A method and system for intelligent analysis of graph data based on enhanced analysis

Abstract

This application relates to an intelligent graph data analysis method and system based on analytics-enhanced generation, comprising: extracting relevance from raw data and constructing a graph based on a user question to obtain graph data for solving the current problem; decomposing the overall analysis objective into multiple sub-tasks and constructing a task-level directed acyclic graph; performing candidate algorithm retrieval, algorithm matching, and parameter configuration for each sub-task in the task-level directed acyclic graph; establishing a call connection between a large language model and a graph computing engine based on MCP, adapting and scheduling the graph data and parameter configuration; filtering and extracting intermediate results generated by the graph algorithm execution; extracting key content from the graph computing results and generating an analysis report in conjunction with the user question. This application achieves end-to-end processing from natural language questions and raw data to analysis report generation by efficiently integrating the generative capabilities of a large language model with deterministic graph analysis algorithms, thereby improving the automation capability of graph data analysis tasks.

Description

Technical Field

[0001] This application relates to the field of graph data processing technology, and in particular to an intelligent analysis method and system for graph data generated based on analysis enhancement. Background Technology

[0002] Graph data, as a data structure capable of effectively representing complex relationships between entities, has been widely applied in fields such as social network analysis, financial risk control (e.g., anti-money laundering detection), knowledge graphs, and recommender systems. Graph data analysis tasks typically include, but are not limited to: node importance assessment, community detection, path finding (e.g., shortest path, negative cycle detection), link prediction, and subgraph matching. Compared to text and image data, graph data has highly heterogeneous topological structures and attribute characteristics, and its analysis often requires the combination of multiple specialized algorithms, demanding extremely high levels of domain knowledge and implementation experience.

[0003] Traditional graph data analysis workflows heavily rely on manual intervention, encompassing multiple stages such as problem definition, algorithm selection, data preprocessing, code implementation, parameter tuning, and result interpretation. As with typical workflows in existing technologies, when faced with complex problems like "detecting potential money laundering paths and estimating the amount involved," analysts need to manually perform multiple iterations. This results in low automation, long processing times, and high technical barriers, significantly limiting the large-scale application of graph technology.

[0004] In recent years, with the rapid development of LLMs (Large Language Models), researchers have attempted to introduce them into the field of graph data analysis, and currently, the following five main technical solutions have been developed:

[0005] (1) RAG (Retrieval-Augmented Generation)

[0006] This type of technology uses a large language model as its core, providing context to the model by retrieving pre-built external knowledge bases (including graph algorithm descriptions, code snippets, or knowledge graph triples) to respond to users' natural language questions. Existing RAG solutions have achieved some success in handling simple knowledge-based question answering or single-hop relation queries. However, in essence, RAG only enhances the large language model's ability to "retrieve" and "memorize" static knowledge, without improving the model's ability to "analyze" and "computate" graph topologies.

[0007] (2) Natural Language to Graph Query Language Conversion Technology

[0008] This type of technology uses a large language model to parse the user's natural language questions into graph query languages ​​(such as Cypher, Gremlin, GQL) or SQL statements, which are then executed by the graph database engine.

[0009] (3) Dedicated graph analysis interactive system

[0010] Some systems provide interactive graph analysis interfaces by deeply integrating large language models with specific graph databases (such as Neo4j and Memgraph). Representative systems include ChatGraph and GraphChat.

[0011] (4) Large Graph Language Model

[0012] This approach attempts to deeply integrate GNNs (Graph Neural Networks) with large language models, directly performing end-to-end modeling and inference on graph structures.

[0013] (5) Data proxy system

[0014] This type of system uses a large language model as a central controller and automates the processing from raw data to analysis results by calling external tools or semantic operation sequences. It has mature applications in the field of tabular data analysis, and some work attempts to extend it to graph data scenarios.

[0015] The above-mentioned technology currently faces the following technical problems:

[0016] 1. Insufficient reliability of algorithm selection and execution process construction: Existing solutions mainly rely on the internal knowledge of large language models or static knowledge bases for algorithm matching. However, graph analysis tasks involve a wide variety of algorithms (including traversal, centrality, community detection, graph neural networks, etc.) and their domain-specific variants, resulting in low matching accuracy and easy selection of suboptimal algorithms.

[0017] 2. There are interaction barriers between graph algorithms and large language models: the input and output formats of different graph algorithms are heterogeneous (e.g., PageRank outputs a node-score dictionary, Louvain outputs a community list), and there is a lack of a unified calling protocol, which makes it difficult for large language models to generate accurate instructions; the data scale of the analysis results often exceeds the length of the context window, and existing solutions lose topological information by sampling or simplification; the results contain multiple modalities such as text, numerical values, and graph structures, and there is a lack of effective fusion and parsing mechanisms.

[0018] 3. The interaction and adaptation between multiple algorithms still faces several challenges: First, it is necessary to identify the data dependencies between algorithms and judge and transform different types of data dependencies; second, in the process of multi-algorithm collaboration, there is a lack of adaptive mechanisms to dynamically adjust parameters based on intermediate results, and the current parameter tuning still mainly relies on manual methods; in addition, the algorithm output is usually complex structured data, but there is still a lack of unified standardized methods for summarizing and interpreting.

