A method and system for detecting and warning of monopolistic violations based on spatiotemporal spectrum

By employing a spatiotemporal spectrum-based method for detecting and warning against monopolistic violations, and utilizing graph neural networks and multi-head attention layers to adaptively adjust edge weights, interpretable warning signals and evidence chains are generated. This solves the problems of accuracy and traceability of spatiotemporal spectrum in existing technologies, thereby improving the regulatory effectiveness of monopolistic violations.

CN123118129APending Publication Date: 2026-11-17SHANGHAI JIAOTONG UNIV +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611626857.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing early warning systems use fixed parameters when constructing dynamic relationship maps between enterprises, making it difficult to accurately reconstruct behavioral relationships at different depths and cycles in the spatiotemporal spectrum. They cannot adaptively adjust time-series weights, resulting in insufficient accuracy and timeliness of early warning signals. Furthermore, they cannot automatically backtrack and verify, thus limiting the regulatory application of anti-monopoly and illegal activities.

Method used

A method for detecting and warning of monopolistic violations based on spatiotemporal spectrum is adopted. An abnormal node is identified by generating embedded vectors through graph neural network. The decay factor is calculated by combining multi-head attention layer and multi-layer perceptron network, and the edge weights are adaptively adjusted to generate warning signals and trace the evidence chain to locate legal provisions.

Benefits of technology

It achieves end-to-end connectivity from multi-source heterogeneous data to executable early warning signals, improving the accuracy and timeliness of early warnings, meeting the interpretability and traceability of evidence required by legal supervision, and solving the technical bias and application bottlenecks of traditional models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118129A_ABST
    Figure CN123118129A_ABST
Patent Text Reader

Abstract

This invention relates to a method and system for detecting and warning of monopolistic violations based on a spatiotemporal spectrum, belonging to the field of data processing technology. It solves the problems of low reliability and high false alarm rate in existing warning systems. The method includes: constructing a temporal multidirected graph with market entities as nodes and each moment's interaction behavior as an edge; calculating edge weights based on the type and timing of the associated interaction behavior; inputting the temporal multidirected graph of the time to be analyzed into a trained graph neural network to generate embedding vectors for each node; identifying abnormal nodes exhibiting monopolistic violations and generating warning signals for them; the warning signals include warning trigger edges and abnormal feature vectors; performing forward tracing and backward expansion using the warning trigger edges as seeds in the temporal multidirected graph, and obtaining the optimal evidence chain based on edge weights; and locating the relevant legal provisions for each abnormal node from a legal knowledge base through multi-level screening. This improves the accuracy and interpretability of the warning.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for detecting and warning of monopolistic violations based on spatiotemporal spectrum. Background Technology

[0002] With the rapid development of the platform economy, the competitive behavior of platform companies is becoming increasingly complex, covert, and dynamic, posing significant challenges to traditional market regulation methods. "Computational antitrust," as an emerging interdisciplinary field, attempts to leverage artificial intelligence and big data technologies to enhance regulatory effectiveness.

[0003] Existing early warning systems often employ a unified Time Series Graph Network (TGN) model when constructing dynamic relationship graphs between enterprises. These models typically pre-determine a fixed, globally uniform time decay parameter when processing historical interaction effects. After identifying anomalies based on the graph neural network, they usually only output a high-dimensional anomaly feature vector or anomaly score, and its generation process is uninterpretable for regulatory personnel.

[0004] Because monopolistic behavior inherently possesses strong temporal dependence, rigorous legal proof, and heterogeneous characteristics, the impact cycles of different types of behavior on the spatiotemporal spectrum vary greatly. For example, "equity investment" is a structural behavior, and its influence in the spectrum decays very slowly; while "instantaneous search ranking adjustment" is a strategic behavior, and its influence smooths out rapidly over time. Existing technologies have significant technical defects and shortcomings in achieving accurate, interpretable, and implementable early warning functions, including: the use of fixed parameters in existing technologies makes it difficult to accurately reconstruct behavioral relationships at different depths and cycles in the constructed spatiotemporal spectrum, easily leading to the loss of key long-term evidence or interference from short-term noise; the fixed and singular temporal perception granularity of the early warning system cannot adaptively adjust the temporal weights according to the nature of the behavior, seriously impairing the accuracy and timeliness of the early warning signal; after identifying anomalies, it cannot automatically trace back a connected evolutionary path that satisfies temporal directional constraints from the spatiotemporal graph, directly resulting in the early warning signal being unreliable in regulatory practice, greatly limiting the practical application value of the technical solution; and the early warning conclusions are disconnected from the mapping of legal provisions, lacking hierarchical and precise positioning capabilities. Summary of the Invention

[0005] In view of the above analysis, the embodiments of the present invention aim to provide a method and system for perceiving and warning of monopolistic violations based on spatiotemporal spectrum, in order to solve the problems of low reliability and high false alarm rate of existing monopolistic violation warnings due to the inability to adapt to behavior type due to temporal decay, the inability to backtrack and verify the warning logic, and the lack of hierarchical and accurate positioning of legal provisions.

[0006] On the one hand, embodiments of the present invention provide a method for detecting and warning of monopolistic violations based on a spatiotemporal spectrum, comprising the following steps: Market entities and their interaction behavior data at various times are extracted from market supervision data. A time-series multidirected graph is constructed with market entities as nodes and interaction behavior at each time as an edge. The edge weight is calculated based on the type of interaction behavior and the interaction time associated with the edge. The temporal multidirected graph of the time to be analyzed is input into a trained graph neural network to generate embedding vectors for each node. Abnormal nodes with monopolistic and illegal behaviors are identified from these vectors, and warning signals are generated for them. The warning signals include warning trigger edges and abnormal feature vectors. In a temporal multidirected graph, forward tracing and backward expansion are performed using the warning trigger edge as the seed, and the optimal evidence chain is obtained based on the edge weight. Based on the abnormal feature vectors and optimal evidence chains of each abnormal node, the legal provisions involved in the monopolistic violations of each abnormal node are located from the legal knowledge base through multi-level screening.

[0007] Based on the above method, the edge weights are calculated using an adaptive decay function based on the interaction behavior type, as shown in the following formula: , in, Head node With tail node During the interaction time The type of interaction behavior that occurred is The edge is waiting to be analyzed. edge weights, Head node With tail node Interaction behavior types The strength of the original relationship, Interaction type The corresponding attenuation factor, Time to be analyzed Interaction time with edges The time difference between them.

[0008] Based on the further improvement of the above method, the decay factor corresponding to the interaction behavior type is obtained in the following way: the interaction behavior type is encoded into a behavior type embedding vector and then fed into a multilayer perceptron network, which is then processed by the Softplus function to obtain the result; the multilayer perceptron network is obtained by joint training with a graph neural network.

[0009] Based on the further improvement of the above method, the edge weights of the incoming edges between each node and each of its neighboring nodes are summarized to obtain the aggregated weights of the incoming edges of each node relative to each of its neighboring nodes; a multi-head attention layer is embedded in the graph neural network; after the temporal multi-directed graph to be analyzed is input into the graph neural network, the aggregated weights of the incoming edges of each node are used as prior coefficients through the multi-head attention layer to weight and aggregate the neighborhood information of each node, generating the embedding vector of each node at the time to be analyzed.

[0010] Based on further improvements to the above method, abnormal nodes exhibiting monopolistic violations are identified, including: Add the nodes corresponding to each merchant entity in the market entity to the candidate node set; By matching the attributes and embedding vectors of candidate nodes, a base node group matching each candidate node is selected from the candidate node set, and the average embedding vector of the base node group is taken as the base vector. Calculate the distance between the embedding vector of each candidate node and its corresponding reference vector to obtain the anomaly score, and identify candidate nodes whose anomaly scores exceed a preset anomaly threshold as anomaly nodes.

