Data mining method and system based on artificial intelligence driving

Abstract

The invention discloses a data mining method and system based on artificial intelligence driving. The method comprises the steps that S1, an original data stream is received, and heterogeneous data feature signals are generated; s2, inputting the heterogeneous data characteristic signal into a dynamic quality sensing module, generating a quality evaluation feedback signal and a purification trigger signal, and outputting a cleaned optimized characteristic signal; s3, in response to the optimization feature signal, synchronously outputting a classification label signal, a clustering correlation signal and an abnormal scoring signal; s4, generating an interpretability enhancement signal based on the mode recognition control signal, and generating a semantic interpretation signal aligned with the business rule; s5, the semantic interpretation signal is received, a closed-loop optimization instruction signal is generated, the feature extraction strategy and the model hyper-parameters are dynamically adjusted through the reinforcement learning agent, and a self-adaptive iterative mining link is formed. According to the data mining method and system based on artificial intelligence driving, the problems of heterogeneous data fusion distortion, poor dynamic environment adaptability and untrusted model decision can be solved.

Description

Technical Field

[0001] This invention relates to the field of cross-modal analysis and adaptive optimization in artificial intelligence data mining, specifically to data mining methods and systems driven by artificial intelligence. Background Technology

[0002] Current AI-driven data mining technologies primarily rely on deep learning models to process structured data. When faced with multi-source, heterogeneous data, manual design of feature fusion rules is required, leading to significant information loss. Existing data cleaning methods using static thresholds struggle to adapt to dynamic noise distributions, often resulting in the accidental deletion of valid features. Most mainstream mining models are end-to-end black-box structures, with their decision-making logic disconnected from business rules, hindering their application in highly sensitive fields such as finance and healthcare. Traditional interpretability technologies like LIME only provide local approximate explanations and cannot trace the global decision-making chain. Model optimization relies on offline parameter tuning, requiring full retraining when data distribution shifts, resulting in slow response and huge resource consumption. Privacy-preserving computation architectures such as federated learning are disconnected from business mining processes, making it difficult to balance privacy protection and model utility. Publicly available patents involve multimodal fusion methods but fail to address the dynamic environment adaptation problem; they propose interpretable model frameworks but lack domain knowledge constraints; and the described online learning systems lack closed-loop feedback mechanisms. While academia has made breakthroughs in cross-modal representation learning, practical engineering still faces bottlenecks such as inaccurate feature alignment and excessive computational overhead. Mainstream industry platforms such as IBM Watson Mining rely on fixed process pipelines, which are insufficiently adaptable to fluctuations in data quality and changes in business objectives. Therefore, there is an urgent need to build a new generation of data mining systems that integrate dynamic cleansing, knowledge guidance, and closed-loop evolution. Summary of the Invention

[0003] In view of the shortcomings of the existing technologies, the purpose of this invention is to provide an artificial intelligence-driven data mining method and system to solve the problems of heterogeneous data fusion distortion, poor adaptability to dynamic environments, and unreliable model decisions. This invention achieves closed-loop intelligent mining through a five-level signal transmission link. First, it constructs a multi-source spatiotemporal feature fusion engine to eliminate modal differences and utilizes generative adversarial networks to achieve noise self-cleaning. A deep model constrained by knowledge graphs is introduced to ensure that the decision logic conforms to domain rules. An attention mechanism is used to parse the decision path and generate interpretable semantics. Finally, reinforcement learning is used to dynamically optimize the parameters throughout the process, forming a continuously evolving system from data input to business decisions, breaking through the performance bottleneck of traditional methods in dynamic and complex environments.

[0004] This invention provides an artificial intelligence-driven data mining method, comprising: S1: Receives the raw data stream and generates heterogeneous data feature signals. The structured data, unstructured data and real-time stream data are normalized by an adaptive multi-source fusion engine to form a feature representation signal with a unified dimension. S2: Input the heterogeneous data feature signal into the dynamic quality perception module to generate a quality assessment feedback signal and a purification trigger signal, and output the optimized feature signal after cleaning; S3: In response to the optimized feature signal, the core generation pattern recognition control signal is mined by AI. The deep neural network constrained by the knowledge graph performs multi-task collaborative learning and simultaneously outputs classification label signal, cluster association signal and anomaly score signal. S4: Generate interpretability enhancement signals based on pattern recognition control signals, where an attention weight parsing engine is used to extract key decision paths and generate semantic interpretation signals aligned with business rules; S5: Receive semantic interpretation signals and generate closed-loop optimization instruction signals, wherein the feature extraction strategy and model hyperparameters are dynamically adjusted through reinforcement learning agents, and the parameter update signals are fed back to steps S1 and S3 to form an adaptive iterative mining link.

