An interpretable knowledge mining system, method, device, and medium

Abstract

This invention provides an interpretable knowledge mining system, method, device, and medium, belonging to the field of interpretable artificial intelligence and intelligent data mining technology. The system includes a data processing module, an algorithm and intelligent interpretation module, a visualization analysis module, and an operation management module. Through multi-branch shallow symbolic feature learning tree structure, dynamic branch expansion strategy, LLM-guided initialization, and a phased gray wolf optimization selection strategy, it generates a concise and easily understandable explicit mathematical model. Furthermore, it introduces LLM and retrieval-enhanced generation (RAG) mechanisms to automatically perform semantic parsing and intelligent interpretation of the generated model. This invention, employing the aforementioned interpretable knowledge mining system, method, device, and medium, effectively lowers the technical threshold for interpretable modeling while improving the model's predictive performance, structural interpretability, and knowledge discovery efficiency. It is suitable for engineering optimization, scientific modeling, and cross-domain data analysis.

Description

Technical Field

[0001] This invention relates to the field of interpretable artificial intelligence and intelligent data mining technology, and in particular to an interpretable knowledge mining system, method, device and medium. Background Technology

[0002] In the fields of data mining and machine learning, while traditional complex models (such as deep neural networks) possess powerful predictive capabilities, their black-box nature makes the model's decision-making process difficult to understand. This limitation is particularly pronounced in many applications requiring knowledge discovery, causal analysis, or mechanism interpretation, such as energy engineering, ecological science, medicine, and other research fields with high requirements for model transparency and interpretability.

[0003] Symbolic regression methods based on genetic programming (GP) are widely used due to their inherent interpretability, as they can automatically generate expressions with clear structure and explicit mathematical form. However, existing genetic programming methods still have significant limitations.

[0004] High barrier to entry: Existing tools (such as gplearn, Distributed Evolutionary Algorithms in Python (DEAP)) are mainly geared towards developers, requiring users to have programming skills and in-depth algorithm knowledge, and to manually configure function sets, parameters and genetic operations. The technical barrier is too high for domain experts.

[0005] Lack of intelligent interpretation mechanism: Existing tools usually only output the final mathematical expression and lack the function of automatic analysis and interpretation of model structure, feature importance and physical meaning. Users still need to rely on their own knowledge to understand and verify the results, resulting in low knowledge extraction efficiency.

[0006] Simple Algorithm Mechanism: Mainstream genetic programming methods mostly adopt standard genetic algorithms and classic tree structures, which often exhibit the following problems: Structural inflation and interpretability degradation: To fit more complex data relationships, traditional symbolic regression (GP) often deepens the representation tree, leading to models that evolve into deep, complex, and highly nested structures. The mathematical structure of these "deep tree models" is difficult for humans to understand and review, weakening the interpretability of symbolic regression methods.

[0007] Search strategies are static and inefficient: Traditional GP mostly uses fixed genetic operation probabilities and selection mechanisms, lacking the ability to dynamically adjust search behavior according to evolutionary stage or population diversity, resulting in an imbalance between exploration and development, which not only easily gets trapped in local optima, but also reduces convergence speed and computational efficiency.

[0008] Lack of prior domain knowledge injection: The initialization and evolution process of traditional GP methods relies almost entirely on random search, failing to effectively utilize domain knowledge or external knowledge bases.

[0009] Therefore, existing technologies have not yet effectively solved the balance between model interpretability and user usability, especially lacking an integrated solution that can automatically combine domain knowledge to intelligently interpret generated models. This greatly limits the promotion and application of interpretable artificial intelligence technology in interdisciplinary research and engineering practice. Summary of the Invention

[0010] The purpose of this invention is to provide an interpretable knowledge mining system, method, device, and medium to solve the problems of poor model interpretability, high usage threshold, and lack of automated intelligent interpretation capabilities in the prior art, and to achieve high-performance, highly interpretable data mining and knowledge discovery with zero programming operations.

[0011] To achieve the above objectives, the present invention provides an interpretable knowledge mining system, comprising: The data processing module is used to load, preprocess, and partition datasets; The algorithm and intelligent interpretation module is used to receive the preprocessed and partitioned structured dataset output by the data processing module, execute the benchmark algorithm to build the control model, and generate an interpretable model with explicit mathematical structure based on the LLM prior knowledge guidance mechanism and the genetic programming hybrid algorithm. At the same time, the LLM and retrieval enhancement generation RAG mechanism is introduced to provide intelligent interpretation of the generated mathematical model. The visualization analysis module is used to receive the data processing results from the data processing module and the model training and interpretation results from the algorithm and intelligent interpretation module, and generate comprehensive visualization results for display. The operation management module connects to all other modules and is used to record the operation history, management result files, and system logs of each module.

[0012] This invention also provides a method for mining interpretable knowledge, comprising the following steps: The raw data is loaded through the data processing module, and data preprocessing and partitioning are performed. By running a hybrid algorithm of benchmark algorithm and genetic programming through an algorithm and intelligent interpretation module, an interpretable mathematical model is generated and intelligently interpreted. The visualization analysis module displays the results of data processing, model training, and interpretation. The operation management module records the entire process operation log and history to ensure that the experiment is reproducible.

[0013] The present invention also provides a computer device, including: a memory and a processor; the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described interpretable knowledge mining method.

[0014] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described interpretable knowledge mining method.

[0015] Therefore, the present invention employs the above-described interpretable knowledge mining system, method, device, and medium, and the beneficial technical effects are as follows: (1) Significantly reduces the barrier to entry for interpretable modeling and improves efficiency. This invention integrates the entire process of data loading, preprocessing, model training, interpretation, and visualization by providing a unified graphical user interface, achieving end-to-end zero-programming operation. This effectively solves the technical barrier problem in the background technology where existing tools (such as gplearn and DEAP) are mainly aimed at developers and require manual configuration and programming skills, enabling domain experts to easily perform high-performance interpretable knowledge mining without delving into algorithm details.

