Big data-based intellectual property infringement clue intelligent mining method and system

Abstract

This invention discloses an intelligent method for mining intellectual property infringement clues based on big data, belonging to the field of big data analysis technology. The method includes: in a federated learning environment, each data participant constructs a local temporal heterogeneous graph based on local big data, and uses a time-aware graph neural network to generate dynamic embedded representations of nodes, which are then encrypted and uploaded to a federated coordination node; the federated coordination node clusters the aggregated global dynamic embedded representations based on a dynamic deep probabilistic clustering model with temporal smoothness constraints to identify potential infringement communities evolving over time; and a visualized evidence chain is generated to display the potential infringement communities, evolution paths, and causal strength. This method enables cross-platform collaborative mining of dynamically evolving infringement communities while protecting data privacy, and significantly improves the accuracy, interpretability, and interactive experience of infringement clue mining through interactive causal tracing and multi-dimensional visualized evidence chains.

Description

Technical Field

[0001] This invention relates to a method and system for intelligent mining of intellectual property infringement clues based on big data, belonging to the field of big data analysis technology. Background Technology

[0002] With the rapid development of internet technology, intellectual property infringement has become increasingly cross-platform, highly concealed, and rapidly evolving. Existing infringement clue mining technologies typically rely on static knowledge graph construction and association analysis, identifying potential infringers through manually set rules or simple clustering algorithms.

[0003] However, this static analysis model struggles to capture the dynamic evolution of infringement over time, leading to delayed warnings of new and covert infringement strategies. Furthermore, due to commercial competition and data privacy regulations, related infringement data is often scattered across different platforms, forming data silos. Existing centralized analysis methods struggle to achieve cross-platform collaborative discovery of infringement communities while ensuring data privacy. At the visualization level, existing tools primarily serve as terminals for displaying results, providing only static graph browsing capabilities and lacking the ability to deeply explain the causal relationships behind the findings. Users find it difficult to trace the root causes of infringement through interactive operations. Simultaneously, facing complex networks composed of massive numbers of nodes, existing rendering engines often experience lag or even crashes when processing data on the scale of hundreds of millions, failing to meet the demands of real-time interactive exploration.

[0004] Therefore, there is an urgent need for an intelligent mining solution that can dynamically track the evolution path of infringement in a federated learning environment, support interactive causal tracing, and has high-performance rendering capabilities. Summary of the Invention

[0005] To address the problems of existing technologies, such as the inability to capture the dynamic evolution of infringement behavior, lack of causal explanatory power, difficulty in achieving cross-platform collaboration while protecting privacy, and poor performance of large-scale data interaction, the present invention aims to provide a method and system for intelligent mining of intellectual property infringement clues based on big data, realizing a technological leap from static association display to dynamic causal interactive tracing.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides an intelligent mining method for intellectual property infringement clues based on big data, comprising: in a federated learning environment, each data participant constructs a local temporal heterogeneous graph containing multiple types of nodes and temporal relationships based on local big data; each data participant processes the local temporal heterogeneous graph using a time-aware graph neural network to generate dynamic embedding representations of nodes containing temporal context, and uploads the encrypted dynamic embedding representations to a federated coordination node; at the federated coordination node, clustering is performed on the aggregated global dynamic embedding representations based on a dynamic deep probabilistic clustering model to identify potential infringement communities that evolve over time; wherein, a temporal smoothness constraint is introduced into the clustering objective of the dynamic deep probabilistic clustering model to ensure that nodes within the same community have a coherent behavioral evolution pattern; in response to user interaction with specific graph elements in the visualized graph, a causal inference model is triggered to calculate the causal strength behind the association represented by the specific graph element; a visualized evidence chain is generated to display the potential infringement community, the evolution path, and the causal strength associated with the specific graph element.

[0007] The aforementioned scheme addresses the limitation of existing technologies in collaboratively mining dynamically evolving communities across platforms while protecting privacy by limiting the learning environment to a federated learning setting. This is achieved by having each participant locally generate encrypted dynamic embeddings using a time-aware graph neural network, and then having a coordinating node perform global clustering based on a deep probabilistic clustering model with temporal constraints. Furthermore, by triggering causal inferences through interactive responses and presenting the results as a visual chain of evidence, the scheme enhances the interpretability of the analysis.