[0011] Based on the further improvement of the above method, the abnormal feature vector is obtained by concatenating the difference vector and the context embedding vector; wherein, the difference vector is obtained by calculating the difference component between the embedding vector of the abnormal node and its corresponding reference vector; the context embedding vector is calculated based on the embedding vectors of the two ends of the warning trigger edge and its first-order neighbor nodes.

[0012] Based on the further improvement of the above method, the warning trigger edge is obtained in the following way: according to the aggregate weight of the incoming edge of the abnormal node and the attention score output by the multi-head attention layer, the incoming edge neighbor node that contributes the most to the embedding vector of the abnormal node is determined; among the multiple directed edges between the incoming edge neighbor node and the abnormal node, the incoming edge with the largest edge weight is taken as the warning trigger edge.

[0013] Based on further improvements to the above method, the optimal chain of evidence is obtained through the following steps: The baseline time is the interaction time associated with the edge that triggers the early warning; Using the warning trigger edge as the initial path, search for candidate incoming edges from the starting point of the initial path in a decreasing direction along the base time, and search for candidate outgoing edges from the ending point of the initial path in a increasing direction along the base time, until the preset termination condition is met. Each candidate incoming edge, initial path, and candidate outgoing edge are combined according to the direction and temporal constraints of the edges to obtain multiple candidate paths. The sum of the weights of all edges in the candidate paths is calculated to obtain the total weight of the path. The candidate path with the largest total weight is selected as the optimal evidence chain.

[0014] Based on further improvements to the above method, the legal provisions related to monopolistic violations at each abnormal node are located from the legal knowledge base through multi-level screening, including: The abnormal feature vectors are mapped to the legal semantic space and matched with the vectors of various legal provisions in the legal knowledge base to recall several candidate legal provisions. Explicit factual text is extracted from the optimal chain of evidence. Semantic similarity is calculated between the explicit factual text and the text at each level of each candidate legal provision. The weighted total score is obtained and the candidate legal provisions are reordered. Several target legal provisions are selected from the candidate legal provisions. The complete original texts of several target legal provisions and the optimal chain of evidence are input into a large language model that has been fine-tuned by law. The large language model performs semantic verification on the clause-level and / or item-level content of each target legal provision and outputs the applicable clause-level and / or item-level legal provisions.

[0015] On the other hand, embodiments of the present invention provide a monopolistic violation perception and early warning system based on a spatiotemporal spectrum, comprising: The data processing module is used to extract market entities and their interaction behavior data at various times from market supervision data. It constructs a time-series multi-directed graph with market entities as nodes and interaction behavior at each time as an edge. The edge weight is calculated based on the type of interaction behavior associated with the edge and the interaction time. The anomaly detection module is used to input the temporal multidirected graph of the time to be analyzed into a trained graph neural network, generate the embedding vector of each node, identify the abnormal nodes with monopolistic and illegal behaviors, and generate warning signals for them. The warning signals include: warning trigger edges and abnormal feature vectors. The evidence extraction module is used to perform forward tracing and backward expansion in a time-series multidirected graph, using the warning trigger edge as a seed, and to obtain the optimal evidence chain based on the edge weight. The legal provision location module is used to locate the relevant legal provisions for monopolistic violations of each abnormal node from the legal knowledge base through multi-level filtering, based on the abnormal feature vectors and optimal evidence chains of each abnormal node.

[0016] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. A complete technical closed loop from multi-source heterogeneous data to executable early warning signals has been constructed. With anomaly feature vectors and early warning trigger edges as the hub, end-to-end connection from raw data to executable early warning signals has been achieved, transforming scattered market supervision data into early warning conclusions that serve law enforcement decision-making. It breaks through the limitation of data collection, anomaly detection and legal application being mutually exclusive in existing technologies, enabling early warning conclusions to directly support the law enforcement decisions of regulatory personnel. From a technical perspective, it solves the bottleneck problem that makes it difficult to implement "computational anti-monopoly" solutions.

[0017] 2. To address the differences in the timeliness characteristics of different interaction behavior types, an adaptive decay factor calculation mechanism was designed to realize adaptive temporal modeling of behavior type perception. By encoding the behavior type identifier into an embedded vector and inputting it into a multilayer perceptron network, the decay factor is automatically optimized during joint training, overcoming the technical bias of fixed decay parameters. This allows long-term structural impact behaviors such as equity investment to maintain a higher weight over a longer period, while short-term strategic behaviors such as search ranking adjustments decay rapidly. This significantly improves the adaptability accuracy and early warning timeliness of the temporal graph network to anti-monopoly regulatory scenarios.

[0018] 3. Guided by the comprehensive contribution determined by the multi-head attention aggregation coefficient and the prior entry edge weight, the core edge that triggers the warning is locked. Through a multi-hop backtracking algorithm with temporal direction constraints, the maximum weight path of the connected multi-directed graph is automatically extracted as the evidence chain, making the generation process of the warning conclusion completely transparent and traceable. This solves the pain point of "black box" decision-making and difficult interpretation of results in traditional AI models, and meets the rigid requirements of legal supervision for the traceability of evidence.

[0019] 4. The abnormal feature vectors are mapped to the legal semantic space for coarse screening and recall. Then, the explicit factual text in the evidence chain is used to perform weighted fine ranking of the legal provisions index at each level. Finally, the large language model verifies the constituent elements of the candidate legal provisions and outputs the clause / item level location. This three-level architecture refines legal retrieval from "global vector matching" to "hierarchical semantic verification", which significantly improves the accuracy and interpretability of legal provision location.

[0020] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0021] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a flowchart of a method for perceiving and warning of monopolistic violations based on spatiotemporal spectrum in Embodiment 1 of the present invention. Detailed Implementation

[0022] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0023] Example 1 A specific embodiment of the present invention discloses a method for detecting and warning of monopolistic violations based on a spatiotemporal spectrum, such as... Figure 1 As shown, it includes steps S1-S4.

[0024] S1. Extract market entities and their interaction behavior data at various times from market supervision data. Construct a time-series multi-directed graph with market entities as nodes and the interaction behavior at each time as an edge. The edge weight is calculated based on the interaction behavior type and interaction time associated with the edge.

[0025] This embodiment collects raw data related to platform enterprises, merchants and their business activities from public data sources and authorized business data sources.

[0026] Specifically, publicly available data sources include, but are not limited to: enterprise registration information in the National Enterprise Credit Information Publicity System; administrative penalty decisions and merger control review announcements published by market supervision departments; periodic reports and interim announcements of listed companies (used to identify equity investment and control relationships); search ranking rules, traffic allocation mechanisms, and operational policy announcements publicly released by platform companies; reports from third-party news media and industry-specific media, as well as publicly available public opinion information on social media platforms; and publicly obtainable operational metrics such as merchant exposure rates, click-through rates, and transaction rankings. Raw data from these publicly available data sources is collected through data crawling or API interfaces.

[0027] Authorized business data sources refer to business operation data with higher data density obtained after authorization from the data subject, including but not limited to: platform internal merchant operation reports, transaction flow summaries, advertising placement records, etc. The original data is obtained after anonymization and compliance verification, such as: masking or generalizing sensitive fields such as personal identification information and trade secrets; confirming the legality of the data source and verifying the scope of data use.

[0028] It should be noted that the collected raw data is mostly unstructured or semi-structured text. Named entity recognition and relation extraction techniques are used, and large language models are used to assist in structured extraction when necessary. Market entities and their interaction behavior data at various times are extracted from the raw data to construct entities, entity attributes, and the relationships between entities (interaction behavior and interaction time).