[0016] (2) This invention achieves synergistic optimization of model structure interpretability and predictive performance. Based on the idea of ​​multi-branch shallow symbolic feature learning, and combined with LLM-guided initialization, adaptive dynamic branch expansion, and phased gray wolf optimization selection strategies, this invention can generate a concise and easy-to-examine explicit mathematical model, effectively overcoming the interpretability degradation problem caused by structural expansion in traditional genetic programming methods. At the same time, by introducing a residual compensation modeling mechanism, the high-order nonlinear parts that the main model fails to characterize are fitted, significantly improving the overall prediction accuracy while maintaining the transparency of the main model's mathematical structure, thus resolving the contradiction between the limited performance of traditional interpretable models and the uninterpretability of complex black-box models.

[0017] (3) It provides automated and intelligent model interpretation and knowledge discovery capabilities. This invention innovatively introduces Large Language Model (LLM) and Retrieval Enhanced Generation (RAG) mechanisms, which not only inject domain prior knowledge into the genetic programming process to guide the search, but also automatically perform semantic parsing on the final generated mathematical model, extracting the meaning of formulas, the mechanism of feature action, and optimization suggestions. This fundamentally solves the problem in the background technology that existing methods lack intelligent interpretation mechanisms and users still need to rely on their own knowledge to manually analyze and verify the results, transforming interpretability from abstract mathematical structure output into directly understandable domain knowledge, thereby improving the efficiency and depth of knowledge discovery. Attached Figure Description

[0018] Figure 1 This is an architecture diagram of an interpretable knowledge mining system according to the present invention; Figure 2 A comparison chart of the KS test results for the distribution of key variables in the training and test sets; Figure 3 This is a diagram of the expression model for traditional TGP-6. Detailed Implementation

[0019] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0021] Example 1 like Figure 1 As shown, an interpretable knowledge mining system includes: (1) Data processing module, used for loading, preprocessing and partitioning datasets, including data loading unit, data preprocessing unit and data partitioning unit.

[0022] Data processing module logic: Prerequisite: The user possesses data in a specific field and wants to perform data modeling and analysis.

[0023] The abstract process is as follows: upload data - clean data - split data (training set and test set).

[0024] Example of user operation scenario demonstration: A user has a CSV and Excel file that needs to be analyzed / mined. The user loads the data in the data loading unit, and then views the data in the corresponding tables and visualizations. Then, the user enters the data preprocessing unit to clean the data. When the data is clean, the user enters the data partitioning unit to divide the dataset into training and test sets, preparing to run the algorithm for data mining.

[0025] Technical implementation characteristics: Based on mainstream scientific computing and machine learning libraries such as pandas, numpy, and scikit-learn, this system implements core functions such as data uploading, data cleaning and preprocessing, statistical analysis, and training and test set partitioning.

[0026] Based on the mainstream matplotlib and seaborn visualization libraries, this tool enables multi-dimensional visualization of box plots, histograms, correlation heatmaps, scatter plot matrices, and more.

[0027] The signal-slot mechanism based on PyQt5 enables data communication and state synchronization between modules, and QThread is used to implement asynchronous background execution of computing tasks, ensuring high-concurrency computing capabilities while maintaining a smooth user interface response.

[0028] 1) Data Loading Unit: Supports reading data in various formats such as CSV, Excel, and TXT. It adopts an asynchronous loading mechanism in the background thread to avoid interface freezing. After loading, it automatically performs data type identification and basic statistical analysis, and provides interactive data preview, descriptive statistical display, and multi-dimensional visualization functions (box plot, histogram, correlation heatmap, scatter plot matrix).

[0029] 2) Data Preprocessing Unit: Provides intelligent data preprocessing functions, including: Missing value analysis: Automatically detects the proportion of missing values ​​in each column and generates visual cards to suggest handling strategies (deletion, mean filling, median filling, mode filling, forward / backward filling).

[0030] Outlier detection and processing: Supports multiple detection algorithms such as IQR interquartile range and Z-Score standard score method.

[0031] Data standardization: Provides methods such as Z-Score standardization, Min-Max normalization, and Robust robust standardization.

[0032] Operation history: Each preprocessing operation is automatically recorded, and supports viewing operation summaries, before-and-after comparison visualizations, and rollback functions.

[0033] 3) Data Partitioning Unit: Supports parameterized training and test set partitioning, allowing users to set parameters such as partition ratio, random seed, whether to shuffle, and stratified sampling. Provides real-time KDE kernel density estimation distribution comparison plots and Kolmogorov-Smirnov (KS) test results to ensure consistency between the training and test set distributions. It also supports loading pre-partitioned datasets from files for easy experimental reproduction.

[0034] (2) The algorithm and intelligent interpretation module is used to receive the preprocessed and partitioned structured dataset output by the data processing module. On the one hand, it executes various benchmark algorithms based on the Scikit-learn framework to build a control model. On the other hand, it generates a mathematical model with explicit structure and good interpretability by using the prior knowledge guidance mechanism of Large Language Model (LLM) and combined with the enhanced hybrid algorithm designed by a novel genetic programming structure. At the same time, it introduces the LLM and Retrieval-Augmented Generation (RAG) mechanism to automatically read the external domain knowledge base and perform intelligent interpretation of the generated mathematical model. Algorithm and Intelligent Interpretation Module Logic: Prerequisite: The user has already correctly divided the dataset into training and test sets in the data processing module.

[0035] Abstract logic: Benchmark algorithm evaluation → Enhanced genetic programming hybrid algorithm mining → Performance comparison analysis → Intelligent interpretation of large models.

[0036] Example of user operation scenario demonstration: The user has correctly divided the dataset into training and test sets in the data processing module. Multiple classic machine learning algorithms are trained in the benchmark algorithm engine, and the performance metrics of the training and test sets are obtained as benchmarks and saved with one click. Deep data mining is performed in a genetic programming hybrid algorithm to obtain the performance metrics of the training and test sets, as well as the learned explicit mathematical structure. The performance metrics of the hybrid algorithm and the benchmark algorithm are compared to quantify the improvement. Finally, a large model is invoked to combine the features of the mathematical model with domain interpretation.