[0008] In one implementation, the dynamic deep probabilistic clustering model runs on the federated coordination node, receives the homomorphically encrypted dynamic embedding representations uploaded by each data participant, and computes global cluster centers under an aggregation protocol with differential privacy protection, wherein the differential privacy protection has a privacy budget. Set to 1.0, with a maximum of 300 iterations.

[0009] The above scheme provides a clear balance between privacy protection and computational efficiency in federated clustering through specific homomorphic encryption, differential privacy budgeting, and iteration round parameters.

[0010] In one implementation, the dynamic embedding representation generated by the time-aware graph neural network is a 256-dimensional vector. The dynamic embedding representation is obtained by mapping nodes of five semantic levels—data, information, knowledge, wisdom, and intent—to a unified vector space using the DIKWP2Vec algorithm.

[0011] The above scheme provides specific parameter guidance and semantic basis for the generation of dynamic embeddings by limiting the dimensions to 256 and introducing the DIKWP2Vec algorithm.

[0012] In one implementation, the specific graph element is an edge in the graph representing a behavioral derivative relationship, and the interactive operation is a click operation; the triggering causal inference model is based on backdoor adjustment or dual machine learning algorithms to estimate the causal effect strength of the behavioral derivative relationship represented by the edge driven by legal provisions or market environment confounding factors.

[0013] The above solution specifies the triggering mechanism for causal tracing by clearly defining the interaction object as an edge and the interaction method as a click, thus solving the problem of vague entry point for causal analysis in traditional solutions.

[0014] In one implementation, generating a visual chain of evidence includes: visually overlaying the causal effect strength as a probability value onto the edge, and simultaneously displaying the contribution of multi-dimensional features to the causal strength in a parallel coordinate view or heatmap; wherein the contribution of the multi-dimensional features is generated by the SHAP algorithm or a layer-by-layer correlation propagation algorithm.

[0015] The above scheme transforms abstract causal inferences into intuitive and verifiable chains of evidence through specific visual coding forms and interpretable AI algorithms (SHAP / LRP).

[0016] In one implementation, the method further includes a multi-view linkage mechanism: defining a unified interactive event protocol so that filtering, highlighting, or zooming operations performed on any of the views in the 3D atlas sand table main view, parallel coordinate view, spatiotemporal heat map, community evolution chord diagram, and causal tracing panel trigger synchronous updates of all other views in real time via an event bus.

[0017] The above solution addresses the limitation of single-view analysis dimensions by defining a multi-view linkage protocol, enabling a multi-angle and three-dimensional exploration of infringement clues.

[0018] In one implementation, the method further includes: providing a clue incubator view for receiving hypothetical operations by users manually adding, deleting, or modifying graph elements on the visualization graph; performing real-time deduction of the impact of the hypothetical operations on the overall infringement network risk diffusion based on a lightweight causal reasoning model, and visually presenting the deduction results in a highlighted or contrasted mode.

[0019] The above solution addresses the problem that existing tools cannot support deductive decision-making by introducing a hypothetical analysis environment, thereby enhancing the system's forward-looking analytical capabilities.

[0020] In one implementation, the rendering of the visualization map includes: real-time monitoring of client performance metrics, dynamically switching between WebGL rendering, Canvas rendering, and hybrid modes based on the performance metrics; and predicting the future viewport range based on user interaction trajectories, and preloading map data within the predicted range using off-screen rendering technology.

[0021] The above solution addresses the rendering performance bottleneck at the scale of hundreds of millions of nodes through adaptive rendering and viewport prediction mechanisms, ensuring smooth interaction.

[0022] Secondly, this invention provides an intelligent mining system for intellectual property infringement clues based on big data, comprising: a graph construction module, used in a federated learning environment, for each data participant to construct a local temporal heterogeneous graph containing multiple types of nodes and temporal relationships based on local big data; a temporal embedding module, used by each data participant to process the local temporal heterogeneous graph using a time-aware graph neural network, generating dynamic embedding representations of nodes containing temporal context, and uploading the encrypted dynamic embedding representations to a federated coordination node; and a federated dynamic clustering module, running on the federated coordination node, used to perform clustering based on a dynamic deep probabilistic clustering model on the aggregated nodes. The system uses a global dynamic embedding representation for clustering to identify potential infringing communities that evolve over time. A temporal smoothness constraint is introduced into the clustering objective of the dynamic deep probabilistic clustering model to ensure that nodes within the same community exhibit coherent behavioral evolution patterns. An interactive response and causal tracing module is used to respond to user interactions with specific graph elements in the visualization graph, triggering a causal inference model to calculate the causal strength behind the association represented by the specific graph element. A visualization rendering and evidence chain generation module is used to generate a visual evidence chain, displaying the potential infringing community, its evolutionary path, and the causal strength associated with the specific graph element.