[0005] In one embodiment of the present invention, the adaptive multi-source fusion engine in step S1 aligns the feature dimensions of heterogeneous data sources through a spatiotemporal attention mechanism. First, it constructs a cross-modal feature mapping matrix to convert unstructured text data into word vector sequence signals, and simultaneously extracts convolutional feature map signals from image data. Then, it uses a temporal alignment module to generate sliding window statistical feature signals for real-time streaming data. Finally, it calculates the weight allocation between different modal features through a multi-head attention layer and outputs a unified feature representation signal that integrates semantic association and spatiotemporal dependency. This signal serves as the input of the dynamic quality perception module to trigger the data quality assessment process.

[0006] In one embodiment of the present invention, the lightweight anomaly detection model in step S2 adopts an adaptive threshold generation network. First, it receives heterogeneous data feature signals and calculates the mutual information entropy value between feature dimensions to generate an entropy change monitoring signal characterizing the stability of data distribution. When the signal exceeds the fluctuation threshold predicted by the hidden Markov model, a data cleanup protocol based on a generative adversarial network is activated. The generator receives damaged feature fragments and outputs reconstructed clean features. The discriminator synchronously generates a feature reconstruction error signal and feeds it back to the generator for adversarial training. Finally, it outputs an optimized feature signal with noise suppression to the subsequent core mining module.

[0007] In one embodiment of the present invention, the knowledge graph constraint in step S3 is implemented through triplet embedding projection. First, the entity relations of the domain knowledge graph are transformed into constraint vector signals in a low-dimensional vector space. The constraint projection matrix is ​​injected into the fully connected layer of the deep neural network to orthogonalize the hidden layer feature vectors and constraint vectors. When the feature vectors deviate from the constraint space, the residual correction mechanism is triggered, and a pattern recognition control signal that conforms to the domain logic is output. This signal simultaneously drives the multi-task learning branch to generate classification probability distribution signals, cluster center distance signals, and anomaly confidence signals.

[0008] In one embodiment of the present invention, the attention weight parsing engine in step S4 constructs an interpretable decision tree. First, it extracts the weight distribution signals of all attention layers in the deep neural network, and recursively aggregates the key attention paths of each layer through a tree structure to generate a complete decision link signal from input features to output results. Then, it inputs the link signal into the natural language generation module, and combines it with the logical predicate template in the business rule knowledge base to automatically synthesize a semantic interpretation signal containing causal relationships. The signal also marks the contribution ranking of key features for manual verification.

[0009] In one embodiment of the present invention, the reinforcement learning agent in step S5 adopts a distributed asynchronous update architecture. First, it receives the business indicator completion score in the semantic interpretation signal and converts it into a reward function increment signal. At the same time, it monitors the model confidence fluctuation value in the pattern recognition control signal to generate an environment state transition signal. The policy network generates a feature selection mask signal and a hyperparameter adjustment signal based on the above signals. The update signal is synchronously fed back to the data fusion engine and the mining core module through multiple parallel executors, forming a cross-step joint optimization mechanism.

[0010] In one embodiment of the present invention, multi-task collaborative learning introduces a gradient conflict resolution mechanism. When the angle between the backpropagation gradient vectors of the classification task and the clustering task exceeds a preset threshold, the gradient projector is activated to calculate the basis vectors of the orthogonal subspace, and the conflicting gradients are projected to the independent subspace for parameter updates. At the same time, a task relevance evaluation signal is generated to control the gradient allocation weight, so that the calculation of abnormal scoring signals can obtain gradient resources first, ensuring the convergence stability of high-value tasks.

[0011] In one embodiment of the present invention, the distributed asynchronous update architecture includes a federated learning coordinator. When the parameter update signal involves multiple data isolation domains, the coordinator encrypts the hyperparameter adjustment signal into homomorphic ciphertext and distributes it to the local executors of each domain. After each domain completes its local model update in the ciphertext state, it generates an encrypted gradient aggregation signal. After the consistency is verified by secure multi-party computation, the signal is decrypted into a global update signal and finally synchronized to the feature extraction strategy library of the central server.