[0037] Technical implementation characteristics: A unified model pool is built based on scikit-learn, XGBoost, LightGBM, and PyTorch libraries, supporting a variety of mainstream machine learning algorithms and deep network models.

[0038] The core logic of genetic programming is implemented based on DEAP, providing evolutionary components such as tree individual representation, fitness evaluation, and genetic operator configuration.

[0039] Inter-module communication is implemented based on the PyQt5 signal-slot mechanism, and background asynchronous execution is achieved through QThread to ensure UI responsiveness and computational efficiency.

[0040] The algorithm and intelligent explanation module includes: 1) Benchmark Algorithm Engine: It supports a variety of scikit-learn algorithms (linear regression, ridge regression, Lasso regression, random forest, gradient boosting, support vector machine regression, decision tree, K-nearest neighbors, etc.) as well as other mainstream models (including XGBoost, LightGBM, CatBoost and deep neural networks (DNN), etc.).

[0041] It provides batch execution capabilities for multiple algorithms, supports unified parameter configuration and random seed control, ensuring the reproducibility of the experimental process; automatic output includes R... 2 It includes multiple evaluation metrics such as MSE, RMSE and MAE, which are displayed in a table format and support visual comparative analysis of the performance of multiple models. 2) Enhanced genetic programming hybrid algorithm engine: Symbolic feature learning unit: A feature pool is constructed based on the genetic programming symbolic framework. A hybrid initialization strategy combining LLM and random strategies is introduced to inject structured prior knowledge while ensuring search diversity. Based on the multi-branch shallow symbolic feature learning tree structure, a dynamic branch expansion strategy is adopted to adaptively adjust the expression complexity within the controlled tree height and high-order operator space, thereby generating high-quality symbolic features that are structurally concise, physically reasonable and have good interpretability.

[0042] The master predictive modeling unit (MPMU) takes excellent features from the feature pool constructed by the symbolic feature learning unit as input and builds a model based on the regression model selected by the user. It evaluates candidate features through cross-validation and uses their predictive performance as a fitness criterion to guide the model selection and parameter optimization process, ultimately evolving to generate a structurally stable, robust, and reviewable master predictive expression.

[0043] Residual Compensation Modeling Unit: Building upon the model constructed by the main prediction modeling unit, this unit uses complex ensemble models or pools of various neural network variants to further capture hidden patterns in the data that were not discovered by the main prediction model. By fitting the residuals, the main model is compensated in series, improving the generalization and accuracy of the overall model without reducing the interpretability of the main model.

[0044] The symbolic feature learning unit includes: Multi-branch shallow symbolic feature learning tree structure: By limiting the maximum depth of a single expression tree, multi-dimensional symbolic features are learned while generating multiple shallow expressions in parallel, thereby improving the interpretability and reviewability of the features and avoiding the structural complexity and interpretation difficulties caused by a single deep expression tree in traditional genetic programming.

[0045] Adaptive dynamic branch expansion strategy: Under the parallel growth framework, this paper proposes to adaptively adjust the number of branches and structural complexity of the expression tree according to the fitness changes of the elite individual group. While keeping the tree under control, richer symbolic features are gradually introduced to capture more complex nonlinear patterns in the data.

[0046] LLM-guided initialization strategy: Based on the prior domain knowledge of the large language model, it generates prior symbolic features that conform to the domain semantics and interpretation constraints, which are used to guide the construction of the initial population of genetic programming, thereby mitigating the unstructured search problem caused by traditional completely random initialization.

[0047] A phased grey wolf optimization selection strategy: To enhance the ability of genetic programming (GWO) to regulate the exploration and development process during the "selection" stage and improve evolutionary convergence efficiency, this invention proposes an improved grey wolf optimization strategy embedded in the selection stage of genetic programming (GP). This framework introduces the social hierarchy and cooperative trapping mechanism of grey wolf optimization (GWO) into the symbolic regression search space of genetic programming. Through innovative phased control and nonlinear convergence adjustment strategies, it adaptively guides the search behavior at different evolutionary stages, thereby achieving a more balanced global exploration and local development capability during the symbolic structure search process. The overall algorithm execution flow is as follows: Step 1: In each generation of evolution, the current population is adjusted based on the fitness value of each individual. Sort them and divide them into four levels: ; in, For the collection of all individuals in the present day; Composed of the best individuals in the current population, serving as the core guides of the evolutionary process; and Each group consists of suboptimal and tertiary-optimal individuals, serving to assist in guiding and stabilizing the search direction. This includes the remaining individuals, who are responsible for maintaining population diversity and expanding the search space; Step 2: To explore and develop the adaptive equilibrium evolution process, introduce a population diversity-based index. Dynamic hierarchical control mechanism: First, define population diversity. This is a normalized metric based on fitness statistics, used to measure the convergence of the population and its ability to cover the search space. ; in, This represents the set of fitness scores for all valid individuals in the current population. Indicates the first The fitness value of each individual; The fitness standard deviation; It has extremely poor adaptability; To prevent division by zero of small constants; The population exhibits high convergence, but the search becomes trapped in a local region; while Individuals exhibit a wide distribution and maintain strong exploratory capabilities.

[0048] Therefore, dynamic regulation can be divided into two scenarios: (1) When At that time, it represents insufficient diversity, and improvement is needed. Adjusting the individual layer ratio enhances global search capabilities and expands the expression structure space; this allows for dynamic hierarchical control. Defined as: ; in, To control the adjustment range of dynamic hierarchical regulation, it is set as a fixed hyperparameter of the algorithm. A larger value results in a larger adjustment range; a smaller value results in a smaller adjustment range. The dynamic update of the proportions of each layer is as follows: ; ; ; ; (2) When At this time, the selection weight of individual leaders is increased, guiding the population to converge faster towards regions of high fitness; this is the dynamic hierarchical regulation. Defined as: ; The dynamic updates of the proportions of each layer are as follows: ; ; ; ; in, These represent the base proportions without adaptive perturbation, which are fixed hyperparameters of the algorithm. Normalization is required after the update. ; in, For four different search layers, the scaling parameter ( This is used to measure its proportion in the current generation; For the sum of four different search layers, For traversal index variables; This represents the proportions of the four different search layers after normalization.