[0023] The system described above achieves all the technical effects described in the method claims through the collaborative work of its various modules, and in particular, it achieves privacy protection and cross-platform collaborative mining in a federated learning environment.

[0024] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any of the preceding claims.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows: By constructing a temporal graph in a federated learning environment and generating dynamic embeddings using a time-aware graph neural network, we can accurately capture the dynamic characteristics of the evolution of infringement over time, while also solving the privacy protection problem of cross-platform data collaboration. By responding to interactive operations to trigger the causal inference model, and generating a visual evidence chain in the form of overlaying SHAP heatmaps and multi-view linkage, the leap from correlation analysis to causal tracing has been realized, making the mining results more explanatory and persuasive. By introducing adaptive hybrid rendering and viewport prediction technologies, the interaction latency issue at the scale of hundreds of millions of nodes was resolved, ensuring a smooth user experience. Furthermore, the introduction of the clue incubator view provides decision-makers with hypothesis-based reasoning capabilities, enhancing the system's forward-looking analytical value. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0027] Figure 1 This is a flowchart of the intelligent mining method for intellectual property infringement clues based on big data according to Embodiment 1 of the present invention; Figure 2 This is a system block diagram of the intelligent mining system for intellectual property infringement clues based on big data, according to Embodiment 4 of the present invention. Figure 3 This is a flowchart of the intelligent rendering optimization scheme of Embodiment 3 of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0029] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0030] Example 1

[0031] This embodiment provides a method for intelligent mining of intellectual property infringement clues based on big data. This method fully realizes a closed-loop process from cross-platform encrypted data collection, temporal graph neural network modeling, federated deep clustering, interactive causal tracing to multi-view visual evidence chain generation. For example... Figure 1 As shown, the method specifically includes the following steps: S100. In a federated learning environment, each data participant constructs a local temporal heterogeneous graph containing multiple types of nodes and temporal relationships based on local big data.

[0032] In this step, multiple data participants, such as e-commerce platform A, social media platform B, and patent database C, build the graph on their respective local servers. The raw data does not leave the local machine, thus meeting privacy protection requirements from the source.

[0033] Specifically, each participant's local data cleaning unit first acquires multi-source heterogeneous data from its local database, including product listing and transaction records, user reviews and sharing logs, publicly available patent claims and litigation documents, and performs noise reduction, deduplication, and format standardization on this raw data. The entity alignment unit uses a pre-trained semantic matching model, such as a BERT-based entity linking model, to associate the same entity from different sources. In one embodiment, it aligns the platform's store name with the assignee in the patent transfer record. Then, the relation extraction unit, based on rule templates and a pre-trained language model, extracts the relationships between subjects, objects, and behaviors from unstructured text such as product descriptions and litigation documents, and assigns a precise timestamp to each relationship. Through the above processing, each participant obtains a temporal heterogeneous graph locally, denoted as... ,in, Identifiers for participating parties. Nodes in the graph. It encompasses various types, including but not limited to infringing parties, infringing objects, behavioral nodes, and legal nodes. Explicitly carrying time attributes .

[0034] S200: Each data participant uses a time-aware graph neural network to process the local temporal heterogeneous graph, generating a dynamic embedded representation of the node containing the time context, and then uploads the encrypted dynamic embedded representation to the federated coordination node.

[0035] The core of this step is to use a time-series graph neural network model on each participant's local machine to transform the graph containing multiple types of nodes and temporal relationships constructed in step S100 into a low-dimensional, dense dynamic vector representation, and then encrypt it before uploading.

[0036] Specifically, each participating party uses its local temporal heterogeneous map. To process an object, the steps include: The system performs type-specific feature transformation. For multi-type nodes in heterogeneous graphs, since different types of nodes have different feature spaces and semantic meanings, directly performing attention calculations on the original feature space will lead to cross-type semantic misalignment. To solve this problem, this embodiment performs type-specific feature transformation for each node type. Define type-specific feature transformation matrices For type neighboring nodes Its eigenvectors First, it undergoes type conversion and is mapped to a unified type. In dimensional embedding space: in, For nodes The original feature vector, For node type The corresponding learnable transformation matrix. In a preferred embodiment, the embedding dimension. It is set to 256. Through this transformation, the features of different types of nodes are mapped to the same vector space, making cross-type attention computation semantically comparable.