[0012] In one embodiment of the present invention, the data sanitization protocol of the generative adversarial network integrates a meta-learning controller. During the model initialization phase, a cross-domain noise pattern feature library is loaded to generate a basic noise recognizer weight signal. When a new noise feature is input, the controller calculates its cosine similarity signal with the feature library and dynamically adjusts the convolution kernel dilation coefficient of the generator, so that the generator can quickly adapt to unknown noise types without retraining, thereby improving the generalization ability of the sanitization module.

[0013] The present invention also includes an artificial intelligence-driven data mining system, comprising: The acquisition module receives the raw data stream and generates heterogeneous data feature signals to form a feature representation signal with a unified dimension. The coordination module inputs heterogeneous data feature signals to the dynamic quality perception module, generates quality assessment feedback signals and purification trigger signals, and outputs optimized feature signals after cleaning. The comparison module responds to the optimized feature signal, and is driven by AI to mine the core generated pattern recognition control signal, and simultaneously outputs the classification label signal, cluster association signal and anomaly score signal. The pre-visualization module generates interpretability enhancement signals based on pattern recognition control signals. It utilizes an attention weight parsing engine to extract key decision paths and generate semantic interpretation signals aligned with business rules. The analysis module receives semantic interpretation signals and generates closed-loop optimization instruction signals. It dynamically adjusts the feature extraction strategy and model hyperparameters through reinforcement learning agents, feeds back parameter update signals, and forms an adaptive iterative mining chain.

[0014] The data mining method and system based on artificial intelligence provided by this invention achieve closed-loop intelligent mining through a five-level signal transmission link. First, a multi-source spatiotemporal feature fusion engine is constructed to eliminate modal differences, and adversarial generative networks are used to achieve noise self-cleaning. A deep model with knowledge graph constraints is introduced to ensure that the decision logic conforms to domain rules. The decision path is parsed based on the attention mechanism to generate interpretable semantics. Finally, the parameters of the entire process are dynamically optimized through reinforcement learning to form a continuously evolving system from data input to business decision-making, breaking through the performance bottleneck of traditional methods in dynamic and complex environments. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1This is a flowchart of a data mining method based on artificial intelligence. Figure 2 This is a system architecture diagram for an AI-driven data mining system. Detailed Implementation

[0017] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0018] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0019] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0020] Please see Figure 1-2The figure shows the data mining method and system based on artificial intelligence (AI) of the present invention. The data mining method based on AI of the present invention includes: S1: receiving raw data stream and generating heterogeneous data feature signals, wherein structured data, unstructured data and real-time streaming data are normalized by an adaptive multi-source fusion engine to form a feature representation signal with a unified dimension; S2: inputting the heterogeneous data feature signals to a dynamic quality perception module to generate quality assessment feedback signals and purification trigger signals, and outputting cleaned optimized feature signals; S3: responding to the optimized feature signals, the AI-driven mining core generates pattern recognition control signals, wherein a deep neural network constrained by a knowledge graph performs multi-task collaborative learning, and synchronously outputs classification label signals, clustering association signals and anomaly scoring signals; S4: generating interpretability enhancement signals based on the pattern recognition control signals, wherein an attention weight parsing engine is used to extract key decision paths and generate semantic interpretation signals aligned with business rules; S5: receiving semantic interpretation signals and generating closed-loop optimization instruction signals, wherein a reinforcement learning agent dynamically adjusts the feature extraction strategy and model hyperparameters, and feeds back the parameter update signals to steps S1 and S3 to form an adaptive iterative mining chain.

[0021] like Figure 1 As shown, the core method achieves closed-loop optimization of data mining through a five-level signal processing chain. Step S1 designs an adaptive multi-source fusion engine to process heterogeneous inputs such as structured database tables, unstructured text and images, and real-time data streams. A feature mapping matrix is ​​used to project text word vectors, image convolutional features, and streaming window statistical features onto a unified space. A multi-head attention mechanism is used to calculate cross-modal feature association weights, outputting a fused feature signal that preserves semantic and temporal relationships. This signal completely eliminates the modal gap problem caused by manual feature engineering in traditional methods. Step S2 constructs a dynamic quality perception module, using a lightweight anomaly detection model to monitor the distribution stability of feature signals in real time. When the entropy change monitoring signal exceeds the hidden Markov prediction threshold, a generative adversarial network is activated to reconstruct and purify noisy feature fragments, outputting an optimized feature signal to ensure the accuracy of subsequent mining. Step S3 deploys a deep neural network constrained by a knowledge graph, transforming domain knowledge into vector space constraints. An orthogonal projection mechanism ensures that hidden layer features conform to business logic, simultaneously outputting multi-task mining signals to solve the problem of multi-objective optimization of a single model. Step S4 constructs an interpretable decision tree based on attention weights, analyzes the internal decision paths of the model, and synthesizes natural language explanation signals, transforming black-box decisions into verifiable business rules. Step S5 uses a reinforcement learning agent to analyze business metrics in the semantic explanation, generates parameter update signals, and feeds them back to the front-end module, forming a continuously optimized closed-loop system.