[0049] Step 3: To finely control the overall evolutionary pace, a three-stage nonlinear convergence factor is introduced. For guidance Layer selection strategy to adaptively adjust the intensity of exploration and development based on the evolutionary stage: ; in, For the current algebra; The maximum number of generations; This represents the percentage at the end of the exploration phase. This represents the percentage at the end of the balancing phase. Let be the initial convergence factor; The mid-term convergence factor; The final convergence factor; and All are adjustable indices.

[0050] The settings of each hyperparameter in this embodiment are shown in Table 1.

[0051] Table 1 Hyperparameter Settings Table

[0052] 3) Large Language Model Collaborative Unit: Technical implementation characteristics: Based on the requests library, this implementation provides a unified encapsulation and calling mechanism for service interfaces of various large language models, supporting session state management, timeout control, and exception capture and retry mechanisms to improve the stability and reliability of interface calls.

[0053] Asynchronous scheduling and execution of background tasks are achieved using PyQt5's QThread and signal-slot mechanism, ensuring continuous responsiveness of the user interface during computationally intensive task execution. At the same time, a rich UI component library enables interactive operation, shortcut key binding, and status feedback.

[0054] Based on the re and json standard libraries, the system performs structured parsing and format validation on the results returned by large language models, enabling automatic extraction and multi-format compatibility processing of formulas, text descriptions, and structured data.

[0055] Unified adapter unit for multiple providers: Extensive compatibility: Supports mainstream large language model providers such as OpenAI GPT series, Alibaba Tongyi Qianwen, Baidu Wenxin Yiyan, Kimi, Gemini, and DeepSeek, and supports a unified API call interface for mainstream high-level models including GPT-5 and Gemini-3 Pro.

[0056] Configuration caching: Encrypt and store user-configured API keys, model names, base URLs, and other information in the user's directory to avoid reconfiguration each time.

[0057] Intelligent prompt engineering building blocks: Formula Analysis Prompt Template: This template integrates structured contextual information such as the mathematical structure generated by enhanced genetic programming, the feature combinations used, fitness scores, and dataset information to construct professional analysis request prompts. The prompt construction adopts a three-part structure: a role setting section (positioning the large language model as a data analysis expert, genetic programming expert, and expert in various fields), a task description section (clarifying analysis requirements and output format), and a contextual information section (structured feature combinations, feature data, fitness metrics, and other key information), ensuring that the large language model can accurately understand the analysis task and provide professional interpretation.

[0058] Domain knowledge mapping: It associates the general analysis results of large language models with specific domain knowledge. For example, it can automatically identify coal quality parameters and physicochemical properties in the field of coal analysis, and identify statistical indicators and machine learning concepts in the field of data analysis, thereby enhancing the professionalism and relevance of the explanation.

[0059] Multi-level analysis types: Supports different analysis dimensions such as single feature importance analysis, domain interpretation of constructed high-level features, and mathematical meaning analysis, and dynamically adjusts prompt word templates and emphasis points for different analysis types.

[0060] Prompt preview and editing: A collapsible prompt preview panel is provided, allowing users to view and edit prompt content before sending to ensure the analysis direction aligns with expectations. Edited prompts can be applied to the actual sent content and saved as templates for later use.

[0061] LLM candidate feature pool construction unit: It is used to combine prompt word templates with RAG technology to retrieve external domain knowledge bases and literature, generate candidate symbol features with domain prior constraints and their semantic interpretations, and provide them for use by the symbol feature learning unit; Model semantic analysis and parsing unit: It is used to analyze mathematical models by combining RAG technology, automatically extract the meaning of formulas, the mechanism of action of features, the explanation of importance and optimization suggestions, and supports interactive natural language explanation of the system usage process.

[0062] Structured information extraction mechanism: The system requires the large language model to return analysis results in JSON format, explicitly including agreed-upon chapter fields during the generation phase. Subsequently, through a combination of keyword matching and field parsing, information such as formula explanations, feature importance descriptions, mathematical insights, and optimization suggestions are automatically extracted from the returned natural language content and mapped into structured data objects. Finally, these structured results are distributed to corresponding display areas according to semantic categories, achieving decoupling and controllable presentation between model output and front-end display.

[0063] (3) Visualization and analysis module, used to receive the data processing results from the data processing module and the model training and interpretation results from the algorithm and intelligent interpretation module, and generate comprehensive visualization results for display, including a baseline algorithm visualization unit, a genetic programming result visualization unit, a comprehensive comparative analysis unit, and an interactive chart unit: Technical implementation characteristics: A visualization engine built on matplotlib and seaborn, supporting Times New Roman font configuration.

[0064] A visual GP tree built using the NetworkX and pygraphviz libraries.

[0065] Based on the NavigationToolbar integrated with PyQt5, it provides professional chart navigation and zooming functions. It implements a page-turning browsing mechanism. Integrated tight_layout automatic layout optimization ensures compact margins and high-quality output when exporting charts. It supports exporting multiple chart formats (PNG, PDF, etc.) to meet the needs of different application scenarios.

[0066] Baseline algorithm visualization unit: Algorithm performance comparison bar chart: A horizontal comparison of R values ​​of multiple baseline algorithms. 2 Metrics such as MSE, RMSE, and MAE provide an intuitive display of performance rankings, and support statistical displays of mean, standard deviation, best value, and worst value.

[0067] Predicted Scatter Plot: Plot a scatter plot of predicted values ​​vs. actual values, add a 45-degree ideal fit reference line, and evaluate the prediction bias.

[0068] Residual analysis plots: Draw scatter plots of residuals vs. predicted values ​​and histograms of residual distribution to test model assumptions.

