as described in claim 1, characterized in that: The visualization step is further specified as follows: the risk transmission subgraph is visualized and rendered, and a structured audit task instruction set containing evidence collection guidance is automatically generated based on the path features in the risk transmission subgraph; in the risk transmission subgraph, nodes and edges are differentiated according to the risk attenuation coefficient, and the location and time of key abnormal documents are marked. The structured audit task instruction set includes: an on-site verification list for enterprises marked as "risk sources", instructions for bank statement lookup for "intermediate bridge" enterprises, and a cooperation notification template for "risk receiving end" enterprises.

4. The enterprise tax risk path tracing method based on multimodal temporal knowledge graph as described in claim 2, characterized in that: The multimodal heterogeneous data acquisition and feature extraction steps specifically include: using a federated learning framework to deploy feature extraction models locally on each data holder's site; only exchanging encrypted model gradients or feature vectors to avoid the original tax data going out of domain, thus achieving multi-source data feature alignment under privacy protection.

5. The enterprise tax risk path tracing method based on multimodal temporal knowledge graph as described in claim 1, characterized in that: The multimodal temporal knowledge graph construction steps further include: for the unstructured audio and video data, the extracted temporal feature vectors include: the vehicle entry and exit frequency change curve and the start-up rate / shutdown event identifier of the enterprise's production and operation site obtained based on video stream analysis; the meeting minutes keywords and sentiment score obtained based on audio analysis; and attaching the vehicle entry and exit frequency change curve, shutdown event identifier, and sentiment score as dynamic attributes to the corresponding enterprise entity or event entity in the form of temporal edges.

6. The enterprise tax risk path tracing method based on multimodal temporal knowledge graph as described in claim 1, characterized in that: The multimodal fusion model in the risk event identification step is a Transformer architecture based on an attention mechanism; The model performs weighted fusion of statistical outliers in structured data, negative sentiment scores in text data, and abnormal fluctuation features in audio and video data to generate a unified risk event representation vector, which is then mapped to a predefined risk category label space.

7. The enterprise tax risk path tracing method based on multimodal temporal knowledge graph as described in claim 1, characterized in that: In the time-constrained transmission path tracing step, the reverse time-series walk algorithm specifically includes: using a path scoring function based on a time-series graph neural network to perform probability evaluation on candidate transmission paths; when it is detected that the flow of funds, invoices, and goods form a closed loop structure in the time series, the path of the closed loop structure is marked as a "circular transmission" risk pattern and included in the output risk transmission subgraph; the reverse time-series walk algorithm introduces a Monte Carlo tree search mechanism during the walk process to balance the breadth and depth of path exploration, and prioritizes the discovery of multi-level bridging transmission paths with a hidden depth exceeding a preset threshold; the Monte Carlo tree search mechanism introduces a path scoring function based on UCT to guide the reverse walk in order to solve the problem of balancing depth and breadth in graph search. For the current node s, its candidate predecessor nodes The UCT value is calculated as follows: ;in: It is a path value function, representing the path from the current node to the destination. The cumulative confidence level tracing back to the source of the risk; It is the total number of times node s was visited. It is the edge The number of times it is selected, C is the exploration constant, which controls the balance between exploration and utilization.

Images