[0019] The root cause of the aforementioned technical problems lies in the fact that existing technologies have not established a deep semantic interaction protocol between large language models and deterministic graph algorithms, have not formed a hierarchical and scalable algorithm knowledge support, and have not designed a data management architecture for the entire matching graph analysis process, thus making it impossible to achieve intelligent end-to-end processing of complex graph analysis tasks. Summary of the Invention

[0020] The technical solution to the above-mentioned technical problems in this application is as follows:

[0021] Firstly, a method for intelligent analysis of graph data based on analytics enhancement is provided, including:

[0022] Based on the user's question, the original data is filtered for relevance, and the large language model is used to perform ad hoc cleaning, extraction and graph construction on the filtered original data to obtain graph data for solving the current problem.

[0023] Semantic understanding of user questions is performed, the overall analysis goal is broken down into multiple sub-tasks, and a task-level directed acyclic graph is constructed based on the logical and data dependencies between the sub-tasks.

[0024] Candidate algorithm retrieval, algorithm matching, and parameter configuration are performed on each subtask in the task-level directed acyclic graph.

[0025] A standardized call connection is established between the large language model engine and the graph computing engine based on MCP. According to the input and output specifications of the matching algorithm, the graph data, parameter configuration and cross-algorithm intermediate results are adapted and scheduled for execution at runtime. At the same time, the key result extraction code is generated based on the large language model to automatically filter and extract the intermediate results generated by the graph algorithm execution.

[0026] The system aggregates key results and evidence fragments from each subtask, extracts key indicators from the graph calculation results and converts them into natural language descriptions, and generates an analysis report by combining the user's question context.

[0027] Optionally, the step of breaking down the overall analysis objective into multiple sub-tasks includes:

[0028] Clearly define the execution order of each subtask;

[0029] Define the output of the preceding subtask as a preset condition for the input of the subsequent subtask.

[0030] Optionally, the step of performing candidate algorithm retrieval, algorithm matching, and parameter configuration for each subtask in the task-level directed acyclic graph includes:

[0031] Based on a hierarchical domain knowledge base that supports automatic expansion, a hybrid retrieval and enhanced generation technique is used to perform multi-level retrieval on each sub-task to obtain background knowledge and candidate algorithms for each sub-task.

[0032] Under the constraints of a directed acyclic graph structure, the goals and requirements of each subtask are semantically refined to identify the computational patterns of each subtask.

[0033] Based on the retrieved domain knowledge, a graph matching algorithm is used for each subtask;

[0034] Based on the problem semantics of the subtask and the parameter semantics in the graph algorithm knowledge base, the parameters of the graph algorithm matched to the subtask are configured.

[0035] Optionally, before performing multi-level retrieval on each subtask, the method further includes:

[0036] A hierarchical domain knowledge base that supports automatic expansion is constructed. The domain knowledge base is constructed using the following mapping relationship: task type corresponds to graph algorithm, and graph algorithm corresponds to multi-dimensional knowledge of the algorithm.

[0037] Secondly, a graph data intelligent analysis system based on an analysis-enhanced generation method is proposed, including:

[0038] The problem transformation module is used to filter the relevance of the raw data based on the user's problem, and to use a large language model to perform ad hoc cleaning, extraction and graph construction on the filtered raw data to obtain graph data for solving the current problem.

[0039] The solution orchestration module is used to perform semantic understanding of user problems, break down the overall analysis goal into multiple sub-tasks, and construct a task-level directed acyclic graph based on the logical and data dependencies between the sub-tasks.

[0040] The algorithm matching module is used to perform candidate algorithm retrieval, algorithm matching, and parameter configuration for each subtask in the task-level directed acyclic graph.

[0041] The data adaptation module is used to establish a standardized call connection between the large language model engine and the graph computing engine based on MCP, and to perform runtime adaptation and scheduling of graph data, parameter configurations and cross-algorithm intermediate results according to the input and output specifications of the matching algorithm; at the same time, it generates key result extraction code based on the large language model to automatically filter and extract intermediate results generated by the execution of graph algorithms.

[0042] The report conversion module is used to aggregate the key results and evidence fragments of each subtask, extract key indicators from the graph calculation results and convert them into natural language descriptions, and generate an analysis report in combination with the user's question context.

[0043] Thirdly, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the intelligent analysis method for graph data based on analysis enhancement as described in the first aspect.

[0044] Fourthly, a computer storage medium is provided, on which a computer program is stored, characterized in that, when the program is executed by a processor, it implements the intelligent analysis method for graph data generated based on analysis enhancement as described in the first aspect.

[0045] Fifthly, a computer program product, comprising a computer program or instructions, characterized in that, when executed by a processor, the computer program or instructions implement the steps in the intelligent analysis method for graph data based on analysis enhancement as described in the first aspect.

[0046] Beneficial effects

[0047] 1. The method described in this application can perform end-to-end complex graph data analysis tasks;

[0048] 2. Compared with database-oriented graph analysis techniques such as Chat2Graph, the method in this application can effectively call graph data analysis algorithms for data analysis;

[0049] 3. Compared with Chat2Graph, the method in this application speeds up the analysis time by more than an order of magnitude on the same complex graph data analysis task;

[0050] 4. Compared with retrieval enhancement generation schemes represented by RAGFlow (retrieval enhancement generation workflow), the method of this application supports complex graph analysis tasks and achieves the completeness and accuracy of analysis results.

[0051] The advantages of this application in its additional aspects will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart illustrating the intelligent analysis method for graph data generated based on analysis enhancement, as shown in an embodiment of this application.