[0037] The system calculates the time code. This embodiment uses a time coding function based on Fourier transform. The time difference between interactions between nodes Mapped to dimensional time-encoded vector: in, and Let be learnable parameters, representing the th , respectively Each frequency component and offset. In a preferred embodiment, the time encoding dimension... It is set to 100. This encoding method can capture time patterns at different scales, with low-frequency components corresponding to long-term trends and high-frequency components corresponding to short-term fluctuations.

[0038] System execution time attention calculation. For the target node. At any moment Its neighboring nodes Time attention score between Taking into account the neighbor type transformation characteristics, the time encoding of interactions, and the characteristics of the target node itself, the calculation formula is as follows: in, This is a learnable attention weight vector. This represents the vector concatenation operation. Based on the attention scores mentioned above, the final temporal attention weights are obtained through softmax normalization. : in, Represents a node The set of neighboring nodes.

[0039] The system generates a dynamically embedded representation. Target node. At any moment Dynamic embedding representation This is generated by aggregating weighted information from all neighbors. This embodiment preferably employs a multi-head attention mechanism, setting... Each of the three attention heads uses different parameters, ultimately resulting in... The outputs of each size are concatenated: in, Indicates splicing, For the first The weight of each attention head, For the first Transformation matrix of the head, The activation function is non-linear, preferably LeakyReLU. In a preferred embodiment, the number of attention heads... It is set to 4. Due to time encoding As the interaction time difference changes, the embedding vector of the same node will change at different times, thus achieving a dynamic embedding effect.

[0040] After generating the dynamic embedding representation, each participant needs to upload the data to the federated coordination node for global analysis. To protect data privacy, each participant uses a homomorphic encryption algorithm to process the locally generated dynamic embedding vectors. Encryption is performed to generate ciphertext. Homomorphic encryption allows direct mathematical operations on the ciphertext, and the decrypted result is identical to the result of performing the same operation on the plaintext. To further enhance privacy protection, differential privacy noise can be applied to the embedding vector before encryption. In a preferred embodiment, the differential privacy budget is set to... After encryption, each participant uploads the ciphertext to the federated coordination node. At this point, the coordination node receives the dynamically embedded ciphertext, which has undergone homomorphic encryption and differential privacy processing, with an embedding dimension of 256.

[0041] S300. At the federal coordination node, cluster the aggregated global dynamic embedding representation based on the dynamic deep probabilistic clustering model to identify potential infringing communities that evolve over time.

[0042] This step is performed on the federated coordination node, which can complete the calculation without decryption, and is the key to breaking down data silos and enabling cross-platform collaborative mining.

[0043] Specifically, the coordinating node first utilizes the additive homomorphic property of homomorphic encryption to perform a global weighted average of the embedding vectors uploaded by all participants in the ciphertext state, obtaining the aggregated global dynamic embedding representation ciphertext. This process does not require decryption, and the coordinating node cannot know the specific embedding value of any node.

[0044] Next, the coordinating node runs an improved dynamic deep probabilistic clustering model. In this embodiment, the dynamic deep probabilistic clustering model is DeepDPM, which clusters the aggregated global dynamic embedding representation. The core innovation of this model lies in the design of its loss function, which consists of two parts: in, Clustering quality loss, such as Evidence Lower Bound (ELBO), is used to ensure that the features of nodes within the same community are close to each other in the vector space. The temporal smoothness constraint is the key component of this invention. It is achieved by calculating the KL divergence of the cluster label distributions of the same node at adjacent time steps: in, Represents a node At any moment Belonging to the community The probability, To adjust the hyperparameters of the constraint strength. Minimize It will penalize the cluster labels for drastic fluctuations in the time series, forcing the model to tend to group nodes with consistent behavioral evolution patterns into the same community.

[0045] The model optimizes the loss function described above, without requiring a pre-set number of clusters. In this scenario, the optimal number of communities and the community affiliation of each node at different time points are automatically inferred. In a preferred embodiment, the maximum number of iterations of the model is set to 300. Finally, the coordinating node outputs a set of potential infringing communities that evolve over time and the evolutionary trajectory of each community, reflecting infringing groups that are cross-platform and exhibit collaborative evolutionary behavior.

[0046] S400: In response to user interaction with specific elements in the visualized graph, the causal inference model is triggered to calculate the causal strength.