[0069] Batch run statistics: Run the random algorithm multiple times, calculate the mean, standard deviation, minimum and maximum values, and evaluate the algorithm's stability.

[0070] Visualization unit for genetic programming results: Evolutionary curve plot: Real-time plotting of the curves showing the changes in optimal fitness and average fitness over generations, monitoring the evolutionary process, and supporting aggregated display of results from multiple batches.

[0071] Expression Tree Structure Diagram: Using NetworkX and pygraphviz, mathematical expressions are converted into a tree-like hierarchical structure. Terminal nodes are displayed in green, and function nodes are displayed in orange. Compact / loose / automatic layout modes are supported.

[0072] Fitting curve plot: Plots the fitting curve between the model's predictions and the actual values, providing confidence intervals.

[0073] Fitness distribution histogram: Displays the distribution of fitness of individuals in the population, and identifies multimodal distributions and outliers.

[0074] Feature Usage Statistics Chart: Displays the usage frequency of original and advanced features, supporting page-turning browsing and horizontal bar chart display.

[0075] Comprehensive comparative analysis unit: The results files of the baseline algorithm and genetic programming are loaded uniformly for horizontal comparison.

[0076] Generate a comprehensive performance dashboard: including a comparison table of basic indicators, radar charts, detailed statistics with pagination, and other multi-dimensional displays.

[0077] Supports multi-chart browsing: Quickly switch between different indicator or feature visualization charts using the Previous / Next button or Jump to dropdown menu.

[0078] Control Panel: Provides on / off control for visualization elements such as grids, legends, titles, axes, and error bars, and supports expanding / collapse to maximize the chart area.

[0079] Chart Export: Supports one-click saving of the current chart as a high-resolution image, automatically using a compact margin layout to reduce blank space, suitable for papers and reports.

[0080] Interactive chart units: Real-time chart interaction: Visual charts support interactive operations such as hovering the mouse to display values, clicking to highlight data points, selecting a local area to zoom in, and double-clicking to reset the view, allowing users to explore data details in depth.

[0081] Before and after processing comparison mode: The charts in the preprocessing module support the overlay display of data before and after processing (distinguished by different colors / transparencies). Users can adjust the display ratio using a slider to dynamically observe the processing effect.

[0082] Dynamic chart type switching: Users can quickly switch between different chart types (box plots, histograms, etc.) on the same dataset without reloading the data, making it easy to analyze distribution characteristics from different perspectives.

[0083] (4) Operation management module, which connects with all other modules and is used to record the operation history, management result files and system logs of each module, including a unified log unit, an operation history unit and a result management unit: Technical implementation characteristics: Implement thread-safe log delivery based on PyQt5's signal and slot mechanism.

[0084] Operation history is stored in JSON format and supports cross-session persistence.

[0085] The log drawer and operation history button, integrated into the main window, provide convenient access.

[0086] It supports the traceability and reversal of operations, ensuring flexibility in data processing.

[0087] Unified logging unit: Multi-level logging system: Supports multiple log levels such as DEBUG, INFO, WARNING, and ERROR, and allows filtering and displaying logs by level in the GUI interface.

[0088] Dual-channel output: Logs are written to a log file and output in real time in the background. The file log includes a continuously appended main log and a session log named by timestamp.

[0089] Standard stream redirection: Redirects Python's stdout and stderr to the logging system, capturing all printed output and exception information.

[0090] Signal bridging: Uses Qt's signal and slot mechanism to pass log messages from background threads to the GUI main thread, enabling real-time log updates.

[0091] Log search and filtering: Provides keyword search and level filtering functions to quickly locate problems.

[0092] Operation history unit: Global operation log: Records every step of the user's operation across modules, including data loading, preprocessing steps, model training parameters, result analysis, etc.

[0093] Structured storage: The operation history is persisted to the user directory in JSON format, recording information such as timestamps, operation types, key parameters, and data change summaries.

[0094] View History: Supports viewing the complete operation history, displayed by module category, with different operation types indicated by color.

[0095] Operation Undo: Supports undoing data preprocessing operations and restoring data to its previous state.

[0096] Results Management Unit: Unified path management: All result files are organized into a unified directory structure according to their type (baseline algorithm results, genetic programming results, etc.), which facilitates management and retrieval.

[0097] Structured result files: Supports storing readable text data and structured data in JSON format, facilitating program parsing.

[0098] Result parser: Automatically scans the result directory, extracts algorithm names, indicator data, best individual expression, training and testing data information, etc., and constructs a structured result object.

[0099] Results file browsing: Supports browsing and selecting results files, loading and generating comparative analysis reports and visualization charts.

[0100] Each module communicates with the data interface through a unified message passing mechanism, realizing an end-to-end zero-programming graphical data mining process.

[0101] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0102] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0103] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0104] Example 2 Ash fusion temperature (AFT) is a key parameter characterizing the softening and melting behavior of coal ash under high-temperature conditions, and is crucial for the design, operation, and optimization of boilers and reactors. Traditional experimental determination methods are time-consuming and costly; therefore, developing accurate and interpretable predictive models based on the chemical composition of coal ash has significant industrial and scientific research value.

[0105] Data preparation and preprocessing.

[0106] The data comes from the Coal Quality Database (version 3.0) of the U.S. Geological Survey (USGS), which integrates multi-source laboratory analysis results and contains 7,659 coal samples and 279 variables.

[0107] Data cleaning and feature engineering.

[0108] General cleaning: Remove quality control markers, non-numerical fields and variables with a missing rate ≥90%, and eliminate samples with physical anomalies or incomplete data, resulting in 110 numerical features of 5474 samples.

[0109] Task-related feature selection: Based on the coal ash melting mechanism, 25 features directly related to AFT were selected, mainly including: The main ash oxides (such as SiO2, Al2O3, Fe2O3, CaO, MgO, etc.).

[0110] The total elemental content of coal (such as Si, Al, Fe, Ca, Mg, etc.).

[0111] Ash content related parameters (such as GSAsh, Standard_Ash, etc.).