[0054] Figure 2This is a schematic diagram illustrating the principle of the intelligent analysis method for graph data based on analysis enhancement, as shown in the embodiments of this application.

[0055] Figure 3 This is an overview diagram of the AAG (Analytics Augmented Generation) architecture shown in the embodiments of this application;

[0056] Figure 4 This is a schematic diagram of the algorithm decision-making mechanism based on a hierarchical graph algorithm knowledge base, as shown in an embodiment of this application.

[0057] Figure 5 This is an online extension of the dependency-driven knowledge base shown in the embodiments of this application. Detailed Implementation

[0058] The technical solutions of the embodiments of this application 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 application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0059] The technical problem this application aims to solve is to address the shortcomings of existing technologies by proposing an intelligent analysis method and system for graph data based on augmented graph generation (AAG). The core technology is AAG, which uses a large language model as the intelligent coordination hub and graph analysis algorithms as the core computing engine to autonomously analyze and process complex and deeply interconnected data from the real world. By efficiently integrating the generative capabilities of the large language model with the deterministic graph analysis algorithm, it achieves end-to-end automated processing from natural language problems and raw data to the generation of final analysis results. The system based on this method achieves end-to-end automated processing from natural language problems to structured graph analysis reports through the collaborative work of five core functional layers (user interaction layer, decision layer, knowledge layer, data interaction layer, and execution interaction layer).

[0060] Example 1

[0061] This application discloses an intelligent analysis method for graph data based on analysis enhancement generation, such as... Figure 1 As shown, it includes the following steps:

[0062] Step S1: Based on the user's question, the original data is filtered for relevance, and the large language model is used to perform ad-hoc cleaning, extraction and graph construction on the filtered original data to obtain graph data for solving the current problem;

[0063] Specifically, through the human-computer interaction page, the system receives users' natural language questions as well as massive amounts of raw or log data; for the unstructured or semi-structured raw data uploaded by users, the system uses the capabilities of a large language model to perform ad-hoc cleaning and extraction.

[0064] The system can automatically identify entities and relationships in the data and transform the raw data into graph structure data, realizing the automated transformation from "raw graph data" to "intelligent graph data management" without the need for tedious manual data cleaning.

[0065] The specific implementation process includes: first, analyzing the user's question, locating / filtering out data fragments related to the question, and then processing these data fragments. The specific process is as follows:

[0066] 1) For relevant data fragments, extract map features according to data type, including:

[0067] For text / knowledge-based data: use a large language model to extract entities and relations (such as triples), and generate nodes and edges;

[0068] For structured data such as transaction logs: identify key objects (such as accounts) as nodes, create edges / event nodes for transaction behaviors, and write the remaining fields as attributes.

[0069] 2) Output a local subgraph (nodes, edges, and attributes) that serves the solution of the current problem, for subsequent graph analysis / reasoning.

[0070] This embodiment first performs relevance trimming based on the problem, then extemporaneously extracts relevant fragments and builds a graph, directly serving the subsequent graph analysis process. This differs from the common technical method of "first extracting all data offline, pre-building a global knowledge graph, and then retrieving / querying".

[0071] Step S2: Perform semantic understanding on the user's question, break down the overall analysis goal into multiple sub-tasks, and construct a task-level directed acyclic graph based on the logical and data dependencies between the sub-tasks to form a structured and executable analysis process;

[0072] In practice, the first step is to perform semantic understanding on the complex natural language questions raised by users, breaking down the overall analysis objective into several sub-tasks with clearly defined functional boundaries, thus providing a foundation for the subsequent analysis process. The main operations performed are as follows:

[0073] Step S2.1: The system organizes the split subtasks into an analysis framework in the form of DAG (Directed Acyclic Graph) to characterize the logical dependencies and data flow relationships in the analysis process;

[0074] Specifically, the execution order of each subtask is clearly defined (e.g., community detection is performed first, followed by loop detection within the community); the output of the preceding subtask is automatically defined as the input of the following subtask (e.g., the identification result of the abnormal community is passed to the path search task).

[0075] Step S2.2: The analysis process is described in a structured manner using DAG, so that the overall analysis scheme is arranged at an algorithm-independent level, providing a unified execution skeleton for subsequent algorithm selection and parameter configuration;

[0076] Structured description is the process of converting natural language analysis requirements into a task-level DAG, while orchestration is the process of fixing the order, dependencies, and input-output relationships of each subtask to form an executable skeleton.

[0077] The structured description mainly includes three layers:

[0078] 1) Subtask nodeization breaks down the overall goal into multiple bounded subtasks, with each node corresponding to a subtask.

[0079] 2) Explicit Dependency Relationships

[0080] Logical dependencies: which comes first and which comes later;

[0081] Data dependency: How the output of the previous task is used as the input of the next task.

[0082] 3) Form an algorithm-independent execution skeleton

[0083] Before deciding on a specific algorithm, the process skeleton is fixed as a DAG to provide a unified framework for subsequent algorithm matching and parameter configuration.