[0029] Specifically, the types of market entities include: platforms and merchants, respectively, as platform entities and merchant entities; among them, platform entities include: e-commerce platforms, search platforms, social platforms, and other platform enterprises. The attributes of platform entities include, but are not limited to: the effective time of the rules extracted from the platform's search ranking rules or operational policy announcements, whether the rules contain exclusivity clauses, bundling clauses, or restricted transaction clauses, the type code of the ranking mechanism or traffic allocation mechanism involved, and the scope of merchants to which the rules apply (such as all merchants or merchants of specific categories). A merchant entity is a merchant or brand that has joined the platform and conducts business activities. The attributes of a merchant entity include, but are not limited to: platform identifier (a unique code recording the platform to which the merchant has joined), main product category code (recording the core product or service category of the merchant, such as "home appliances", "food and beverage", "clothing and apparel", etc.), and business scale level code (a scale range divided according to objective indicators such as the merchant's annual total transaction volume or registered capital, such as "large", "medium", "small"); exposure rate time series, click-through rate time series, sales time series extracted from operational data, and statistical characteristics further extracted from these time series data, such as mean, mutation point location, month-on-month growth rate, etc.

[0030] In addition, public opinion attributes are extracted from reports by third-party news media, industry vertical media, and public opinion information on social media platforms. These attributes include: the time of the event, the type of emotional polarity reflected in the report (positive / negative / neutral, which can be automatically determined using a pre-trained sentiment classification model), whether the report mentions bundled sales or restricted transactions, and the names of the entities involved in the report. These public opinion attributes are then added to the attributes of the corresponding entities as shared attributes of the platform entities and / or merchant entities involved.

[0031] It should be noted that the relationships between entities include: interaction type and interaction time. Interaction type includes, but is not limited to: bundling (including traffic bias and self-preferential treatment implemented by improving one's own or specific merchants' search rankings, providing special traffic entrances, etc.), equity investment, and cooperation. Since entities may engage in the same or one type of interaction at different times, the relationships between entities are represented by a four-tuple, including: head entity, interaction type, tail entity, and interaction time.

[0032] Furthermore, after completing entity identification, attribute extraction, and relationship extraction, a temporal multi-directed graph is constructed using the extracted market entities as nodes (i.e., market nodes and merchant nodes) and each extracted interaction behavior at any given moment as an edge. "Temporal" refers to the fact that each edge is associated with an interaction moment (i.e., an interaction timestamp), recording the specific time when the interaction behavior occurred. "Multi-" refers to the fact that multiple edges may be generated between the same pair of nodes due to different behavior types (such as bundling, equity investment, and cooperation), or due to the same behavior type occurring repeatedly at different times. "Directed" refers to the fact that each edge uniquely corresponds to a four-tuple of relationship information in the graph, with the head entity pointing to the tail entity.

[0033] As can be seen from the above, in this embodiment, "space" refers to the multidimensional structured attribute feature space of the platform and merchant nodes; "time" refers to the sequence of interaction moments. The spatiotemporal spectrum is the dynamic relational topology of entity nodes evolving over time in a high-dimensional feature space. The constructed temporal multidirected graph is the specific physical implementation of this "spatiotemporal spectrum" at the data structure level.

[0034] It is important to emphasize that each edge in the temporal multidirected graph of this embodiment is assigned a numerical value reflecting its importance at the time of analysis, i.e., edge weight. The edge weights evolve dynamically over time, and the evolution rates of different interaction types are fundamentally different.

[0035] Specifically, the edge weight is calculated using an adaptive decay function based on the interaction type and interaction time associated with the edge, as shown in the following formula: , in, Head node With tail node During the interaction time The type of interaction behavior that occurred is The edge is waiting to be analyzed. The edge weights have a range of [0,1]. The larger the value, the higher the contribution intensity of the interaction behavior at the time to be analyzed. Head node With tail node Interaction behavior types The original relation strength, with a value range of [0,1]; Interaction type The corresponding attenuation factor, Time to be analyzed Interaction time with edges The time difference between them.

[0036] It should be noted that in the formula for calculating edge weights, the original relationship strength of the interaction type between nodes reflects the absolute scale when the interaction occurs, is not affected by time decay, and is pre-calculated using the following method: For equity investment behavior (head node) Investment tail end ), with head node The shareholding ratio or investment amount is used as the benchmark value; for tying arrangements (head entity) Tail-end entity Apply flow tilt), to tail body The exposure rate deviation (i.e., the excess of the exposure rate compared to the average of similar categories during the period of traffic tilt) is used as the benchmark value; for cooperative behavior (head node) With tail node (Cooperation), with head node The frequency or amount of cooperation initiated is used as the baseline value; each type of interaction behavior is independently normalized using Min-Max to map the corresponding baseline value to the [0,1] interval to obtain the original relationship strength.

[0037] It should be noted that the decay factor corresponding to the interaction behavior type is obtained as follows: After encoding the interaction behavior type into a behavior type embedding vector, it is fed into a multilayer perceptron network. After transformation by several fully connected layers and nonlinear activation functions, a scalar value is output, which is then processed by the Softplus function to obtain the final value, as shown in the formula below: , in, This represents a multilayer perceptron; Indicates the type of interaction behavior The corresponding behavior type embedding vector; The function is defined as The output range is (0, +∞), which ensures that the decay factor is always positive, in accordance with the physical constraints of time decay.

[0038] The multilayer perceptron network is obtained through end-to-end joint training with a graph neural network that identifies monopolistic violations. This allows the decay factor to be adaptively adjusted according to different regulatory scenarios: for long-term structural impact behaviors such as equity investment, the multilayer perceptron network is trained to output a smaller decay factor value, so that its influence can be maintained over a longer period of time; for strategic short-term behaviors such as search ranking adjustments, the multilayer perceptron network is trained to output a larger decay factor value, so that its influence fades quickly over time, thereby focusing on recent key changes.

[0039] Preferably, the multilayer perceptron network introduces prior knowledge such as "the equity investment decay rate is relatively small and the search ranking adjustment decay rate is relatively large" as an initialization or regularization constraint in the early stage of training to accelerate network convergence and improve training stability.

[0040] After calculating the edge weights, the topological information in the graph data structure is organized into an input format usable by the graph neural network, including: constructing a node feature matrix, an adjacency index table, an edge attribute feature matrix, and direction labels.

[0041] Specifically, since platform nodes and merchant nodes have different attribute types (discrete flags, time-series values, natural language text) and different dimensions, a pre-trained multimodal attribute encoder is used to uniformly map all kinds of original attributes to a feature space of the same dimension. In the multimodal attribute encoder, discrete categorical attributes (such as mechanism type encoding, flags) are mapped into dense vectors using an embedding layer; time-series numerical attributes (such as exposure rate, click-through rate sequences) are mapped into fixed-length vectors using a time-series encoder (such as an LSTM or Transformer encoder); and natural language text attributes (such as the main name of a public opinion report) are encoded into semantic vectors using a text encoder (such as BERT). The output dimension of each encoder is unified, and after concatenation or fusion, the initial feature representation of each node is obtained, constructing a node feature matrix.

[0042] Furthermore, a globally unique index is generated for each node, and a globally unique edge number is generated for each directed edge, corresponding one-to-one with the quadruple association of the directed edge. The quadruple association of the directed edges is then stored in a sparse format, including: Convert each edge into a tuple of (head node index, tail node index), forming a shape of... The adjacency index tensor, where The total number of edges in the graph; if there are multiple directed edges between the same pair of nodes (corresponding to different interaction behavior types or different interaction times), since each directed edge has a different edge number, multiple records are retained in the adjacency index tensor, each corresponding to a different interaction behavior.

[0043] Simultaneously, for each edge, its interaction behavior type identifier and interaction time are encoded into a dimension of . Numerical vectors, numbered by edge Each of the adjacent index tensors corresponds one-to-one, forming a shape of The edge attribute feature matrix is ​​calculated; the edge weight of each edge is calculated as an additional attribute of the edge and stored separately by edge number.

[0044] The direction of each directed edge is fixed so that it points from the head entity (behavior initiator) to the tail entity (behavior receiver).

[0045] S2. Input the temporal multidirected graph of the time to be analyzed into the trained graph neural network to generate the embedding vector of each node. Identify abnormal nodes with monopolistic and illegal behaviors and generate warning signals for them. The warning signals include: abnormal feature vectors and warning trigger edges. S21. Generate the embedding vector for each node.