[0047] This step establishes an on-demand response mechanism for front-end interaction and back-end computation, delaying computationally intensive causal inference tasks until they are triggered at the moment when the user is truly interested.

[0048] Specifically, the system's event listener continuously monitors user interactions. In a preferred embodiment, the specific graph element is an edge in the graph representing a behavioral derivation relationship, and the interaction operation is a click operation. When a user clicks an edge connecting a product circumvention behavior node and an original patent infringement behavior node, the system immediately captures the edge click event, encapsulates the event signal into a request, and sends it to the backend causal inference engine.

[0049] The causal inference engine first loads the local causal graph structure containing the edge. Then, the system preferably employs a backdoor adjustment algorithm. This algorithm identifies all backdoor paths from the causal variable to the outcome variable based on the causal graph—that is, a non-causal path connecting cause and effect, representing the influence of confounding dependent variables. The algorithm mathematically blocks these backdoor paths through calculation and weighted averaging, thereby eliminating the interference of confounding variables such as market popularity and seasonality, and estimating the strength of the pure causal effect driven by confounding factors such as legal provisions or market environment in the behavioral derivative relationship represented by the edge. Alternatively, the system can employ a dual machine learning algorithm, which uses machine learning models to fit the outcome variable and the treatment variable respectively, and estimates the causal effect based on the residuals, particularly suitable for scenarios with high-dimensional confounding factors. Finally, the system outputs a quantified causal effect strength value, which can be represented as a probability value between 0 and 1.

[0050] S500 generates a visual chain of evidence, showcasing potential infringing communities, evolutionary paths, and the causal strength associated with specific graph elements.

[0051] This step transforms all the aforementioned analysis results into an intuitive and interactive visual form, forming a complete evidence loop.

[0052] Specifically, the visualization rendering module first acquires the evolutionary community data output by S300 and the causal strength values ​​output by S400. For the evolutionary community, the system maps different communities to different node colors, and the spatial position of each node is initialized based on the dimensionality reduction projection of its dynamic embedding vector. The community evolution path is presented as a dynamic trajectory line in the main view of the 3D atlas sandbox. Users can observe the merging and splitting of communities by dragging the timeline slider.

[0053] For causal strength, the system visually encodes it based on its numerical value: the closer the strength is to 1, the darker the edge color and the thicker the line. Simultaneously, when the user hovers over or clicks to view details, the system generates an overlay panel. This panel displays a feature contribution heatmap generated by the SHAP algorithm. The SHAP algorithm calculates the contribution of each input feature to the final causal strength prediction. In the heatmap, each feature occupies a row, with the warmth or coolness of the color indicating the positive or negative contribution, and the darkness of the color indicating the absolute value of the contribution. As another optional implementation, the contribution of multi-dimensional features can also be generated by a layer-by-layer correlation propagation algorithm. Furthermore, the system defines a unified interactive event protocol. When a user selects a community in the main view, related views such as the parallel coordinate view, the spatiotemporal heatmap, and the community evolution chord diagram are all updated synchronously in real time via the event bus, thus forming a multi-angle, three-dimensional visual evidence chain.

[0054] Example 2

[0055] This embodiment, based on embodiment 1, further details the implementation process of the two advanced interactive functions: multi-view linkage and hypothesis deduction.

[0056] Specifically, for multi-view interaction, the system defines a unified interaction event protocol, including standard event types such as nodeSelected, edgeClicked, timeRangeChanged, and clusterFiltered. In one embodiment, when a user selects a feature range on the parallel coordinate view, this operation triggers the clusterFiltered event, which, along with its filtering parameters, is published to the global event bus. The rendering module of the 3D map sandbox main view subscribes to this event, and upon receiving it, updates the scene in real time, highlighting only nodes and edges that meet the filtering criteria, while other elements are made semi-transparent. Simultaneously, the spatiotemporal heatmap and community evolution chord diagram are also refreshed, displaying only the spatiotemporal distribution and evolutionary relationships of the filtered subset of data.

[0057] For the clue incubator, the system provides an independent sandbox analysis environment on the client side. In this environment, users can manually add a new cooperative edge or delete a legally restricted node based on the current graph. These operations, as hypothetical inputs, are fed into a lightweight causal inference model deployed on the client side. This model can be a simplified student model distilled from the original global model, or a pre-trained graph Markov network, capable of simulating the impact of the change on the local network risk diffusion within seconds of latency. The system visualizes the change in risk value and the scope of impact caused by the hypothetical operation, either by highlighting the path or by comparing it to the original graph.