[0022] Furthermore, the anomaly detection model in step S2 innovatively employs an adaptive threshold generation network. It constructs a data distribution stability quantification index by calculating the mutual information entropy value between feature dimensions. This index drives the Hidden Markov Model to predict dynamic fluctuation thresholds, significantly improving noise recognition sensitivity compared to fixed threshold schemes. When the entropy change monitoring signal exceeds the threshold, the activated purification protocol uses a dual-channel adversarial mechanism. The generator receives damaged feature fragments and outputs reconstructed features, while the discriminator simultaneously calculates the reconstruction error to generate adversarial training signals. This closed-loop optimization structure allows the generator to continuously evolve to adapt to new noise patterns. A specially designed feature fragment segmentation strategy dynamically segments the input signal according to the information entropy value, ensuring that local noise regions are accurately located and purified, avoiding the loss of effective information caused by global processing. The discriminator uses a multi-scale convolutional structure to extract abnormal patterns in the feature space. Its output error signal constrains the generator's update direction through a gradient penalty mechanism, preventing pattern collapse. The purified optimized feature signal carries an integrity verification label into the downstream module. This label serves as a quality traceability basis for final decision auditing.

[0023] like Figure 1 As shown, the knowledge graph constraint implementation mechanism in step S3 includes two core technologies: triplet embedding projection and residual correction. First, the entity relationships in the domain knowledge graph are transformed into low-dimensional constraint vectors, which are then injected into the fully connected layer of the neural network through a projection matrix. When the hidden layer feature vector deviates from the constraint space, an orthogonal decomposition algorithm is triggered to separate the principal components conforming to the business logic and the abnormal residual components. The residual correction module performs a nonlinear transformation on the abnormal components, generating a correction signal conforming to the knowledge graph relationships, which is then re-injected into the feature stream. Multi-task collaborative learning employs a gradient conflict resolution strategy, monitoring the angle between the gradient vectors of classification and clustering tasks in real time. When the conflict exceeds a threshold, the subspace projector is activated, allocating the conflict gradients to independent optimization channels. Simultaneously, the gradient allocation weights are dynamically adjusted based on the task value evaluation signal. The design of prioritizing computational resources for the anomaly scoring task is implemented through a gradient gating mechanism, whose output confidence signal controls the gradient scaling coefficients of other tasks. This scheme overcomes the performance balancing problem of traditional multi-task learning, and real-world testing in a financial anti-fraud scenario shows a significant improvement in anomaly detection recall.

[0024] Furthermore, the interpretability enhancement stage constructs a tree-like attention parsing engine, recursively aggregating the attention weight distribution signals of each layer of the deep neural network, and generating a complete decision-making chain from input features to output results through a path integral algorithm. This chain signal is input into the natural language generation module and undergoes multiple rounds of matching with logical predicate templates in the business rule knowledge base: first, key feature node sequences in the decision path are extracted and their similarity is calculated with conditional predicates in the knowledge base; second, highly matched predicates are combined according to the decision order to form an initial interpretation framework; finally, feature contribution values ​​are injected to generate structured interpretation text. The dynamic evolution mechanism, after receiving human feedback signals, scans high-frequency feature patterns in historical decision chains, and automatically expands the knowledge base predicate set when the confidence level of a new pattern meets the standard. The version control module records the model decision consistency score after each rule update, ensuring that interpretability and accuracy are improved simultaneously. In medical diagnostic scenarios, the semantic interpretation signals generated by this solution achieve a medical knowledge verification rate of over 97%. The message queue tracing system constructs a full lifecycle monitoring chain, designed based on a distributed log architecture and signal lineage analysis algorithm. All signals transmitted between modules carry four-dimensional metadata: a globally unique operation identifier (including module ID and session UUID), a nanosecond-level version timestamp, a signal type encoding, and a pointer to the preceding signal. The message queue employs a sharded storage strategy; structured signals are stored on the Apache Kafka streaming platform, while unstructured signals are stored in MinIO object storage. When a closed-loop optimization command signal triggers a parameter update, the tracing engine performs a three-stage analysis: in the backtracking stage, it retrieves the version sequence of homogeneous feature signals in the historical queue along the signal pointer; in the differential stage, it compares the KL divergence and covariance shift index of the feature distribution before and after the parameter change; and in the attribution stage, it constructs a multilayer perceptron model to locate key influencing paths. The generated model iteration tracing report signal contains structured fields: parameter change batch number, list of affected feature dimensions, data distribution shift quantification index, and decision consistency change rate. To improve audit efficiency, the system establishes a real-time early warning dashboard: when a single parameter update causes a decrease in decision consistency exceeding a threshold, a parameter rollback signal is automatically triggered and the operations and maintenance module is notified. In securities trading monitoring scenarios, this solution reduces the model iteration compliance audit time from an average of 40 hours to 2 hours per person.