[0046] It should be noted that the temporal multidirected graph constructed in step S1 is a continuously accumulating and evolving dynamic graph structure. New nodes and edges are constantly added to the graph as market participants and their interactions occur. In actual early warning inference, only the time period to be analyzed needs to be considered. (i.e., the current moment or any specified historical backtracking moment) is used for slice analysis. Specifically, all interaction moments are extracted from the complete graph structure. The edges and all nodes associated with these edges constitute the time interval to be analyzed. The temporal multidirected graph is obtained, and the corresponding node feature matrix, adjacency index table, edge attribute feature matrix and direction identifier are obtained.

[0047] Since the temporal multidirected graph constructed in step S1 allows for multiple directed edges with different directions between the same pair of nodes (corresponding to different interaction types or different interaction times), and in the message passing process of the graph neural network, the information of neighboring nodes is aggregated in node pairs as the basic unit. Therefore, based on the direction of the directed edges, the weights of the incoming edges between each node and each of its neighboring nodes are summarized to obtain the aggregated weights of the incoming edges of each node relative to each of its neighboring nodes, as shown in the following formula: , in, For nodes Relative to its neighboring nodes At the moment of analysis The aggregate weight of the incoming edges, For neighboring nodes Pointing to node The set of incoming edges.

[0048] The graph neural network in this embodiment embeds a multi-head attention layer and adopts a prior-guided multi-head graph attention architecture. This network is an improvement on the standard multi-head graph attention network. The aggregation coefficient is no longer determined entirely by the attention score, but by the attention score and the prior aggregation weight. That is, through the multi-head attention layer, the aggregation weight of the incoming edge of each node is used as the prior coefficient to perform weighted aggregation of the neighborhood information of each node, generating the embedding vector of each node at the time to be analyzed.

[0049] Specifically, the input layer of the graph neural network receives the node feature matrix, adjacency index table, and aggregation weights between node pairs corresponding to the temporal multidirected graph at the time to be analyzed.

[0050] Furthermore, in the multi-head attention layer, the number of attention heads is... For the first A person's attention In the Layer, for nodes and its incoming neighbor nodes First, the original attention score under this attention head is calculated using a learnable transformation matrix.

[0051] Specifically, the nodes and its incoming neighbor nodes In the eigenvectors of the layer and Through learnable transformation matrices Perform a linear transformation, then concatenate the transformed vectors, and finally combine them with the learnable attention vector. The original attention score is obtained by calculating the dot product and then activating it with LeakyReLU. : , in, This represents a vector concatenation operation. This indicates the transpose operation.

[0052] Furthermore, Softmax normalization is performed on the original attention scores of all neighboring nodes to obtain the... Attention score under each attention head This reflects the strength of the association between nodes, which is automatically learned from the data.

[0053] Furthermore, the incoming edge aggregation weight of each node is used as a prior coefficient and multiplied by the attention score to obtain the prior-guided incoming edge aggregation coefficient. This coefficient is then used to perform a weighted summation of the neighborhood information (i.e., features of the incoming edge neighboring nodes) of each node, and the summation is then processed through a learnable value transformation matrix. and nonlinear activation functions (such as LeakyReLU, ELU, Sigmoid, etc.) to obtain nodes In the New feature representation under attention head: , , in, For the first Each attention node Relative to the incoming neighbor node In the Layer polymerization coefficient, For nodes The incoming neighbor node.

[0054] Multi-head attention mechanism through Each independent attention head executes the above process in parallel, and the output features of each head are concatenated to form a node. In the The new embedding vector of the layer, after being propagated through all multi-head attention layers, yields the node. At the moment of analysis Dynamic embedding vectors .

[0055] This embodiment introduces a multi-head attention mechanism with prior coefficients, which organically integrates data and priors, enabling the generated node embedding to maintain data-driven flexibility while possessing clear physical interpretability.

[0056] It should be noted that the graph neural network (GNN) is pre-trained jointly with the multilayer perceptron network (MLP) that outputs the decay factor corresponding to the interaction behavior type in step S1 to optimize the learnable parameters in the network.

[0057] Before training, training sample sets for two networks were constructed based on historical market supervision cases. Each training sample contains merchant nodes, a temporal multidirected graph within a specific time interval, and corresponding anomaly annotation information.

[0058] The sample sources are divided into two categories: positive sample cases are from historically concluded cases that were found to have engaged in monopolistic violations and were subject to administrative penalties or judicial rulings, and are marked as "abnormal"; negative sample cases are from groups of normally operating enterprises that have been manually verified and have not engaged in any monopolistic violations within the same time span, and are marked as "normal". Before being input into the model, each training sample completes data extraction, temporal multidirected graph construction, and feature encoding according to the method in step S1, and is organized into an input format supported by the graph neural network.

[0059] When constructing the training sample set, all interactive behavior data are extracted from the positive sample cases within a time window before the incident (e.g., 12 months before the incident) to construct a time-series multidirected graph; for negative samples, normal operating periods with similar time spans and market entity sizes to the positive samples are selected to construct corresponding time-series multidirected graphs.

[0060] During training, the embedding vectors of each node in the graph are obtained during forward propagation. During backpropagation, the model parameters are updated by calculating the contrastive learning loss based on the positive and negative examples corresponding to each sample, and then using a gradient descent optimizer. For example, for each merchant node in a sample, the embedding vectors of several nodes are randomly sampled from merchant nodes of the same category and size that have not engaged in monopolistic or illegal activities as negative examples, while the embedding vectors of that node itself under different time windows are used as positive examples.

[0061] After training, the transformation matrices, attention vectors, value transformation matrices of each attention head in the graph neural network and the network parameters of the multilayer perceptron are permanently saved.

[0062] S22. Identify abnormal nodes with monopolistic or non-compliant behavior and generate early warning signals.

[0063] Specifically, abnormal nodes are identified through the following steps: ① Add the nodes corresponding to each merchant entity in the temporal multidirected graph at the time to be analyzed to the candidate node set.

[0064] ② By matching the attributes and embedding vectors of candidate nodes, a baseline node group matching each candidate node is selected from the candidate node set, and the average embedding vector of the baseline node group is taken as the baseline vector.

[0065] It should be noted that each merchant node corresponds to a baseline node group, which is used to measure whether the merchant's behavioral characteristics deviate significantly from the normal level.

[0066] The baseline node group is selected as follows: For each candidate node, merchant nodes that are on the same platform, in the same product category, and at the same business scale level as the candidate node and have not engaged in monopolistic or illegal activities are selected as candidate nodes. The cosine similarity between the candidate node and the embedding vector of each candidate node is calculated, and after sorting them in descending order, several candidate nodes with a cosine similarity greater than or equal to the similarity threshold are selected to form the baseline node group.

[0067] If the number of nodes in the baseline node group is insufficient (e.g., less than the preset minimum number threshold), the constraints should be adjusted appropriately, such as increasing the selection range of business scale levels and lowering the feature similarity threshold, to ensure that the baseline node group has sufficient statistical representativeness.

[0068] Furthermore, the mean of the embedding vectors of all nodes in the reference node group is calculated, that is, the average embedding vector is taken as the reference vector.

[0069] ③ Calculate the distance between the embedding vector of each candidate node and its corresponding reference vector to obtain the anomaly score, which ranges from [0,2]. Candidate nodes whose anomaly scores exceed the preset anomaly threshold are designated as anomaly nodes.

[0070] In this embodiment, cosine distance is preferably used as the metric for anomaly score. Cosine distance measures the degree of difference between two vectors in direction, is insensitive to the absolute value of features, and can effectively avoid interference from differences in feature magnitude caused by objective factors such as business scale between different businesses in anomaly detection.

[0071] The anomaly threshold is adjusted by regulators based on actual regulatory needs, or calibrated using historical labeled data to achieve a balance between accuracy and recall.

[0072] If the anomaly score is less than or equal to the anomaly threshold, the candidate node is considered normal and no warning signal is generated; otherwise, the candidate node is considered abnormal, added to the abnormal node set, and a warning signal is generated for each abnormal node in the set. The warning signal includes: the warning trigger edge and the anomaly feature vector.