[0058] Example 3

[0059] like Figure 3As shown, this embodiment provides an intelligent rendering optimization solution to address the rendering performance bottleneck at a scale of hundreds of millions of nodes.

[0060] Specifically, the system deploys performance monitoring probes within the visualization rendering engine. When the system detects that the client's current frame rate is consistently below a preset threshold, such as 30fps, and the GPU memory usage exceeds 80%, it determines that rendering degradation is necessary. At this point, the rendering controller automatically switches the current WebGL full rendering mode to Canvas 2D degraded rendering mode, or enables a hybrid rendering mode, i.e., maintaining WebGL rendering for core interactive areas and using Canvas 2D rendering for background areas. Simultaneously, a level-of-detail strategy is implemented, simplifying nodes farther from the viewpoint into monochrome dots and disabling shadows and particle effects.

[0061] Furthermore, the system's viewport prediction module continuously analyzes the user's interaction trajectory. Employing a kinematic model-based trajectory prediction algorithm, it calculates the predicted viewport range and position within the next 150 milliseconds based on the current mouse speed and acceleration. The data preloading module then requests or reads the graph node and edge data within this predicted range from the server or from the local cache in a background asynchronous thread and pre-renders it on the off-screen Canvas. When the user's viewport moves to the predicted area, the main rendering thread directly composites the content of the off-screen Canvas, achieving seamless, real-time display and providing users with a smooth, zero-wait interactive experience when dragging.

[0062] Example 4

[0063] This embodiment provides an intelligent mining system for intellectual property infringement clues based on big data. This system is used to implement the intelligent mining process described in the above-mentioned method embodiments. Figure 2 As shown, the system uses a modular architecture design to decouple and coordinate data acquisition, algorithm processing, interactive response, and visualization rendering, ensuring processing efficiency and smooth interaction in large-scale data scenarios.

[0064] Specifically, the system includes a graph construction module, a temporal embedding module, a federated dynamic clustering module, an interactive response and causal tracing module, and a visualization rendering and evidence chain generation module.

[0065] The graph construction module is used in a federated learning environment to build a local time-series heterogeneous graph containing multiple types of nodes and temporal relationships based on local big data by various data participants. This module integrates a data cleaning unit, an entity alignment unit, and a relation extraction unit. The data cleaning unit removes noise and redundancy from the raw data; the entity alignment unit identifies the same entity across different data sources; and the relation extraction unit, based on natural language processing techniques, extracts the relationships between subjects, objects, and behaviors from unstructured text and adds timestamps. Finally, this module outputs structured time-series graph data for downstream modules to use.

[0066] The temporal embedding module is used by various data participants to process the local temporal heterogeneous graph using a time-aware graph neural network, generating dynamic embedding representations of nodes that include temporal context, and uploading the encrypted dynamic embedding representations to the federated coordination node. This module encapsulates a time encoding unit, a type transformation unit, and a multi-head attention computation unit. The time encoding unit maps interaction time differences into high-dimensional time-encoded vectors; the type transformation unit maps different types of nodes in the heterogeneous graph to a unified embedding space; and the multi-head attention computation unit integrates neighbor features, temporal information, and target node features to generate dynamic embedding vectors containing temporal context information. After generating the embedding vectors, this module calls the encryption unit to encrypt the embedding vectors using a homomorphic encryption algorithm and a differential privacy mechanism before uploading them to the federated coordination node.

[0067] The federated dynamic clustering module, running on the federated coordination node, is used to cluster the aggregated global dynamic embedding representation based on a dynamic deep probabilistic clustering model to identify potential infringing communities that evolve over time. This module includes a secure aggregation unit and a deep clustering unit. The secure aggregation unit utilizes the additive homomorphic property of homomorphic encryption to perform a globally weighted average of the embedding vectors of each participant in the ciphertext state. The deep clustering unit runs an improved DeepDPM algorithm, which introduces a temporal smoothness constraint term into its clustering objective function. By minimizing the KL divergence of cluster label distributions at adjacent times, it ensures that nodes within the same community have a coherent behavioral evolution pattern. The model automatically outputs the optimal number of communities and the evolutionary trajectory of each community without requiring a preset cluster number K.