[0025] like Figure 1As shown, the closed-loop optimization phase adopts a distributed reinforcement learning architecture, comprising three main components: a policy network, an environment simulator, and multiple executors. The policy network receives the business KPI completion score from the semantic interpretation signal and converts it into an incremental signal for the reward function. The environment simulator collects indicators such as model confidence and feature stability from the pattern recognition control signal in real time, constructing an environment state transition matrix. The policy network outputs a dual-channel instruction containing a feature selection mask signal and a hyperparameter adjustment signal. The feature selection mask acts on the feature projection layer of the data fusion engine through a sparsity matrix, dynamically closing low-value feature channels. The hyperparameter adjustment signal updates the learning rate, regularization coefficient, and other parameters of the mining core in a differential form. The multiple executor architecture supports parallel trials of different parameter combinations and accelerates policy optimization through an asynchronous gradient aggregation mechanism. In cross-domain scenarios, the federated learning coordinator encrypts the adjustment signal into homomorphic ciphertext for distribution. After local updates in each domain, the signals are securely aggregated to generate a global optimization signal. This scheme improves the system's adaptive efficiency by more than 10 times in data distribution drift scenarios.

[0026] Specifically, the closed-loop optimization mechanism achieves dynamic parameter adjustment across modules through a distributed reinforcement learning architecture. Its core comprises a policy network, an environment state simulator, and a group of parallel executors. The policy network receives business indicator completion scores (such as user conversion rate and risk coverage) from the semantic interpretation signal, normalizes them into an incremental signal for the reward function, and simultaneously monitors model confidence fluctuations and feature stability indices in the pattern recognition control signal to construct a multi-dimensional environment state transition signal. The policy network adopts a dual-channel output architecture: the first channel generates a feature selection mask signal, which is a sparse binary matrix acting on the feature projection layer of the data fusion engine, optimizing computational resources by hard-closing low-contribution feature channels; the second channel outputs hyperparameter adjustment signals, which are transmitted in differential encoding form to the learning rate scheduler, regularization coefficient controller, and knowledge constraint projection weights of the core mining module. The multi-executor architecture is deployed in a cloud computing environment. Each executor loads a copy of the current policy network and independently executes parameter update experiments. The policy gradient signals of each executor are aggregated to the central server through an asynchronous gradient aggregation mechanism. The federated learning coordinator is activated in cross-institutional collaboration scenarios. It transforms hyperparameter adjustment signals into ciphertext vectors using fully homomorphic encryption and distributes them to each participant. Each participant completes model updates locally in their ciphertext state, generating encrypted gradient aggregation signals. These signals are then decrypted and reconstructed after consistency verification via a secure multi-party computation protocol, ultimately synchronizing with the central policy repository to form a closed loop. This design improves the system's adaptive response speed to the second level in scenarios with abrupt data distribution changes, while simultaneously ensuring privacy compliance.