[0073] It should be noted that the warning trigger edge is the specific interaction event that causes the target node to be judged as abnormal. Based on the aggregate weight of the incoming edge of the abnormal node and the attention score output by the multi-head attention layer (the last layer), the incoming edge neighbor node that contributes the most to the embedding vector of the abnormal node is determined. Among the multiple directed edges between the incoming edge neighbor node and the abnormal node, the incoming edge with the largest edge weight is taken as the warning trigger edge.

[0074] Specifically, due to abnormal nodes Embedded vector It is determined by its incoming neighbor nodes. The embedding vectors are aggregated using prior-guided input edge aggregation coefficients; for all input edge neighbors of an abnormal node, the coefficients of each input edge neighbor are calculated separately. The average aggregation coefficient under each attention head is used to select the incoming neighbor node with the largest average aggregation coefficient as the incoming neighbor node that contributes the most to the embedding vector of the abnormal node. .

[0075] Furthermore, from the node pair Among the multiple directed edges, the incoming edge with the largest weight is selected as the warning trigger edge. Based on the edge change corresponding to the warning trigger edge, the head node index, interaction behavior type identifier, tail node index (i.e., abnormal node) and interaction time are obtained.

[0076] It should be noted that the abnormal feature vector is obtained by concatenating the difference vector and the context embedding vector. The difference vector is obtained by calculating the difference component between the embedding vector of the abnormal node and its corresponding reference vector. It reflects the deviation direction and magnitude of the abnormal node and its corresponding reference node group in the feature space and is the core basis for measuring the abnormal deviation of market indicators caused by monopolistic behavior. The context embedding vector is calculated based on the embedding vectors of the two ends of the warning trigger edge and its first-order neighbor nodes. The first-order neighbor nodes are all nodes directly connected to any end node of the warning trigger edge, regardless of the direction of the edge. This provides the most complete local network environment semantics possible and avoids missing circumstantial information related to monopolistic behavior due to directional limitations.

[0077] Specifically, the two ends of the warning trigger edge As two root nodes, each root node's first-order neighbor node is taken as a child node, and both the root node and its child nodes are placed into the context node set. In the first step, calculate the shortest path distance from each child node to each root node, and then calculate the weighting coefficient for that child node based on the shortest path distance. Next, sum the weighted embedding vectors of all nodes in the context node set, and then divide by the sum of the weighting coefficients to obtain a fixed-length context embedding vector. The formula is shown below: , in, For the context node set nodes The weighting coefficients (the shortest distance from the root node to itself) (The weighting factor is 1, and the value is 0). For the context node set nodes The embedding vector.

[0078] The difference vector and the context embedding vector are concatenated along the feature dimension to obtain the complete anomaly feature vector.

[0079] S3. In a time-series multidirected graph, the early warning trigger edge is used as a seed for forward tracing and backward expansion, and the optimal evidence chain is obtained based on the edge weight.

[0080] It should be noted that, using the warning trigger edge as a seed, a multi-hop backtracking expansion is performed along the temporal direction in the temporal multidirected graph to generate a connected and traceable optimal evidence chain, which is used to explain the complete causal evolution process of this monopolistic violation.

[0081] Specifically, the interaction time associated with the warning trigger edge is taken as the base time; the warning trigger edge is taken as the initial path, and candidate incoming edges are searched from the starting point of the initial path in a decreasing direction along the base time, and candidate outgoing edges are searched from the ending point of the initial path in an increasing direction along the base time, until the preset termination condition is met.

[0082] It should be noted that during forward tracing, the first step is to search the set of incoming edges connected to the starting point of the initial path. Preferably, these edges are sorted by weight, and the edge with the highest weight is selected first. Candidate inbound edges ( For example, to preset the extended width or Add it to the current path. For each candidate incoming edge added, the head node of each candidate incoming edge is used as the endpoint of the next round of forward tracing.

[0083] During backward expansion, the first step is to search the set of outgoing edges connected to the endpoint of the initial path. Preferably, these outgoing edges are sorted by weight, and the first outgoing edge with the highest weight is selected. Candidate out edges ( For example, to preset the extended width or Add it to the current path. For each candidate outgoing edge added, the tail node of each candidate outgoing edge is used as the starting point for the next round of backward expansion.

[0084] Repeat the forward tracing and backward expansion process described above to form a bidirectional, alternating iterative expansion process. Each round starts from the two endpoints of the current path (the endpoint with the earliest timestamp and the endpoint with the latest timestamp), allowing the chain of evidence to grow simultaneously in both the past and future directions.

[0085] Before each round of expansion, check if any of the following termination conditions are met; if they are met, stop further expansion in the corresponding direction: there are no candidate incoming edges or candidate outgoing edges, or the total number of edges in the current path reaches the preset maximum search depth (e.g., 5).

[0086] Each candidate incoming edge, initial path, and candidate outgoing edge are combined according to the direction and temporal constraints of the edges to obtain multiple candidate paths. The sum of the weights of all edges in the candidate paths is calculated to obtain the total weight of the path. The candidate path with the largest total weight is selected as the optimal evidence chain.

[0087] Furthermore, based on a pre-set evidence template, the structured information carried by the optimal evidence chain is transformed into natural language evidence items that can be directly read and understood by regulatory personnel.

[0088] Specifically, the following rendering operations are performed on the optimal chain of evidence: Extract the head node index, interaction behavior type label, tail node index, and interaction time for each edge, thereby obtaining the subject name corresponding to the head node and tail node, as well as the behavior description text mapped to the interaction behavior type; obtain the fingerprint of the original data source corresponding to the edge, including but not limited to: database table name and primary key ID, file storage path, official announcement URL, or link and archive identifier of third-party reports; mark the role of each edge in the optimal evidence chain, including: cause (edge ​​obtained by tracing forward from the warning trigger edge), core event (warning trigger edge), and consequence (edge ​​obtained by expanding backward from the warning trigger edge). For the warning trigger edge in the optimal evidence chain, extract the attribute change information of the abnormal node (i.e., the tail node) before and after the event from its structured attributes. This information serves as supporting evidence for the quantitative analysis of market damage caused by the core event, including: The interaction time of the warning trigger edge is taken as the start time of the observation window, and the interaction time of the last backward expansion edge in the time to be analyzed or the optimal evidence chain is taken as the end time of the observation window; and the time period of the same length before the observation window is taken as the baseline window. Obtain the mean, peak, and trend (e.g., month-on-month growth rate) of key operational metrics (at least including exposure rate, click rate, and sales) of abnormal nodes within the observation window, and compare them with the corresponding metrics in the baseline window to calculate the magnitude of change in key operational metrics.

[0089] For example, the evidence template is as follows: Evidence [Edge Number] ([Role]): [The subject name of the head node of the current evidence] at [the moment of interaction] and [the subject name of the tail node of the current evidence] [interaction behavior description text]; Data source: [Original data source fingerprint].

[0090] Supporting evidence: During the period from the start time of the observation window to the end time of the observation window, the attribute name of the [abnormal node] shows the [direction and magnitude of change], which is [percentage change] compared to the baseline period from the start time of the baseline window to the end time of the baseline window.

[0091] The optimal chain of evidence after rendering is as follows: Evidence 1 (Background): Platform A and Merchant B released new search ranking rules on January 15, 2025; Data source: Announcement system; Announcement ID: XZ-2025-001.

[0092] Evidence 2 (Core Event): Platform A established a bundling relationship with Merchant B on January 20, 2025; Data source: Platform rule base; Rule record ID: Rule-045; Extraction time: 2025-01-21.

[0093] Evidence 3 (Consequence): Merchant B signed a distribution agreement with downstream distributor D on January 21, 2025; Data source: Enterprise contract database; Contract number: HT-2025-0121; Registration date: 2025-01-22.