[0068] The interaction response and causal tracing module is used to respond to user interactions with specific graph elements in the visualization graph, triggering the causal inference model to calculate the causal strength behind the relationships represented by those specific graph elements. This module internally maintains an event listener and a causal inference engine. The event listener monitors user actions on the front-end visualization interface in real time. When it detects a user clicking on an edge in the graph representing a behavioral relationship, it immediately parses the interaction event and generates a causal inference task request. Upon receiving the request, the causal inference engine estimates the causal effect strength based on a backdoor adjustment algorithm or a dual machine learning algorithm and returns the calculation result.

[0069] The visualization rendering and evidence chain generation module is used to generate a visual evidence chain, showcasing the potential infringing community, evolutionary path, and causal strength associated with the specific graph element. This module integrates a 3D rendering engine, an adaptive degradation controller, a viewport predictor, and a multi-view interactive evidence chain assembler. The 3D rendering engine is responsible for rendering the main body of the graph sandbox; the adaptive degradation controller monitors the client frame rate and video memory usage in real time, dynamically switching between WebGL and Canvas rendering modes; the viewport predictor predicts the future viewport range based on user interaction trajectories and drives preloading; the evidence chain assembler assembles the heatmaps of evolutionary communities, causal strength, and SHAP feature contribution according to a unified interaction protocol and outputs them to each associated view.

[0070] Through the collaborative work of the aforementioned graph construction module, temporal embedding module, federated dynamic clustering module, interactive response and causal tracing module, and visualization rendering and evidence chain generation module, this embodiment constructs a complete closed-loop system from data input and intelligent analysis to interactive presentation. The modules communicate through standardized data interfaces, ensuring efficient data flow and decoupling functional logic. This results in a system with good scalability and maintainability, capable of adapting to the needs of intellectual property infringement clue mining in different scales and application scenarios.

[0071] Example 5: Computer-readable storage medium This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the intelligent mining method for intellectual property infringement clues based on big data as described in any of the above embodiments.

[0072] Specifically, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this embodiment, the computer-readable storage medium may include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any suitable combination of the foregoing. More specific examples of computer-readable storage media include: portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, static random access memory, portable compact disk read-only memory, digital multifunction disks, memory sticks, floppy disks, mechanical encoding devices, and any suitable combination of the foregoing.

[0073] In this embodiment, the computer program stored on a computer-readable storage medium includes program code for performing the steps in the method embodiment of the present invention. When the program is executed by the processor, the processor will perform the following operations sequentially according to the logical instructions of the program code: First, it reads and parses the graph construction instructions stored in the medium, controlling the data interface to obtain local time-series heterogeneous graph data containing multiple types of nodes and temporal relationships from the local multi-source heterogeneous database by each data participant in the federated learning environment; Second, it reads the dynamic embedding generation instructions, calls the time-aware graph neural network model to perform feature extraction and vectorization processing on the time-series graph data, generates dynamic embedding representations of nodes containing temporal context, and uploads the encrypted dynamic embedding representations to the federated coordination node; Next, it reads the federated clustering identification instructions, and at the federated coordination node, it performs clustering on the aggregated global dynamic embedding representations based on the dynamic deep probabilistic clustering model to identify potential infringement communities that evolve over time; Subsequently, it reads the interaction response instructions, monitors user input interaction events, and triggers the causal inference model to calculate causal strength when user interaction with specific graph elements in the visualized graph is detected; Finally, it reads the visualization rendering instructions, controlling the display device to generate and display a visualized evidence chain containing evolving communities and causal strength.

[0074] The computer-readable storage medium provided in this embodiment stores a computer program that implements the intelligent mining method of the present invention, so that the technical solution of the present invention can exist and be distributed in the form of a software product, which facilitates the dissemination and deployment of the technology. At the same time, it also provides flexible carrier support for implementing the present invention through various carriers such as cloud servers and edge computing nodes.

[0075] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligently mining intellectual property infringement clues based on big data, characterized in that, include: In a federated learning environment, each data participant constructs a local temporal heterogeneous graph containing multiple types of nodes and temporal relationships based on local big data. Each data participant uses a time-aware graph neural network to process the local temporal heterogeneous graph, generating a dynamic embedded representation of the node containing the temporal context, and then uploads the encrypted dynamic embedded representation to the federated coordination node. At the federal coordination node, the aggregated global dynamic embedding representation is clustered based on a dynamic deep probabilistic clustering model to identify potential infringing communities that evolve over time. Temporal smoothness constraints are introduced into the clustering objective of the dynamic deep probabilistic clustering model to ensure that nodes within the same community have a coherent behavioral evolution pattern. In response to user interaction with specific graph elements in the visualization graph, the causal inference model is triggered to calculate the causal strength behind the association represented by the specific graph element; Generate a visual chain of evidence to demonstrate the potential infringing community, its evolutionary path, and the causal strength associated with the specific graph element.