[0112] Final dataset: After the above processing, a modeling dataset containing 4466 complete samples is obtained, with no missing values ​​for any features.

[0113] Data partitioning.

[0114] The training and test sets were split using scikit-learn, and the Kolmogorov–Smirnov (KS) test was introduced to verify the consistency of the distributions of key variables, ensuring the robustness and fairness of the model evaluation. The results are as follows: Figure 2 .

[0115] Comparative experimental design.

[0116] To comprehensively evaluate the effectiveness of the present invention, multiple sets of comparative experiments were designed.

[0117] Compare with the baseline model.

[0118] Linear interpretable baselines include Linear Regression, RidgeCV, and ElasticNet.

[0119] Traditional genetic programming (TGP) baseline: Six different hyperparameter configurations (TGP-1 to TGP-6) were set, covering different population sizes, generations, and tree depths to represent the performance range of traditional symbolic regression methods, as shown in Table 2.

[0120] Table 2 Baseline Configuration of Traditional Genetic Programming (TGP)

[0121] Mainstream black-box models include: Random Forest, Extremely Random Tree, Gradient Boosting Tree, AdaBoost, XGBoost, LightGBM, CatBoost, Support Vector Regression, K-Nearest Neighbors, and Deep Neural Networks (DNN).

[0122] Configuration of the method of the present invention.

[0123] LLM-IGWO-DAE-MBSFL-GPR: A complete white-box interpretable model, which is the core method.

[0124] LLM-IGWO-DAE-MBSFL-GPR+Residual: Based on the white-box model, the ExtraTrees model is introduced to compensate for the prediction residuals in order to explore the upper limit of performance.

[0125] As shown in Table 3, the experimental results indicate that TGP-6 achieves the best overall prediction performance among all traditional genetic programming configurations. However, its results depend on extremely high structural complexity and computational cost, leading to a significant reduction in model readability and practicality. Therefore, this embodiment selects it as the performance upper limit baseline to rigorously compare it with the best performance of traditional methods, thereby more clearly highlighting the substantial improvements of this invention in improving prediction performance, computational efficiency, and structural interpretability.

[0126] Table 3. Performance comparison of different TGP configurations in predicting ash distortion temperature

[0127] Table 4 shows the performance comparison results of the linear model (R²). 2 While stable, the R² value (≈0.65) has limited capabilities and cannot fully characterize complex nonlinear relationships; traditional genetic programming (TGP-6) has a best R² value. 2The R² value was improved to 0.7577, but it suffers from a large standard deviation, long training time (5915.80 seconds), and a complex and difficult-to-understand structure. In contrast, the LLM-IGWO-DAE-MBSFL-GPR method proposed in this invention achieves Best R². 2 (0.8035), predictive stability (Mean R) 2 The method demonstrates superior performance in both computational efficiency (0.7955±0.0052) and training efficiency (382.14 seconds, only about 1 / 15 of TGP-6). Through the synergy of shallow symbolic feature learning and improved optimization strategies, this method significantly enhances the modeling capability and computational practicality of nonlinear relationships while maintaining clear and interpretable structures.

[0128] Table 4. Performance comparison between the method of this invention and the baseline model in predicting ash deformation temperature.

[0129] To systematically evaluate the structural interpretability of different models in predicting ash deformation temperature, this invention conducts a comprehensive comparison from three levels: linear models, traditional genetic programming (TGP), and the method of this invention. The results clearly demonstrate the breakthroughs of this invention in terms of performance and interpretability.

[0130] Interpretable comparative analysis of LLM-IGWO-DAE-MBSFL-GPR.

[0131] (1) Linear baseline model: clear structure but limited capabilities.

[0132] Select the best performing ElasticNet model (R) 2 =0.6549), its expression is an explicit linear combination of no interaction terms: ; Although the model has complete mathematical transparency, it is limited by the linear assumption and cannot characterize the complex nonlinear coupling and threshold effect between gray components, resulting in a clear upper limit to the prediction accuracy.

[0133] (2) Traditional genetic programming models: improved predictive ability but unreadable structure (e.g.) Figure 3 (As shown).

[0134] The optimal model of traditional TGP-6 (R 2 While the coefficient (=0.7577) can capture some nonlinear modes, its expression is usually a complex structure with deep nesting and a large number of constants and redundant operators. This "programmatic stacking" makes it difficult to decompose the model into substructures with independent physical meaning, resulting in its interpretability remaining only at the formal level and failing to effectively support mechanism analysis and knowledge extraction.

[0135] (3) The method of the present invention achieves a balance between high performance and high interpretability.

[0136] The LLM-IGWO-DAE-MBSFL-GPR method proposed in this invention can generate explicit symbolic models with clear structure and semantic representation. For example, two high-performance models (R... 2 The expressions (0.8035 and 0.8012 respectively) are composed of a linear combination of multiple shallow symbolic features: Model 1 (R) 2 =0.8035).

[0137] ; Model 2 (R) 2 =0.8012).

[0138] ; While both models demonstrate good predictive performance, Model 1 is superior in overall performance and structural simplicity. Therefore, this embodiment showcases the interactive features in Model 1 and their accompanying large-scale semantic representation, providing domain experts with possible mechanistic analysis approaches, facilitating the verification of feature action paths, and inspiring subsequent research.

[0139] This describes the limited synergistic effect between refractory oxides and fluxing oxides. In the coal ash melting reaction, It participates in the formation of the low-melting-point iron silicate phase, and These phases participate in the formation of high-melting-point phases such as mullite and spinel. When both participate in the same melting phase transformation pathway, the progress of the reaction system is often limited by the one with the lower content, exhibiting "weakest link" melting control. This reflects the mechanistic characteristics of mineral phase evolution kinetics under compositional constraints.

[0140] Logarithmic transformation of Ca emphasizes its strong fluxing contribution in low-content regions, consistent with the chemical mechanism by which Ca participates in the formation of calcium silicate glass phases and Ca–Fe–Si composite low-melting phases. In coal ash, Ca content often exhibits a skewed distribution, with low-Ca regions showing strong sensitivity and high-Ca regions potentially gradually transforming into stable calcium silicates. Logarithmic mapping captures this nonlinear segmental behavior and the "low-dose effect" of Ca on the silicon-oxygen network.