[0094] Supporting evidence: Merchant B's exposure rate remained consistently high from February 1, 2025 to February 28, 2025, averaging about 120% higher; compared with the baseline period from January 1, 2025 to January 31, 2025, it increased by 180%.

[0095] S4. Based on the abnormal feature vectors and optimal evidence chains of each abnormal node, locate the legal provisions involved in the monopolistic violations of each abnormal node from the legal knowledge base through multi-level screening.

[0096] S41. Construct a legal knowledge base.

[0097] It should be noted that each legal text involved is pre-analyzed hierarchically. Considering the differences in the internal structure of different laws (e.g., the Anti-Monopoly Law is structured as "Chapter → Article → Clause → Item," while the Civil Code is structured as "Part → Chapter → Section → Article → Clause → Item"), and that some "Articles" may only contain "Clause" or "Item," or both, this embodiment extracts the core hierarchy common to all laws as the index hierarchy to achieve unified cross-legal retrieval, namely: Chapter → Article → Block.

[0098] Among them, the "Chapter" level is the chapter title of the law (such as "Chapter 3 Abuse of Market Dominance" in the Anti-Monopoly Law); the "Article" level is the number and introductory text of each article of the law; the "Block" level uniformly refers to the smallest semantic unit under the "Article", including "Section" and "Item", and the two are treated as a "Block" without further distinguishing between Section and Item.

[0099] The original text structure of each law is parsed using regular expressions, extracting the text content at the chapter, article, and block levels, and generating a globally unique level identifier for each level to maintain the relationship between levels.

[0100] Specifically, chapter-level text is extracted by chapter, using the chapter title as the unit. For laws with a "section" structure, the chapter-level text is "section name + chapter name" (e.g., "Part One General Provisions Chapter One Basic Provisions"); for laws without a "section", the chapter-level text is the chapter name (e.g., "Chapter Three Abuse of Market Dominance"). When a chapter title is scanned, a new chapter identifier is generated by combining it with the identifier of the relevant law (e.g., LAW001_CH03). The text at the item level is organized by "item". The introductory phrase (the first paragraph of the main text, including the item number but excluding clauses and items) is extracted from each item, such as "Article 22 prohibits operators with a dominant market position from engaging in the following acts of abusing their dominant market position." When an introductory phrase is detected, a new item identifier (e.g., LAW001_CH03_ART22) is generated, and the chapter identifier to which that item belongs is recorded. The block-level text treats each clause and item under each entry as an independent "minimum semantic block," extracting the main text of that block (i.e., the complete paragraph text of the "clause" or the complete sentence text of the "item"), such as "(iv) Without justifiable reason, the counterparty to the transaction is restricted to only transacting with it." When a clause or item is scanned, a new block identifier (such as LAW001_CH03_ART22_BLK04) is generated, and the clause identifier and chapter identifier to which the block belongs are recorded. Through the above sequential scanning and context maintenance, a complete hierarchical path record is constructed for each level entry.

[0101] Furthermore, a pre-trained legal language model (such as Lawformer or Legal-BERT) is used to encode the text content at each level into corresponding vectors. Each chapter-level entry records its chapter identifier and corresponding chapter-level vector; each article-level entry records its article identifier, its chapter identifier, and corresponding article-level vector; and each block-level entry records its block identifier, its article identifier, its chapter identifier, and corresponding block-level vector. This information is then uniformly stored in a legal knowledge base.

[0102] It should be noted that weights are preset for three different levels. , and Furthermore, the sum of the three weights is 1. The weights at each level reflect their differences in importance in semantic matching. In principle, the "block" level (i.e., the specific behavioral description of a clause or item) contributes the most to the matching of legal provisions, followed by the "article" level, with the "chapter" level serving as a general thematic reference. For example, the default value is set to... , , .

[0103] S42. Filter legal provisions.

[0104] It should be noted that this step, based on the abnormal feature vectors and optimal evidence chains of each abnormal node, uses multi-level screening to locate the legal provisions related to the monopolistic violations of the abnormal nodes from the legal knowledge base, including: ① Map the abnormal feature vectors to the legal semantic space, perform similarity matching with the vectors of various legal provisions in the legal knowledge base, and recall several candidate legal provisions.

[0105] It should be noted that this embodiment employs a multi-layer fully connected network, taking the anomaly feature vector as input and mapping it to a behavioral semantic vector with the same dimension as the vectors at each level in the legal knowledge base. The multi-layer fully connected network and the legal language model form a dual-tower structure, undergoing joint fine-tuning within a contrastive learning framework. This aligns the output vectors of both within a unified "legal semantic space," supporting cross-modal similarity comparisons.

[0106] Specifically, the legal language model, as a legal provision-side encoder, has undergone large-scale self-supervised pre-training on a large number of legal documents and regulations (such as the MLM task), and has the basic semantic understanding ability of legal language. It can accurately understand the meaning of legal terms such as "abuse of market dominance", "merger of undertakings", and "tying". The multi-layer fully connected network, as a behavior-side encoder, receives abnormal feature vectors and outputs behavior semantic vectors.

[0107] Each training sample includes: anomaly feature vectors constructed from historically concluded, verifiable monopoly cases, with the corresponding applicable legal texts (article-level text content) as positive samples, and inapplicable legal texts randomly selected from the legal knowledge base as negative samples.

[0108] During training, the contrastive learning InfoNCE loss function is used to maximize the cosine similarity of positive sample pairs while minimizing the similarity of negative sample pairs, so that the vectors output by the two encoders are aligned in a unified semantic space.

[0109] After training, a trained legal language model is obtained, which is used in step S41 to perform offline batch encoding of legal text content at each level.

[0110] During the screening process, the abnormal feature vector from step S3 is input into the trained multi-layer fully connected network and mapped to a query vector. The cosine similarity is then calculated between the query vector and the text vector at each level in the legal knowledge base, and the multiple candidate legal provisions with the highest cosine similarity are recalled.

[0111] ② Extract explicit factual text from the optimal evidence chain, calculate the semantic similarity between the explicit factual text and the text at each level of each candidate legal provision, obtain the weighted total score at each level, reorder each candidate legal provision, and select several target legal provisions from several candidate legal provisions.

[0112] It should be noted that extracting explicit factual text from the optimal chain of evidence includes: based on predefined keywords, traversing each evidence item in the optimal chain of evidence, and extracting the keywords contained therein, such as "tying," "restricted transaction," "exclusivity clause," "traffic bias," etc., to form explicit factual text.

[0113] Based on the identifier of each candidate legal provision, obtain all blocks under that provision and their corresponding block identifiers; for each block, through its recorded "corresponding chapter identifier" and "corresponding article identifier", index the corresponding chapter-level vector and article-level vector from the legal knowledge base to obtain the chapter-level vector, article-level vector and block-level vector to which the block-level entry belongs, respectively.

[0114] A trained lightweight semantic matching model (such as Sentence-BERT) is used to obtain the vectors of explicit factual text. The semantic similarity of the explicit factual text vectors with chapter-level vectors, article-level vectors, and block-level vectors is calculated. The weighted hierarchical total score for each block is calculated using the following formula: , in, , and The semantic similarity is calculated between the vectors of explicit factual text and the chapter-level, article-level, and block-level vectors, respectively.

[0115] Since a single legal provision may contain multiple blocks, to prevent a single legal provision from crowding out the recall quota of other legal provisions due to excessive blocks, this embodiment performs max pooling aggregation on a "provision" basis: the highest weighted level total score among all blocks under the same legal provision is taken as the fine-ranking score of that candidate legal provision. Finally, the multiple candidate legal provisions with the highest fine-ranking scores are selected as the target legal provisions. If the number of candidate legal provisions is less than the number threshold, the candidate legal provisions are sorted from high to low according to their fine-ranking scores.

[0116] ③ Input the complete original text of several target legal provisions and the optimal chain of evidence into the large language model that has been fine-tuned by law. The large language model performs semantic verification on the clause-level and / or item-level content under each target legal provision and outputs the applicable clause-level and / or item-level legal provisions.