[0084] Step S3: Perform candidate algorithm retrieval, algorithm matching, and parameter configuration for each subtask in the task-level directed acyclic graph;

[0085] Step S3.1: Based on a hierarchical domain knowledge base that supports automatic expansion, a hybrid retrieval and enhanced generation technique is used to perform multi-level retrieval on each sub-task to obtain the background knowledge and candidate algorithms for each sub-task;

[0086] The hybrid retrieval enhancement generation technology used in this embodiment is specifically designed as follows:

[0087] The first layer of hybrid search: keyword search + literature citation relationship search;

[0088] When knowledge is insufficient, the system obtains candidate algorithm information by combining keyword retrieval and literature citation relationships (reference / cited links);

[0089] The second layer of hybrid retrieval: hierarchical multi-level retrieval;

[0090] In the existing hierarchical knowledge base: first, perform a coarse screening at the "task type / algorithm family" layer; then drill down to the specific algorithm and variant layer.

[0091] The hybrid retrieval enhancement generation technology is employed, and the specific retrieval process is as follows:

[0092] Step S3.1.1: Determine the subtask objectives based on the subtasks in the DAG;

[0093] Step S3.1.2: Retrieve matching algorithm categories at the task type level;

[0094] Step S3.1.3: Retrieve candidate algorithms and variants at the algorithm layer;

[0095] Step S3.1.4: Read the fine-grained knowledge of the candidate algorithm, such as its applicable conditions, parameter constraints, and input / output structure;

[0096] Step S3.1.5: If the existing knowledge is insufficient, trigger the online extended search of "keywords + literature citation relationship".

[0097] Step S3.2: Under the constraints of the directed acyclic graph structure, semantically refine the goals and requirements of each subtask and identify the computational patterns of each subtask;

[0098] Step S3.3: Based on the retrieved domain knowledge (i.e., the background knowledge and candidate algorithms of the above sub-tasks), match the graph algorithm for each sub-task.

[0099] In practice, the main purpose of this step is to combine problem understanding with graph algorithm matching based on a hierarchical knowledge base; this is achieved through the following project:

[0100] 1) Hierarchical knowledge base construction and retrieval:

[0101] The system pre-builds a hierarchical domain knowledge base that supports automatic expansion. This knowledge base is constructed with "task type - graph algorithm - multi-dimensional knowledge of the algorithm" as the core mapping relationship. Once the analysis process framework is determined, the system uses hybrid RAG (Retrieval-Augmented Generation) technology to perform multi-level searches on each sub-task to obtain relevant background knowledge and candidate algorithms.

[0102] in, Figure 5This diagram illustrates the automatic expansion mechanism of the graph algorithm knowledge base proposed in this embodiment. When the system performs algorithm matching, if the existing knowledge base lacks sufficient information to support the current task, the system will trigger the automatic knowledge base expansion mechanism. The main workflow of this mechanism is as follows: First, search keywords are generated based on the current task requirements; then, relevant algorithm information is obtained through multi-source information retrieval, including academic literature databases, technical blogs and algorithm documents, and open-source code repositories; further, the system simultaneously uses keyword retrieval and literature citation relationship retrieval to obtain candidate algorithms; after obtaining candidate algorithm information, the system filters them, only including algorithm information relevant to the current task in the knowledge base, while maintaining the original knowledge structure; finally, the newly added algorithm knowledge is structured and stored so that it can be retrieved and utilized in subsequent algorithm matching processes. Through this mechanism, the system can continuously expand the graph algorithm knowledge base.

[0103] The construction method for an automatically expanding domain knowledge base includes: establishing a hierarchical knowledge structure based on "task type - graph algorithm - multi-dimensional knowledge of algorithms"; modeling each graph algorithm as a knowledge node and establishing relationships between algorithms in terms of differences, applicable scenarios, and mechanisms; associating each algorithm node with multi-dimensional knowledge such as applicable conditions, parameter constraints, and input / output structures; triggering an online expansion module when existing knowledge is insufficient to support the current subtask, and obtaining candidate algorithm information through keyword retrieval and literature citation relationship retrieval; selectively incorporating candidate algorithm information related to the current task into the hierarchical knowledge structure while maintaining the original hierarchical organization.

[0104] 2) Task semantic refinement and requirements assessment:

[0105] Under the constraints of the DAG structure, the system semantically refines the analysis objectives and computational requirements of each subtask and identifies its core computational patterns (such as traversal, community detection, centrality measurement, etc.).

[0106] Supported by task DAG constraints and a hierarchical graph algorithm knowledge base, subtasks are refined into graph analysis and computation modes, which serve to match subsequent algorithms.

[0107] 3) Graph Algorithm Matching and Parameter Configuration:

[0108] Based on the retrieved domain knowledge, the system automatically matches the most suitable graph algorithm for each subtask (e.g., matching the PageRank algorithm for the influence assessment task), and completes the initial setting of key hyperparameters in combination with task characteristics;

[0109] Given the analysis objectives, computational patterns, input data characteristics, dependencies between preceding and following nodes, and parameter constraints of the current subtask, the candidate algorithm that meets the applicable conditions and has the highest overall matching degree is selected. The matching process includes: first, coarse-grained screening at the task type layer, then drilling down to the specific algorithm and variant layer, and determining the target graph algorithm by combining the algorithm's applicable conditions, key hyperparameter constraints, and input-output structure;

[0110] Parameter registration is mainly based on the problem semantics of the subtask and the parameter semantics in the graph algorithm knowledge base, and the parameters of the graph algorithm matched to the subtask are configured.

[0111] Step S4: Establish a standardized call connection between the large language model engine and the graph computing engine based on MCP (Model Context Protocol), and perform runtime adaptation and scheduling execution of graph data, parameter configuration and cross-algorithm intermediate results according to the input and output specifications of the matching algorithm; at the same time, generate key result extraction code based on the large language model to automatically filter and extract the intermediate results generated by the graph algorithm execution.