[0027] like Figure 2As shown, the gradient conflict resolution mechanism addresses resource competition in multi-task learning. Its implementation comprises two main components: a gradient angle monitor and a subspace projector. The gradient angle monitor collects real-time spatial direction data of the backpropagation gradient vectors of the classification task and the gradient vectors of the clustering task, calculates their cosine similarity, and generates a conflict coefficient signal. When this coefficient is below a preset threshold, the projector is activated. The subspace projector first performs singular value decomposition on the features of the current hidden layer of the neural network, extracts orthogonal basis vector sets representing task-irrelevant information, constructs an independent optimization subspace, and projects the conflicting gradients onto this subspace for parameter updates to avoid disturbance to the main task space. The task relevance evaluation module uses a sliding window to statistically analyze the value contribution of historical task output signals and generates a dynamic weight allocation signal: the anomaly scoring task receives the maximum weight coefficient due to its high business priority, while the classification and clustering tasks allocate remaining resources according to their real-time confidence ratios. The gradient gating controller generates a channel scaling coefficient matrix based on the weight signal, compressing the gradients of low-priority tasks layer by layer while preserving the complete gradient path for the anomaly scoring task. In real-world testing of e-commerce anti-fraud scenarios, this mechanism improved the convergence speed of anomaly detection tasks by 40% and reduced the accuracy fluctuation range to ±0.5%, effectively solving the seesaw effect of traditional multi-task learning. The meta-learning controller endows the noise removal module with cross-domain transfer capabilities; its core consists of a noise pattern feature library and a dynamic architecture adjuster. During initialization, the feature library loads typical noise spectrum features from fields such as industrial equipment vibration, financial transaction time series, and medical imaging. These features are compressed into low-dimensional pattern vectors using a variational autoencoder and a vector retrieval tree is constructed. When a new noise feature is input, the controller calculates its cosine similarity signal with all pattern vectors in the feature library and selects Top-K nearest neighbor vectors to generate a pattern matching degree distribution map. The dynamic architecture adjuster performs three real-time optimizations based on the matching degree distribution: adjusting the expansion coefficient of the generator convolution kernel to expand the receptive field and capture long-range noise patterns, modifying the channel attention weights of the discriminator to focus on abnormal frequency bands, and resetting the balance coefficient between reconstruction error and adversarial loss in the loss function of the generative adversarial network. The incremental feature library update mechanism is activated after each cleanup task: it extracts the frequency domain wavelet coefficients and spatial autocorrelation features of novel noise, filters them through outlier detection, injects them into the feature library, and reconstructs the index. In cross-plant migration tests for equipment fault diagnosis, this scheme improved the accuracy of unknown noise cleanup from 63% to 89%, and reduced the model cold start time to 1 / 5 of the original scheme.

[0028] like Figure 2As shown, the federated learning coordinator implements cross-privacy domain security optimization, comprising three main subsystems: a ciphertext transformation engine, a local executor cluster, and a secure aggregation verification ring. When the parameter update signal involves multiple data isolation domains, the coordinator first inputs the hyperparameter adjustment signal into the fully homomorphic encryption engine, using the CKKS algorithm based on lattice ciphers to generate a homomorphic ciphertext vector packet. This ciphertext packet retains the ability to perform addition and multiplication operations on floating-point vectors. After being distributed to the local executors in each domain, the executors perform three key operations in the ciphertext state: loading a local model copy and applying the parameter update instructions, running the forward propagation of local data batches to calculate the loss gradient, and encrypting the gradient values ​​into a homomorphic gradient aggregation signal. The secure aggregation verification ring adopts a three-layer protection architecture: the first layer injects Gaussian noise into the gradient signal through a differential privacy mechanism; the second layer uses Shamir secret sharing to shard the key to the verification nodes; and the third layer runs a secure multi-party computation protocol to verify the consistency of the signals from each node. The verified aggregation signal is reconstructed into a global gradient update vector after joint decryption. This vector is then synchronized to the central server after fusing the contributions of each domain using a confidence-weighted algorithm. The coordinator automatically rotates the encryption key and updates the noise distribution parameters every 24 hours to prevent long-term tracking attacks. In the bank's joint risk control system, this solution meets GDPR data residency requirements while maintaining 98% model accuracy. A dynamic knowledge base evolution strategy enables the autonomous evolution of business rules, comprising three main modules: a conflict detector, a rule mining engine, and a version controller. The conflict detector compares semantic interpretation signals with manual review feedback signals in real time. When the number of logical contradictions reaches the sliding window statistical threshold (e.g., in medical diagnosis, model recommendations differ from expert judgments), the rule mining engine is triggered to scan historical decision-making link signals from the past three months. The scanning process employs a distributed graph computing framework: first, a decision path graph is constructed with feature nodes as vertices and attention weights as edges; then, a frequent subgraph mining algorithm is used to extract high-frequency feature combination patterns with support exceeding the confidence threshold; finally, the patterns are transformed into candidate rule vectors (containing premise feature sets, inference operators, and conclusion labels). Confidence verification is performed in two steps: the statistical verification stage calculates the rule's coverage and accuracy in historical data, and the logical verification stage checks its compatibility with the existing knowledge base rule set. After the validated rule vectors are injected into the knowledge base, the version controller generates an update notification signal containing the following elements: a unique identifier for the new rule, an effective timestamp, a list of affected models, and automatically optimizes the predicate matching algorithm weights of the natural language generation module. In insurance claims cases, the system generates an average of 12 new rules per month, continuously improving the consistency rate between automated decision-making and manual review from 82% to 96%.