[0117] This step involves using a large language model to verify the constituent elements of multiple selected target legal provisions, and then precisely pinpointing the warning conclusion to a specific "section" or "item".

[0118] First, based on the complete text of several target legal provisions and the optimal chain of evidence, structured prompts are constructed, including: The text corresponding to the best chain of evidence is used as factual basis to provide a complete description of the market behavior process, the relationship between the parties and the consequences; Each target legal provision is marked with a serial number, and the complete hierarchical path of the target legal provision (law name → chapter name → article number) and the complete original text are attached as the legal basis to provide the legal normative text to be verified. Define verification criteria for the subject, interactive behavior, and interactive result; Define the output format requirements for various verification results.

[0119] For example, the prompt words are as follows: You are an assistant in determining the applicability of legal provisions. Below is a complete set of factual evidence (chain of evidence) and several legal provisions (target provisions). Based on the facts and referring to the provisions, please examine each target provision to determine whether the abnormal behavior meets the constituent elements of a particular clause or provision.

[0120] Factual basis: [Embedded optimal chain of evidence text] Legal basis: Target provision 1: [Complete hierarchical path] [Complete original text] Target provision 2: [Complete hierarchical path] [Complete original text] ... Target legal provision J: [Full hierarchical path] [Full original text] Judgment criteria: Subjective requirement: Whether the subject of the act in the evidence meets the subject type stipulated in the legal provisions; Behavioral requirement: Whether the conduct in the evidence falls under the category of conduct prohibited by law; Results requirement: Does the evidence contain market damage consequences or exclusion / restriction of competition effects required by law?

[0121] Output format requirements: If the target legal provision is applicable as a whole, output the complete hierarchical path of the target legal provision, and output all specific clauses or items that match the target legal provision (if there are multiple, they are arranged in descending order of matching degree), and briefly explain the reason for the match of each clause / item; the output format is: [Law Name] [Chapter Name] [Article Introduction Summary] [Block Number and Text Summary]; If the target legal provision is not applicable as a whole, explicitly output "Not Applicable"; If none of the target legal provisions apply, explicitly output "No match".

[0122] The constructed prompts are input into a legally fine-tuned large language model. The large language model verifies the constituent elements of each target legal provision based on the complete facts in the chain of evidence. Since the prompts provide multiple target legal provisions, the large language model performs a horizontal comparison of each target legal provision, prioritizing the legal provision that best matches the evidentiary facts and is precise down to the clause or item, along with the corresponding reasoning. If the large language model determines that multiple clauses or items are applicable, all are retained and arranged in descending order of matching degree. If the large language model determines that a certain target legal provision is applicable but does not explicitly specify the clause or item, the entire target legal provision is recorded as a "provisional-level hit".

[0123] This embodiment leverages the strength of large language models in understanding the semantics and logical relationships of long texts. It inputs complete chains of evidence (rather than fragmented keywords) and the full text of multiple target legal provisions, enabling it to perform element reasoning based on a complete legal normative context and a complete factual narrative, thus avoiding misjudgments caused by taking things out of context. Furthermore, inputting multiple target legal provisions simultaneously for horizontal comparison by the large language model can improve the accuracy and exclusivity of its judgments.

[0124] Preferably, this embodiment also calculates the comprehensive confidence level of the warning signal and outputs a complete warning conclusion to the regulatory personnel.

[0125] It should be noted that the overall confidence score is obtained by normalizing the evidence chain integrity score, path strength score, anomaly mapping score, and fine-ranking score, then weighting and summing them, and finally mapping them using the Sigmoid function.

[0126] Among them, the score for the completeness of the chain of evidence. The value is quantified into different preset values ​​based on whether the warning trigger edge has a forward-in edge and a backward-out edge, including: only a warning trigger edge, only a warning trigger edge and at least one forward-in edge or backward-out edge, and having a warning trigger edge, a forward-in edge, and a backward-out edge simultaneously; for example: if there is only a warning trigger edge, then If there is only an edge that triggers the warning and at least one forward in edge or backward out edge, then If there are warning trigger edges, forward inbound edges, and backward outbound edges simultaneously, then .

[0127] Path strength score It is obtained by performing Min-Max normalization on the total weight of the path corresponding to the optimal evidence chain.

[0128] Abnormal mapping score This is obtained by linearly mapping the abnormal scores of abnormal nodes to the [0,1] interval, for example, by dividing the abnormal scores with values ​​in the range [0,2] by 2.

[0129] Ranking score The result is obtained based on the maximum refined score of the target legal provision. If the large language model confirms that one or more target legal provisions are applicable, the maximum refined score is taken from the target legal provisions that are confirmed to be applicable; otherwise, the maximum refined score is taken from all target legal provisions.

[0130] The confidence level is calculated using the following formula. : , in, For the first The preset weights of the items satisfy The default values ​​are set to 0.25, 0.25, 0.2 and 0.3 respectively, and the specific values ​​are optimized through verification using historical annotation data.

[0131] Finally, a complete early warning conclusion is output, including: an early warning summary, the optimal chain of evidence, and the location of legal provisions. The early warning summary includes: the name of the abnormal node, its platform, its anomaly score, and its overall confidence level. The optimal chain of evidence is a traceable natural language evidence item generated through rendering. The location of legal provisions includes: the applicable legal provisions confirmed by the large language model, and a list of specific clauses or items that match under that legal provision (if there are multiple, they are arranged in descending order of matching degree), along with the reasons for the matching of each clause / item.

[0132] Compared with existing technologies, this embodiment provides a method for detecting and warning of monopolistic violations based on spatiotemporal spectrum. It constructs a complete technical closed loop from multi-source heterogeneous data to executable warning signals. Using anomaly feature vectors and warning trigger edges as hubs, it achieves end-to-end connectivity from raw data to executable warning signals, transforming scattered market regulatory data into warning conclusions that serve law enforcement decision-making. It overcomes the limitations of existing technologies where data collection, anomaly detection, and legal application are disconnected, enabling warning conclusions to directly support regulatory personnel's law enforcement decisions. This technically solves the bottleneck problem of the difficulty in implementing "computational anti-monopoly" solutions. Addressing the differences in the timeliness characteristics of different interaction behavior types, an adaptive decay factor calculation mechanism is designed, achieving adaptive temporal modeling of behavior type perception. By encoding behavior type identifiers as embedded vectors and inputting them into a multilayer perceptron network, the decay factor is automatically optimized during joint training, overcoming the technical bias of fixed decay parameters. This allows long-term structural impact behaviors such as equity investment to maintain higher weights over a longer period, while short-term strategic behaviors such as search ranking adjustments decay rapidly. This significantly improves the adaptability accuracy of the temporal graph network to anti-monopoly regulatory scenarios and the timeliness of warnings. Guided by the comprehensive contribution determined by the multi-head attention aggregation coefficient and the prior incoming edge weights, the core edge triggering the early warning is identified. A multi-hop backtracking algorithm with temporal directional constraints automatically extracts the connected, maximum-weighted path from the temporal multidirected graph as the evidence chain, making the generation process of the early warning conclusion completely transparent and traceable. This solves the pain points of traditional AI models' "black box" decision-making and difficulty in interpreting results, meeting the rigid requirements of legal supervision for the traceability of evidence. Abnormal feature vectors are mapped to the legal semantic space for coarse screening and recall. Then, explicit factual text in the evidence chain is used to perform weighted fine-grained ranking of legal provisions at each level. Finally, a large language model verifies the constituent elements of candidate legal provisions and outputs the clause / item level location. This three-level architecture refines legal retrieval from "global vector matching" to "hierarchical semantic verification," significantly improving the accuracy and interpretability of legal provision location.