[0141] This quadratic term reflects the silicon-oxygen network enhancement effect in the high-silicon region. At higher silicon content, the Si–O–Si bridging bonds increase, forming a continuous three-dimensional network structure, which increases the viscosity of coal ash and the liquid phase formation temperature, thereby raising the ash melting point. This nonlinear expression reflects a network enhancement mechanism where the high-silicon effect is more pronounced than in the low-silicon region, consistent with the viscous flow properties of coal ash.

[0142] :reflect and The system exhibits synergistic refractory action between different components. It may form spinel (MgA2O4) or a stable, highly fusible phase with mullite, resulting in high viscosity and a high melting point in the coal ash. This characteristic reveals a synergistic mineral phase mechanism of high-melting-point oxides, a typical expression of a refractory system.

[0143] Both K and Fe2O3 are fluxing components, and their combination term reflects the synergistic fluxing effect. + K can disrupt the silicon-oxygen network to form a low-viscosity glassy phase, while multiple oxidation states of Fe participate in the formation of low-melting-point iron silicates. When both coexist, they can accelerate the formation of the low-melting-point phase and advance the ash melting deformation temperature. Furthermore, in coal geochemistry, K and Fe exhibit a covariant relationship, which may characterize the coupling between sedimentary diagenesis and mineral occurrence.

[0144] The cube root transformation suggests that this involves the control of scale or specific surface area on melting kinetics. If GSA is related to ash particle size or specific surface area, smaller particles are more reactive, generate a liquid phase, and lower the melting temperature, while larger particles exhibit refractoriness and delayed phase transformation. The cube root compression of extreme differences while preserving kinetic trends reflects the expansion from compositional control to structural kinetic control during coal ash melting.

[0145] The results show that LLM-IGWO-DAE-MBSFL-GPR not only outperforms existing interpretable modeling methods in terms of numerical performance, but also achieves a fundamental enhancement of interpretability at the model structure level, enabling the model output to exist as a readable, discussable, and inheritable form of knowledge representation.

[0146] A comparison between GWO and IGWO.

[0147] Table 5 shows that, to verify the effectiveness of the improved gray wolf optimization strategy (IGWO) in this framework, it is compared with the traditional GWO under the same settings. Experimental results show that IGWO outperforms the traditional GWO in prediction performance, stability, and efficiency: its Best R... 2 Increased to 0.8035, Mean R 2 The standard deviation of IGWO was 0.7955 with a significant reduction (0.0052 vs 0.0160), and both RMSE and MAE improved accordingly. Meanwhile, the average training time of IGWO was slightly lower than that of GWO (382.14 seconds vs 391.7 seconds). This indicates that IGWO can achieve more stable and efficient symbolic search and feature evolution at a similar computational cost, making it more suitable for symbolic regression tasks that require both interpretability and performance.

[0148] Table 5 Performance Comparison of Improved Gray Wolf Optimization (IGWO) and Traditional Gray Wolf Optimization (GWO)

[0149] Comparison with the mainstream black box model.

[0150] As shown in Table 6, in direct comparison with various mainstream black-box models (such as CatBoost, DNN, RandomForest, etc.), the method of this invention demonstrates its unique advantages: its pure white-box version (LLM-IGWO-DAE-MBSFL-GPR, Mean R) 2 =0.7955) Without relying on uninterpretable structures, the prediction performance is comparable to high-performance ensemble models; and after introducing residual compensation (LLM-IGWO-DAE-MBSFL-GPR+Residual, Mean R 2 =0.8154), its performance even surpasses all compared black-box models, achieving the best prediction accuracy. Although the training time of this method (approximately 7 minutes) is longer than that of black-box models trained in seconds, the process aims to generate explicit symbolic models with clear structure, reusability, and explicit physical interpretation, rather than simply performing parameter fitting. After training, the model's prediction speed is extremely fast, comparable to other black-box models in practical deployment. Therefore, this invention achieves remarkable progress in prediction performance, interpretability, and engineering practicality by exchanging a moderate one-time computational investment for complete transparency of model decisions, verifiability of results, and long-term reusability of derived knowledge.

[0151] Table 6. Comparison of the overall performance of the method of this invention and mainstream black box models in predicting ash deformation temperature.

[0152] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0153] Therefore, the present invention employs the above-mentioned interpretable knowledge mining system, method, device and medium, which can solve the problems of poor model interpretability, high usage threshold and lack of intelligent interpretation ability in the prior art, realize the full-process automation and interpretable mining from data to knowledge, and improve the efficiency and depth of cross-domain data mining.