[0029] The present invention relates to an AI-driven data mining method and system that achieves closed-loop intelligent mining through a five-level signal transmission link. First, a multi-source spatiotemporal feature fusion engine is constructed to eliminate modal differences, and adversarial generative networks are used to achieve noise self-cleaning. A deep model constrained by knowledge graphs is introduced to ensure that the decision logic conforms to domain rules. An attention mechanism is used to parse the decision path and generate interpretable semantics. Finally, reinforcement learning is used to dynamically optimize the parameters of the entire process, forming a continuously evolving system from data input to business decision-making, breaking through the performance bottleneck of traditional methods in dynamic and complex environments.

[0030] Therefore, the data mining method and system based on artificial intelligence of the present invention can solve the problems of distorted heterogeneous data fusion, poor adaptability to dynamic environments, and unreliable model decisions.

[0031] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A data mining method based on artificial intelligence, characterized in that, include: S1: Receives the raw data stream and generates heterogeneous data feature signals. The structured data, unstructured data and real-time stream data are normalized by an adaptive multi-source fusion engine to form a feature representation signal with a unified dimension. S2: Input the heterogeneous data feature signal into the dynamic quality perception module to generate a quality assessment feedback signal and a purification trigger signal, and output the optimized feature signal after cleaning; S3: In response to the optimized feature signal, the AI-driven core generation pattern recognition control signal is generated, wherein a deep neural network constrained by a knowledge graph performs multi-task collaborative learning and simultaneously outputs classification label signal, cluster association signal and anomaly score signal; S4: Generate an interpretability enhancement signal based on the pattern recognition control signal, wherein the attention weight parsing engine is used to extract key decision paths and generate a semantic interpretation signal aligned with business rules; S5: Receive the semantic interpretation signal and generate a closed-loop optimization instruction signal, wherein the feature extraction strategy and model hyperparameters are dynamically adjusted through a reinforcement learning agent, and the parameter update signal is fed back to steps S1 and S3 to form an adaptive iterative mining link.

2. The data mining method based on artificial intelligence driven according to claim 1, characterized in that, The adaptive multi-source fusion engine in step S1 aligns the feature dimensions of heterogeneous data sources through a spatiotemporal attention mechanism. First, it constructs a cross-modal feature mapping matrix to convert unstructured text data into word vector sequence signals. At the same time, it extracts convolutional feature map signals from image data. Then, it uses a temporal alignment module to generate sliding window statistical feature signals for real-time streaming data. Finally, it calculates the weight distribution between different modal features through a multi-head attention layer and outputs a unified feature representation signal that integrates semantic association and spatiotemporal dependency. This signal serves as the input of the dynamic quality perception module to trigger the data quality assessment process.

3. The data mining method based on artificial intelligence driven according to claim 1, characterized in that, The lightweight anomaly detection model in step S2 employs an adaptive threshold generation network. It first receives heterogeneous data feature signals and calculates the mutual information entropy value between feature dimensions to generate an entropy change monitoring signal that characterizes the stability of data distribution. When this signal exceeds the fluctuation threshold predicted by the Hidden Markov Model, a data cleanup protocol based on a generative adversarial network is activated. The generator receives damaged feature fragments and outputs reconstructed clean features. The discriminator synchronously generates a feature reconstruction error signal and feeds it back to the generator for adversarial training. Finally, it outputs an optimized feature signal with noise suppression to the subsequent core mining module.