[0112] This step involves the collaboration of the execution interaction layer and the data interaction layer, primarily based on the execution interaction and dynamic adaptation of the MCP protocol to ensure the efficient execution and data flow of the heterogeneous graph algorithm;

[0113] The main tasks are as follows:

[0114] 1) Interactive execution based on MCP:

[0115] MCP is used to establish a connection standard between the large language model analysis engine and the underlying graph computing engine (such as Graph Analysis Engine, Graph Learning Engine) and external tools (Pandas, NumPy, etc.).

[0116] Specifically, in this embodiment, the "connection standard" established in this embodiment is essentially a set of context protocols / structured frameworks, serving as a runtime contract between the large language model engine and the graph computing engine, and used for closed-loop verification of parameter configuration and output understanding.

[0117] This standard consists of at least two parts:

[0118] A: InputSpec: Declares the name, type, constraints / value range, whether it is required, and semantic description for each algorithm parameter;

[0119] B: OutputSpec: Declares the structure type and semantic description of the algorithm's output, guiding large language models to stably parse results.

[0120] Setup steps: Read the algorithm comments and user documentation of the graph engine → Organize / generate corresponding InputSpec and OutputSpec for each algorithm → Register as a unified connection standard.

[0121] After algorithm matching, the graph data / call configuration is adapted and verified according to the connection standard (ensuring that the parameters are valid and the output is interpretable), thereby realizing reliable interaction between the large language model and the graph algorithm.

[0122] This embodiment of the method is geared towards graph analysis tasks, integrating algorithm registration, input / output specification, runtime verification, cross-algorithm adaptation, and result interpretation into a unified execution chain.

[0123] 2) Automatic data adaptation across algorithms:

[0124] To address the issue of data format incompatibility between different algorithms (such as PageRank and GNN classification models), the data interaction layer provides an automatic cross-algorithm data adaptation function to ensure seamless data flow in the DAG process;

[0125] In practice, "connecting standard-adapted graph data" is not a simple format conversion, but rather aligning the current graph data with the algorithm requirements, allowing a particular algorithm to run directly and its results to be sent to the next algorithm. The specific steps are as follows:

[0126] Step S4.1: Read the input specification and input structure of the target algorithm;

[0127] Clearly define what kind of graph representation the algorithm requires, what parameters it needs, and what the input field format should be.

[0128] Step S4.2: Select a suitable graph data storage structure or representation for the algorithm;

[0129] Based on the algorithm's access patterns and computational characteristics, select the most suitable graph data storage structure or representation method, and use it as the input form for that node.

[0130] Step S4.3: Map the current graph data to an input that the algorithm can accept;

[0131] This includes: node / edge / attribute field alignment; parameter population; necessary data trimming, reorganization, or type conversion;

[0132] Step S4.4: Perform parameter adaptation;

[0133] Map the constraints in the user problem to algorithm parameters to generate executable parameter configuration;

[0134] Step S4.5: Perform cross-algorithm data adaptation;

[0135] When the output structure of the upstream algorithm and the input structure of the downstream algorithm are inconsistent, a conversion is performed to complete the connection between nodes;

[0136] Step S4.6: Generate key result extraction code:

[0137] To address the specific requirements of each subtask, the code for extracting key results is automatically generated using large language model technology, and the intermediate data results generated by the graph algorithm are automatically filtered, verified, and extracted.

[0138] The main difference between this embodiment and the prior art is that, under the constraints of task DAG and unified structure, it is oriented towards multi-algorithm collaborative scenarios for graph analysis, and performs a combined process of graph representation selection, parameter mapping, cross-algorithm data adaptation, and key result filtering.

[0139] 3) Error handling and process optimization:

[0140] If any abnormalities are encountered during the above execution process, the system will automatically correct the execution plan or adjust the parameters based on the feedback using a reinforcement learning mechanism, thereby achieving closed-loop process optimization.

[0141] Step S5: Gather the key results and evidence fragments from each subtask, extract the key indicators from the graph calculation results and convert them into natural language descriptions, and generate an analysis report by combining the user's question context.

[0142] The specific implementation process includes: extracting key indicators from the adapted graph data, and using graph-to-text generation technology to convert the key indicators into natural language descriptions; and using a large language model to generate an analysis report by combining the context of the problem data with the natural language descriptions.

[0143] The main tasks are as follows:

[0144] 1) Convert key results data in the graph to text (Data2Text):

[0145] The system automatically extracts key indicators (such as the amount involved, key nodes, and path length) from complex graph calculation results and uses the text module to convert them into natural language descriptions.

[0146] 2) Report generation integrating data and inference:

[0147] The large language model combines the context of the original problem with the structured analysis results to generate a complete analysis report that includes data support, risk assessment, and disposal recommendations.

[0148] To better understand the principles of this embodiment ( Figure 2 As shown in the figure, an example of applying the method of this embodiment to process a complex graph analysis task is given below:

[0149] (I) Implementation Method for Constructing a Hierarchical Graph Algorithm Knowledge Base

[0150] The system constructs a hierarchical graph algorithm knowledge base for graph analysis task planning and orchestration, providing support at the algorithm semantic and computational levels. This enables the model to perform queryable, structured reasoning on algorithm selection, applicability conditions, and trade-offs, rather than simply serving as static document storage. It includes the following:

[0151] 1. Knowledge Base Structure and Representation

[0152] The knowledge base is organized using a hierarchical knowledge graph, forming a top-down "coarse-to-fine" retrieval structure:

[0153] 1) The upper layer is the task type layer.