2. The method for intelligent mining of intellectual property infringement clues based on big data according to claim 1, characterized in that, The dynamic deep probabilistic clustering model runs on the federated coordination node, receives the homomorphically encrypted dynamic embedding representation uploaded by each data participant, and calculates the global cluster centers under an aggregation protocol with differential privacy protection. The privacy budget for the differential privacy is... Set to 1.0, with a maximum of 300 iterations.

3. The method for intelligent mining of intellectual property infringement clues based on big data according to claim 1, characterized in that, The dynamic embedding representation generated by the time-aware graph neural network is a 256-dimensional vector. The dynamic embedding representation is obtained by mapping nodes of five semantic levels—data, information, knowledge, wisdom, and intent—to a unified vector space through the DIKWP2Vec algorithm.

4. The method for intelligent mining of intellectual property infringement clues based on big data according to claim 1, characterized in that, The specific graph element is an edge in the graph representing a behavioral derivative relationship, and the interactive operation is a click operation; the triggering causal inference model is based on backdoor adjustment or dual machine learning algorithms to estimate the causal effect strength of the behavioral derivative relationship represented by the edge driven by legal provisions or market environment confounding factors.

5. The method for intelligent mining of intellectual property infringement clues based on big data according to claim 4, characterized in that, The generation of the visual evidence chain includes: visualizing the causal effect strength as a probability value and overlaying it on the edge, and simultaneously displaying the contribution of multi-dimensional features to the causal strength in a parallel coordinate view or heatmap; wherein the contribution of the multi-dimensional features is generated by the SHAP algorithm or the layer-by-layer correlation propagation algorithm.

6. The method for intelligent mining of intellectual property infringement clues based on big data according to claim 1, characterized in that, The method also includes a multi-view linkage mechanism: defining a unified interactive event protocol so that filtering, highlighting or zooming operations performed on any of the views in the 3D map sandbox main view, parallel coordinate view, spatiotemporal heat map, community evolution chord diagram and causal tracing panel can trigger synchronous updates of all other views in real time through the event bus.

7. The method for intelligent mining of intellectual property infringement clues based on big data according to claim 1, characterized in that, The method further includes: providing a clue incubator view for receiving hypothetical operations by users manually adding, deleting, or modifying graph elements on the visualization graph; and performing real-time inference of the impact of the hypothetical operations on the overall risk spread of the infringement network based on a lightweight causal reasoning model, and visually presenting the inference results in a highlighted or contrasted mode.

8. The method for intelligent mining of intellectual property infringement clues based on big data according to claim 1, characterized in that, The rendering and generation of the visualization map includes: real-time monitoring of client performance metrics, dynamically switching between WebGL rendering, Canvas rendering, and hybrid modes based on the performance metrics; and predicting the future viewport range based on user interaction trajectories, and preloading map data within the predicted range using off-screen rendering technology.

9. A big data-based intelligent system for mining intellectual property infringement clues, characterized in that, include: The graph construction module is used in a federated learning environment to enable each data participant to construct a local temporal heterogeneous graph containing multiple types of nodes and temporal relationships based on local big data. The temporal embedding module is used by each data participant to process the local temporal heterogeneous graph using a time-aware graph neural network, generate a dynamic embedding representation of the node containing the temporal context, and upload the encrypted dynamic embedding representation to the federated coordination node. The federated dynamic clustering module, running on the federated coordination node, is used to cluster the aggregated global dynamic embedding representation based on the dynamic deep probabilistic clustering model to identify potential infringing communities that evolve over time. The dynamic deep probabilistic clustering model introduces a temporal smoothness constraint in its clustering objective to ensure that nodes within the same community have a coherent behavioral evolution pattern. The interactive response and causal tracing module is used to respond to user interaction with specific graph elements in the visualization graph, triggering the causal inference model to calculate the causal strength behind the association represented by the specific graph element; The visualization rendering and evidence chain generation module is used to generate a visual evidence chain, which displays the potential infringing community, the evolution path, and the causal strength associated with the specific graph element.

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

Images