4. The data mining method based on artificial intelligence as described in claim 1, characterized in that, The knowledge graph constraints in step S3 are implemented through triplet embedding projection. First, the entity relationships of the domain knowledge graph are transformed into constraint vector signals in a low-dimensional vector space. Then, a constraint projection matrix is ​​injected into the fully connected layer of the deep neural network to orthogonalize the hidden layer feature vectors with the constraint vectors. When the feature vectors deviate from the constraint space, a residual correction mechanism is triggered, and a pattern recognition control signal that conforms to the domain logic is output. This signal simultaneously drives the multi-task learning branch to generate classification probability distribution signals, cluster center distance signals, and anomaly confidence signals.

5. The data mining method based on artificial intelligence driven according to claim 1, characterized in that, The attention weight parsing engine in step S4 constructs an interpretable decision tree. First, it extracts the weight distribution signals of all attention layers in the deep neural network. Then, it recursively aggregates the key attention paths of each layer through a tree structure to generate a complete decision link signal from input features to output results. Subsequently, it inputs the link signal into the natural language generation module and combines it with the logical predicate templates in the business rule knowledge base to automatically synthesize a semantic explanation signal containing causal relationships. This signal also marks the contribution ranking of key features for manual verification.

6. The data mining method based on artificial intelligence as described in claim 1, characterized in that, The reinforcement learning agent in step S5 adopts a distributed asynchronous update architecture. First, it receives the business indicator completion score from the semantic interpretation signal and converts it into an incremental signal of the reward function. At the same time, it monitors the model confidence fluctuation value in the pattern recognition control signal to generate an environment state transition signal. The policy network generates a feature selection mask signal and a hyperparameter adjustment signal based on the above signals. The update signal is synchronously fed back to the data fusion engine and the core mining module through multiple parallel executors, forming a cross-step joint optimization mechanism.

7. The data mining method based on artificial intelligence as described in claim 1, characterized in that, The multi-task collaborative learning introduces a gradient conflict resolution mechanism. When the angle between the backpropagation gradient vectors of the classification task and the clustering task exceeds a preset threshold, the gradient projector is activated to calculate the basis vectors of the orthogonal subspace. The conflicting gradients are projected onto the independent subspace for parameter updates. At the same time, a task relevance evaluation signal is generated to control the gradient allocation weights, so that the calculation of abnormal scoring signals can obtain gradient resources first, ensuring the convergence stability of high-value tasks.

8. The data mining method based on artificial intelligence as described in claim 1, characterized in that, The distributed asynchronous update architecture includes a federated learning coordinator. When the parameter update signal involves multiple data isolation domains, the coordinator encrypts the hyperparameter adjustment signal into homomorphic ciphertext and distributes it to the local executors of each domain. After each domain completes its local model update in the ciphertext state, it generates an encrypted gradient aggregation signal. After the consistency is verified by secure multi-party computation, the signal is decrypted into a global update signal and finally synchronized to the feature extraction strategy library of the central server.

9. The data mining method based on artificial intelligence as described in claim 1, characterized in that, The data sanitization protocol of the generative adversarial network integrates a meta-learning controller. During the model initialization phase, a cross-domain noise pattern feature library is loaded to generate a basic noise recognizer weight signal. When a new noise feature is input, the controller calculates its cosine similarity signal with the feature library and dynamically adjusts the convolution kernel dilation coefficient of the generator, enabling the generator to quickly adapt to unknown noise types without retraining and improving the generalization ability of the sanitization module.

10. A data mining system using the AI-driven data mining method according to any one of claims 1-9, characterized in that, include: The acquisition module receives the raw data stream and generates heterogeneous data feature signals to form a feature representation signal with a unified dimension; The coordination module inputs the heterogeneous data feature signals to the dynamic quality perception module, generates a quality assessment feedback signal and a purification trigger signal, and outputs the optimized feature signals after cleaning. The comparison module, in response to the optimized feature signal, is driven by AI to mine the core to generate a pattern recognition control signal, and simultaneously outputs a classification label signal, a clustering association signal, and anomaly score signal; The pre-visualization module generates an interpretability enhancement signal based on the pattern recognition control signal, wherein the attention weight parsing engine is used to extract key decision paths and generate a semantic interpretation signal aligned with business rules. The analysis module receives the semantic interpretation signal and generates a closed-loop optimization instruction signal, wherein the feature extraction strategy and model hyperparameters are dynamically adjusted through a reinforcement learning agent, the parameter update signal is fed back, and an adaptive iterative mining link is formed.

Images