[0154] This is used to cluster graph analysis algorithms according to analysis objectives and usage scenarios, so as to tailor the search space during the task planning stage;

[0155] 2) The middle layer is the algorithm layer.

[0156] Each algorithm is modeled as a knowledge graph node; edges are used to encode the key differences and relationships between algorithms, including but not limited to differences in analysis objectives, differences in use cases, and differences in key mechanisms (e.g., differences between similar ranking methods in terms of personalization, topic sensitivity, etc.).

[0157] 3) The underlying layer is a multi-dimensional knowledge layer of algorithms.

[0158] Each algorithm node is associated with detailed algorithmic knowledge information to support subsequent parameter adaptation, participation in optimization, and result interpretation. This detailed information is only accessed as needed after candidate algorithms have been selected at the upper level, thus achieving a separation mechanism of "upper-level navigation and lower-level detailed examination," thereby avoiding the exposure of unnecessary low-level details in the early planning stage and improving matching efficiency.

[0159] 2. On-demand dynamic expansion and update mechanism

[0160] Given the rapid evolution of graph analysis methods, the knowledge base adopts an on-demand expansion strategy: when the existing knowledge is insufficient to support the current task, the system triggers the retrieval expansion module to obtain candidate algorithm information related to the task by jointly utilizing keyword retrieval and literature citation relationships (including citation and cited links); subsequently, only methods related to the current task are selectively included in the knowledge base, and its hierarchical organizational structure is maintained to achieve continuous adaptation.

[0161] 3. Storage format

[0162] The knowledge base can be implemented using a graph database or an equivalent structured storage, and its node attributes and hierarchical relationships can be persisted in the form of JSON, YAML, or database tables. When a new algorithm, variant, or relationship is added, it can be taken effect by updating the knowledge base content without modifying the task planning and execution orchestration logic.

[0163] (II) Implementation of Graph Algorithm Registration and Standardized Invocation

[0164] 1. Graph Algorithm Registration

[0165] In this example, the system registers all executable graph algorithms uniformly. Each algorithm's registration information includes:

[0166] Algorithm unique identifier; input / output JSON structure;

[0167] 2. MCP Protocol Execution

[0168] The system uses MCP as the unified protocol for algorithm invocation. Each invocation includes a task ID, context digest, error code dictionary, and input / output JSON structure. The scheduler validates the input data based on the structure to ensure that the algorithm engines (graph traversal, centrality, community detection, GNN inference, etc.) can directly consume it.

[0169] (III) Problem Analysis Process Implementation Method

[0170] 1. Natural Language Understanding and Task DAG Construction

[0171] After receiving a user's natural language question, the system uses a large language model for semantic understanding and decomposes the solution steps to construct a task-level Directed Acyclic Graph (DAG). In the DAG: nodes represent subtasks; directed edges represent logical and data dependencies between tasks, used to limit the execution order and input / output transmission relationships.

[0172] 2. Hierarchical graph algorithm knowledge base driven graph matching

[0173] The system performs algorithm retrieval and matching on each node of the DAG based on a hierarchical graph algorithm knowledge base: first, coarse-grained screening is performed at the task type / algorithm family layer to prune the candidate space, and then drill down to the specific algorithm and variant layer to determine the candidate algorithm set, and output the algorithm's applicable conditions, key hyperparameter constraints and expected input-output structure as the basis for subsequent collaborative computing.

[0174] 3. DAG execution scheduling, batch processing and exception control

[0175] The fusion scheduling module sorts the DAG topology and triggers execution according to dependency order; it also schedules nodes without dependencies in parallel to improve throughput. The system supports batch problem processing and can statistically analyze the time consumption, number of node executions, and resource consumption metrics for each process. If invalid parameters, unavailable algorithms, or runtime anomalies occur during execution, the system captures the exceptions and outputs interpretable error messages, and continues processing subsequent nodes or requests after the recovery conditions are met.

[0176] 4. MCP-based collaborative computing and runtime orchestration

[0177] The system utilizes the MCP mechanism to enable interactive computation between the large language model and graph algorithms. For each DAG node, the following steps are executed sequentially:

[0178] Graph data storage structure selection: Based on the access patterns and computational characteristics of candidate algorithms, select the most suitable graph data storage structure or representation and use it as the computational input form for that node.

[0179] Algorithm parameter adaptation: Map the available constraints in the problem to algorithm parameters to generate an executable parameter configuration;

[0180] Cross-algorithm data adaptation: When the input and output structures of adjacent node algorithms are inconsistent and there is a data dependency, data transformation is performed to complete the cross-algorithm connection;

[0181] Key result filtering: Based on the user's concerns and the algorithm's output structure, generate result extraction / filtering logic to retain only key intermediate results and evidence fragments related to the subtask, reducing the interference of irrelevant output on subsequent reasoning.

[0182] 5. Strengthen learning optimization

[0183] In a preferred embodiment, the system introduces reinforcement learning to optimize the execution strategy. The optimization targets include: algorithm and variant selection, key parameter settings, data adaptation strategies between nodes, and DAG structure (such as node merging / splitting and dependency adjustment) to improve overall accuracy and execution efficiency.