[0154] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An interpretable knowledge mining system, characterized in that, include: The data processing module is used to load, preprocess, and partition datasets; The algorithm and intelligent interpretation module is used to receive the preprocessed and partitioned structured dataset output by the data processing module, execute the benchmark algorithm to build the control model, and generate an interpretable model with explicit mathematical structure based on the LLM prior knowledge guidance mechanism and the genetic programming hybrid algorithm. At the same time, the LLM and retrieval enhancement generation RAG mechanism is introduced to provide intelligent interpretation of the generated mathematical model. The algorithm and intelligent interpretation module includes: an enhanced genetic programming hybrid algorithm engine for building interpretable mathematical models that are presented in explicit mathematical structures and can be reviewed and verified; The enhanced genetic programming hybrid algorithm engine includes: a symbolic feature learning unit, which is used to build a feature pool based on the genetic programming symbolic framework, introduces a hybrid initialization strategy that combines LLM and random strategies, and generates symbolic features with concise structure and interpretability based on a multi-branch shallow symbolic feature learning tree structure and a dynamic branch expansion strategy. The symbolic feature learning unit includes: A multi-branch shallow symbolic feature learning tree structure is used to generate multiple shallow expressions in parallel to learn multidimensional symbolic features by limiting the maximum depth of a single expression tree. An adaptive dynamic branching expansion strategy is used to adaptively adjust the number of branches and structural complexity of the representation tree based on changes in the fitness of elite individual groups. LLM-guided initialization strategy is used to generate prior symbolic features based on domain knowledge from a large language model to guide the construction of the initial population in genetic programming. A phased gray wolf optimization selection strategy is used to embed the improved gray wolf optimization algorithm into the selection stage of genetic programming. Through a dynamic hierarchical regulation mechanism based on population diversity indicators and a three-stage nonlinear convergence factor, the search behavior is adaptively guided to balance global exploration and local development capabilities. The phased gray wolf optimization selection strategy is implemented in the following steps: Step 1: In each generation of evolution, the current population is adjusted based on the fitness value of each individual. Sort them and divide them into four levels: ； in, For the collection of all individuals in the present day; It consists of the best individuals in the current population; and It consists of second-best and third-best individuals respectively; Includes the remaining individuals; Step 2: Define population diversity For use of a normalized metric based on fitness statistics to guide The proportion of each layer: ； in, The set of fitness scores for all valid individuals in the current population; For the first The fitness value of each individual; The fitness standard deviation; It has extremely poor adaptability; To prevent division by zero of small constants; Dynamic regulation can be divided into two scenarios: when Dynamic hierarchical regulation Defined as: ； in, To control the adjustment range of dynamic hierarchical regulation, The diversity threshold is used; the dynamic update of the proportions of each layer is as follows: ； ； ； ； when Dynamic hierarchical regulation Defined as: ； The dynamic updates of the proportions of each layer are as follows: ； ； ； ； in, These are the base ratios under the condition of no adaptive perturbation. After the update, normalization processing needs to be performed: ； in, Four different search layers; proportional parameters Used to measure its proportion in the current generation, ; The sum of four different search layers; The proportions of the four different search layers after normalization; Step 3: Introduce a three-stage nonlinear convergence factor For guidance Layer selection strategy: ； in, The maximum number of generations; For the current algebra; This represents the percentage at the end of the exploration phase. This represents the percentage at the end of the balancing phase. Let be the initial convergence factor; The mid-term convergence factor; The final convergence factor; and All are adjustable indices; The visualization analysis module is used to receive the data processing results from the data processing module and the model training and interpretation results from the algorithm and intelligent interpretation module, and generate comprehensive visualization results for display. The operation management module connects to all other modules and is used to record the operation history, management result files, and system logs of each module.

2. The interpretable knowledge mining system according to claim 1, characterized in that, The data processing module includes: The data loading unit supports asynchronous loading of data in various formats and provides data preview, descriptive statistics, and multi-dimensional visualization functions. The data preprocessing unit performs missing value analysis, outlier detection and processing, and data standardization operations, and records the operation history to support rollback. The data partitioning unit is used to parameterize the partitioning of the training and test sets, and provides distribution consistency tests, including distribution comparison plots based on kernel density estimation and statistical results of the Kolmogorov–Smirnov test.

3. The interpretable knowledge mining system according to claim 1, characterized in that, The algorithm and intelligent explanation module also includes: A benchmark algorithm engine to support batch execution, parameterized configuration, and performance evaluation of various machine learning algorithms; The Large Language Model Collaboration Unit is used to build an LLM-guided pool of prior knowledge and feature candidates for the genetic programming process, and to provide intelligent interpretation of the mathematical models generated by the enhanced genetic programming hybrid algorithm engine.

4. The interpretable knowledge mining system according to claim 3, characterized in that, The enhanced genetic programming hybrid algorithm engine also includes: The master prediction modeling unit is used to take features from the feature pool constructed by the symbolic feature learning unit as input, model them through a linear regression model, and generate a master prediction expression that can be reviewed. The residual compensation modeling unit is used to fit and compensate the prediction residuals using a black-box model based on the model built by the main prediction modeling unit.

5. The interpretable knowledge mining system according to claim 3, characterized in that, Large language model co-working units include: A unified adaptation unit for multiple providers is used for unified access and scheduling of APIs for various large language models; The intelligent prompting engineering construction unit is used to construct prompt words in a structured manner using a unified template, and to comprehensively set the roles, task requirements and context information of the large language model; The LLM candidate feature pool construction unit is used to combine prompt word templates with RAG technology to retrieve external domain knowledge bases and generate candidate symbol features with domain prior constraints. The model semantic analysis and parsing unit provides domain background knowledge for LLM through RAG technology, parses the generated mathematical model, and extracts the feature mechanism and optimization suggestions.

6. The interpretable knowledge mining system according to claim 1, characterized in that, The visualization and analysis module includes: The baseline algorithm visualization unit is used to generate bar charts comparing algorithm performance, scatter plots of predictions, residual analysis plots, and batch run statistical charts. The genetic programming results visualization unit is used to draw evolutionary curves, expression tree structure diagrams, fitting curves, feature importance bar charts, and fitness distribution histograms. The comprehensive comparative analysis unit is used to load baseline algorithm and genetic programming result files, generate a comprehensive performance dashboard, and support multi-chart page-turning browsing; Interactive chart units are used to enable real-time chart interaction, before-and-after comparison modes, and legend control and management.

7. A method for mining interpretable knowledge, characterized in that, An application to the interpretable knowledge mining system as described in any one of claims 1-6 includes the following steps: The raw data is loaded through the data processing module, and data preprocessing and partitioning are performed. By running a hybrid algorithm of benchmark algorithm and genetic programming through an algorithm and intelligent interpretation module, an interpretable mathematical model is generated and intelligently interpreted. The visualization analysis module displays the results of data processing, model training, and interpretation. The operation management module records the entire process operation log and history to ensure that the experiment is reproducible.

8. A computer device, comprising: Memory and processor; The memory stores a computer program, characterized in that when the processor executes the computer program, it implements the steps of the interpretable knowledge mining method as described in claim 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When a computer program is executed by a processor, it implements the steps of the interpretable knowledge mining method described in claim 7.

Images