[0133] Example 2 Another embodiment of the present invention discloses a spatiotemporal spectrum-based anti-monopoly behavior perception and early warning system, thereby implementing the spatiotemporal spectrum-based anti-monopoly behavior perception and early warning method in Embodiment 1. The specific implementation of each module is described in the corresponding description in Embodiment 1. The system includes: The data processing module is used to extract market entities and their interaction behavior data at various times from market supervision data. It constructs a time-series multi-directed graph with market entities as nodes and interaction behavior at each time as an edge. The edge weight is calculated based on the type of interaction behavior associated with the edge and the interaction time. The anomaly detection module is used to input the temporal multidirected graph of the time to be analyzed into a trained graph neural network, generate the embedding vector of each node, identify the abnormal nodes with monopolistic and illegal behaviors, and generate warning signals for them. The warning signals include: warning trigger edges and abnormal feature vectors. The evidence extraction module is used to perform forward tracing and backward expansion in a time-series multidirected graph, using the warning trigger edge as a seed, and to obtain the optimal evidence chain based on the edge weight. The legal provision location module is used to locate the relevant legal provisions for monopolistic violations of each abnormal node from the legal knowledge base through multi-level filtering, based on the abnormal feature vectors and optimal evidence chains of each abnormal node.

[0134] Since the monopolistic violation perception and early warning system based on spatiotemporal spectrum in this embodiment and the aforementioned monopolistic violation perception and early warning method based on spatiotemporal spectrum can be mutually referenced, this description is redundant and will not be repeated here. Because this system embodiment shares the same principle as the aforementioned method embodiment, it also possesses the corresponding technical effects of the aforementioned method embodiment.

[0135] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0136] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monopoly violation behavior perception early warning based on space-time spectrum, characterized in that, Includes the following steps: Market entities and their interaction behavior data at various times are extracted from market supervision data. A time-series multidirected graph is constructed with the market entities as nodes and the interaction behavior at each time as an edge. The edge weight is calculated based on the interaction behavior type and interaction time associated with the edge. The temporal multidirected graph of the time to be analyzed is input into a trained graph neural network to generate embedding vectors for each node. Abnormal nodes with monopolistic violations are identified from these vectors, and warning signals are generated for them. The warning signals include: warning trigger edges and abnormal feature vectors. In the temporal multidirected graph, forward tracing and backward expansion are performed using the warning trigger edge as the seed, and the optimal evidence chain is obtained according to the edge weight. Based on the abnormal feature vectors of each abnormal node and the optimal evidence chain, the legal provisions involved in the monopolistic violations of each abnormal node are located from the legal knowledge base through multi-level screening.

2. The method of claim 1, wherein the method is characterized by, The edge weights are calculated using an adaptive decay function based on the interaction behavior type, as shown in the following formula: , wherein, is the head node and the tail node interact at the interaction time , the interaction behavior type between the head node and the tail node at the analysis time is the edge weight, is the original relationship strength between the head node and the tail node of the interaction behavior type is the decay factor corresponding to the interaction behavior type , and is the time difference between the analysis time and the interaction time of the edge.

3. The method for detecting and warning of monopolistic violations based on spatiotemporal spectrum according to claim 2, characterized in that, The decay factor corresponding to the interaction behavior type is obtained by encoding the interaction behavior type into a behavior type embedding vector, then feeding it into a multilayer perceptron network and processing it through the Softplus function; the multilayer perceptron network is obtained by joint training with the graph neural network.

4. The method for detecting and warning of monopolistic violations based on spatiotemporal spectrum according to claim 1, characterized in that, The weights of the incoming edges between each node and each of its neighboring nodes are aggregated to obtain the aggregated weights of the incoming edges of each node relative to each of its neighboring nodes; the graph neural network has a multi-head attention layer embedded in it; after the temporal multi-directed graph at the time to be analyzed is input into the graph neural network, the aggregated weights of the incoming edges of each node are used as prior coefficients through the multi-head attention layer to perform weighted aggregation of the neighborhood information of each node, and generate the embedding vector of each node at the time to be analyzed.

5. The method for detecting and warning of monopolistic violations based on spatiotemporal spectrum according to claim 1, characterized in that, The identified abnormal nodes exhibiting monopolistic violations include: Add the nodes corresponding to each merchant entity in the market entity to the candidate node set; By matching the attributes and embedding vectors of candidate nodes, a base node group matching each candidate node is selected from the candidate node set, and the average embedding vector of the base node group is taken as the base vector. Calculate the distance between the embedding vector of each candidate node and its corresponding reference vector to obtain the anomaly score, and identify candidate nodes whose anomaly scores exceed a preset anomaly threshold as anomaly nodes.

6. The method for detecting and warning of monopolistic violations based on spatiotemporal spectrum according to claim 5, characterized in that, The abnormal feature vector is obtained by concatenating the difference vector and the context embedding vector; wherein, the difference vector is obtained by calculating the difference component between the embedding vector of the abnormal node and its corresponding reference vector; the context embedding vector is calculated based on the embedding vectors of the two ends of the warning trigger edge and its first-order neighbor nodes.

7. The method for detecting and warning of monopolistic violations based on spatiotemporal spectrum according to claim 4, characterized in that, The warning trigger edge is obtained in the following way: based on the aggregate weight of the incoming edge of the abnormal node and the attention score output by the multi-head attention layer, the incoming edge neighbor node that contributes the most to the embedding vector of the abnormal node is determined; among the multiple directed edges between the incoming edge neighbor node and the abnormal node, the incoming edge with the largest edge weight is selected as the warning trigger edge.

8. The method for detecting and warning of monopolistic violations based on spatiotemporal spectrum according to claim 1, characterized in that, The optimal chain of evidence is obtained through the following steps: The interaction time associated with the aforementioned warning trigger edge is used as the reference time; Using the warning trigger edge as the initial path, search for candidate incoming edges from the starting point of the initial path along the decreasing direction of the reference time, and search for candidate outgoing edges from the ending point of the initial path along the increasing direction of the reference time, until the preset termination condition is met. The candidate incoming edges, the initial path, and the candidate outgoing edges are combined according to the direction and temporal constraints of the edges to obtain multiple candidate paths. The sum of the weights of all edges in the candidate paths is then calculated to obtain the total weight of the path. The candidate path with the largest total path weight is selected as the optimal evidence chain.

9. The method for detecting and warning of monopolistic violations based on spatiotemporal spectrum according to claim 1, characterized in that, The legal provisions involved in locating monopolistic violations at various abnormal nodes from the legal knowledge base through multi-level screening include: The abnormal feature vectors are mapped to the legal semantic space and matched with the vectors of various legal provisions in the legal knowledge base to recall several candidate legal provisions. Explicit factual text is extracted from the optimal evidence chain. Semantic similarity is calculated between the explicit factual text and the text at each level of each candidate legal provision. The weighted total score is obtained and the candidate legal provisions are reordered. Several target legal provisions are selected from the candidate legal provisions. The complete original texts of several target legal provisions and the optimal chain of evidence are input into a large language model that has been fine-tuned by law. The large language model performs semantic verification on the clause-level and / or item-level content under each target legal provision and outputs the applicable clause-level and / or item-level legal provisions.

10. A system for detecting and warning of monopolistic violations based on a spatiotemporal spectrum, characterized in that, include: The data processing module is used to extract market entities and their interaction behavior data at various times from market supervision data. The market entities are used as nodes and the interaction behavior at each time moment is used as an edge to construct a time-series multidirected graph. The edge weight is calculated based on the interaction behavior type and interaction time associated with the edge. An anomaly identification module is used to input the temporal multidirected graph of the time to be analyzed into a trained graph neural network, generate the embedding vector of each node, identify the abnormal nodes with monopolistic and illegal behaviors, and generate an early warning signal for them. The early warning signal includes: an early warning trigger edge and an abnormal feature vector. The evidence extraction module is used to perform forward tracing and backward expansion using the warning trigger edge as a seed in the time-series multidirected graph, and to obtain the optimal evidence chain based on the edge weight. The legal provision location module is used to locate the legal provisions involved in the monopolistic violations of each abnormal node from the legal knowledge base through multi-level filtering based on the abnormal feature vectors of each abnormal node and the optimal evidence chain.