[0184] 6. Results aggregation and analysis report generation

[0185] After the process is completed, the system aggregates the key results and evidence chains of each node. The large language model generates a final analysis report based on the aggregated results. The report includes core conclusions, key findings, a summary of the main algorithms and parameters used, as well as data sources and intermediate evidence, which facilitates verification and reproduction.

[0186] Compared with existing graph data analysis techniques based on large language models, this invention achieves end-to-end complex graph data analysis tasks. Furthermore, it has the following advantages:

[0187] Compared to database-oriented graph analysis solutions like Chat2Graph, which essentially perform operator-level matching and are limited by SQL, making it difficult to effectively call graph data analysis algorithms for data analysis;

[0188] For the same complex graph data analysis task, AAG is more than an order of magnitude faster than Chat2Graph in analysis time;

[0189] Compared to retrieval enhancement generation schemes represented by RAGFlow, which mainly focus on vector retrieval and text generation and lack a mechanism for scheduling and executing graph data analysis algorithms as deterministic computing engines, this approach does not support complex graph analysis tasks, resulting in limited completeness and accuracy of analysis results. The method in this application supports complex graph analysis tasks and achieves high completeness and accuracy of analysis results.

[0190] Example 2

[0191] This embodiment provides an intelligent analysis system for graph data generated based on analysis enhancement, including:

[0192] The problem transformation module is used to filter the relevance of the raw data based on the user's problem, and to use a large language model to perform ad hoc cleaning, extraction and graph construction on the filtered raw data to obtain graph data for solving the current problem.

[0193] The solution orchestration module is used to perform semantic understanding of user problems, break down the overall analysis goal into multiple sub-tasks, and construct a task-level directed acyclic graph based on the logical and data dependencies between the sub-tasks.

[0194] The algorithm matching module is used to perform candidate algorithm retrieval, algorithm matching, and parameter configuration for each subtask in the task-level directed acyclic graph.

[0195] The data adaptation module is used to establish a standardized call connection between the large language model engine and the graph computing engine based on MCP, and to perform runtime adaptation and scheduling of graph data, parameter configurations and cross-algorithm intermediate results according to the input and output specifications of the matching algorithm; at the same time, it generates key result extraction code based on the large language model to automatically filter and extract intermediate results generated by the execution of graph algorithms.

[0196] The report conversion module is used to aggregate the key results and evidence fragments of each subtask, extract key indicators from the graph calculation results and convert them into natural language descriptions, and generate an analysis report in combination with the user's question context.

[0197] In one feasible implementation, an intelligent analysis system for graph data based on analytics enhancement generation, such as... Figure 3 As shown, the system architecture mainly includes the following five functional layers:

[0198] (1) User interaction layer

[0199] The user interaction layer receives user input questions and related data, and outputs the final analysis results to the user. Users can describe their analysis needs in natural language, such as influence analysis, community discovery, or path analysis.

[0200] (2) Decision-making level

[0201] The decision-making level is responsible for understanding user problems and planning tasks. Its main functions include: semantic parsing of problems, decomposition of analysis tasks, structured arrangement of subtasks, integration of analysis results and generation of reports.

[0202] This layer enables automatic planning for complex analysis tasks through a large language model.

[0203] (3) Knowledge layer

[0204] The knowledge layer stores algorithmic knowledge and related background information in the field of graph analysis, assisting large language models in algorithm selection. The knowledge layer is typically organized in a hierarchical structure, including task type, algorithm category, and algorithm-related knowledge.

[0205] (4) Data Interaction Layer

[0206] The data interaction layer is responsible for data processing and cross-algorithm data adaptation, mainly including: raw data extraction and graph structure construction, data format conversion between different algorithms, and intermediate result filtering and transmission;

[0207] (5) Execute the interaction layer

[0208] The execution interaction layer is responsible for connecting the large language model and the graph computing engine, implementing algorithm calls and execution control through a unified interface, and returning the calculation results.

[0209] Through the above-mentioned multi-layered collaboration, this embodiment realizes a natural language-driven automatic graph analysis system.

[0210] Design a matching operating mechanism for the automatic graph analysis system, such as Figure 4 This is a schematic diagram of the algorithm decision-making mechanism based on a hierarchical algorithm knowledge base proposed in this embodiment.

[0211] When a user submits an analysis request, such as "assess the influence of Tom and Alice," the system first enhances the user's question through the RAG-Prompt module and then inputs it into the algorithm's decision-making big language model for semantic understanding.

[0212] The algorithmic decision-making big language model classifies analysis tasks based on task semantics and categorizes tasks into predefined graph analysis task types, such as: traversal and search tasks, community detection and clustering tasks, and centrality and importance measurement tasks.

[0213] After the task type is determined, the system selects the corresponding candidate graph algorithm at the algorithm layer. For example:

[0214] Traversal and search tasks: BFS, DFS;

[0215] Community discovery tasks: Louvain, Label Propagation;

[0216] Centrality computation tasks: PageRank, Degree Centrality;

[0217] During the algorithm selection process, the system simultaneously accesses the underlying knowledge layer (evidence and background knowledge), which includes: the original algorithm paper, the algorithm implementation document, the algorithm's applicable scenarios, and the algorithm's performance characteristics.

[0218] With the background information provided by the knowledge layer, the large language model can more accurately determine the applicability of the algorithm and provide the recommendation algorithm and its explanation.

[0219] For example, in an influence assessment task, the system might recommend the PageRank algorithm and explain that it measures node importance through random walks.

[0220] Through the above mechanism, this embodiment realizes the automatic mapping from natural language tasks to graph algorithm selection.

[0221] Example 3

[0222] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0223] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the intelligent analysis method for graph data based on analysis enhancement as described in the various embodiments of this application.

[0224] The aforementioned storage media include: flash memory, hard disks, multimedia cards, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disks, optical discs, servers, APP (Application) application stores, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the intelligent analysis method for graph data based on enhanced analysis.

[0225] Example 4

[0226] This embodiment proposes an electronic device, including: one or more processors, and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the intelligent analysis method for graph data based on analysis enhancement.

[0227] The electronic device may be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the intelligent analysis method for graph data based on analysis enhancement as described in the embodiments. It is understood that the electronic device may also include input / output (I / O) interfaces and communication components.

[0228] The processor is used to execute all or part of the steps in the intelligent analysis method for graph data based on analysis enhancement as described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in an electronic device, as well as application-related data.

[0229] The processor can be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the intelligent analysis method for graph data generated based on analysis enhancement as described in the above embodiments.

[0230] Example 5

[0231] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the intelligent analysis method for graph data generated based on analysis enhancement.

[0232] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0233] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0234] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of equivalent technology of this disclosure, then the intent of this disclosure also includes such modifications and variations.

[0235] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

[0236] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application.

Claims

1. An intelligent analysis method based on analysis of enhanced generated graph data, characterized in that, include: Based on the user's question, the original data is filtered for relevance, and the large language model is used to perform ad hoc cleaning, extraction and graph construction on the filtered original data to obtain graph data for solving the current problem. Semantic understanding of user questions is performed, the overall analysis goal is broken down into multiple sub-tasks, and a task-level directed acyclic graph is constructed based on the logical and data dependencies between the sub-tasks. Candidate algorithm retrieval, algorithm matching, and parameter configuration are performed on each subtask in the task-level directed acyclic graph. A standardized call connection is established between the large language model engine and the graph computing engine based on MCP. According to the input and output specifications of the matching algorithm, the graph data, parameter configuration and cross-algorithm intermediate results are adapted and scheduled for execution at runtime. At the same time, the key result extraction code is generated based on the large language model to automatically filter and extract the intermediate results generated by the graph algorithm execution. The system aggregates key results and evidence fragments from each subtask, extracts key indicators from the graph calculation results and converts them into natural language descriptions, and generates an analysis report by combining the user's question context. 2.The method of claim 1, wherein, The overall analysis objective is broken down into multiple sub-tasks, including: Clearly define the execution order of each subtask; Define the output of the preceding subtask as a preset condition for the input of the subsequent subtask. 3.The method of claim 1, wherein, The process of performing candidate algorithm retrieval, algorithm matching, and parameter configuration for each subtask in the task-level directed acyclic graph includes: Based on a hierarchical domain knowledge base that supports automatic expansion, a hybrid retrieval and enhanced generation technique is used to perform multi-level retrieval on each sub-task to obtain background knowledge and candidate algorithms for each sub-task. Under the constraints of a directed acyclic graph structure, the goals and requirements of each subtask are semantically refined to identify the computational patterns of each subtask. Based on the retrieved domain knowledge, a graph matching algorithm is used for each subtask; Based on the problem semantics of the subtask and the parameter semantics in the graph algorithm knowledge base, the parameters of the graph algorithm matched to the subtask are configured.

4. The intelligent analysis method for graph data based on analysis enhancement generation according to claim 3, characterized in that, Before performing multi-level retrieval on each subtask, the following is also included: A hierarchical domain knowledge base that supports automatic expansion is constructed. The domain knowledge base is constructed using the following mapping relationship: task type corresponds to graph algorithm, and graph algorithm corresponds to multi-dimensional knowledge of the algorithm.

5. A graph data intelligent analysis system based on enhanced analysis generation, characterized in that, include: The problem transformation module is used to filter the relevance of the raw data based on the user's problem, and to use a large language model to perform ad hoc cleaning, extraction and graph construction on the filtered raw data to obtain graph data for solving the current problem. The solution orchestration module is used to perform semantic understanding of user problems, break down the overall analysis goal into multiple sub-tasks, and construct a task-level directed acyclic graph based on the logical and data dependencies between the sub-tasks. The algorithm matching module is used to perform candidate algorithm retrieval, algorithm matching, and parameter configuration for each subtask in the task-level directed acyclic graph. The data adaptation module is used to establish a standardized call connection between the large language model engine and the graph computing engine based on MCP, and to perform runtime adaptation and scheduling of graph data, parameter configurations and cross-algorithm intermediate results according to the input and output specifications of the matching algorithm; at the same time, it generates key result extraction code based on the large language model to automatically filter and extract intermediate results generated by the execution of graph algorithms. The report conversion module is used to aggregate the key results and evidence fragments of each subtask, extract key indicators from the graph calculation results and convert them into natural language descriptions, and generate an analysis report in combination with the user's question context.

6. A computer storage medium having a computer program stored thereon, characterized in that, When executed by a processor, this program implements the intelligent analysis method for graph data based on analysis enhancement as described in any one of claims 1 to 4.

7. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the intelligent analysis method for graph data based on analysis enhancement as described in any one of claims 1 to 4.

8. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by the processor, they implement the steps in the intelligent analysis method for graph data based on analysis enhancement as described in any one of claims 1 to 4.

Images