CN-122020257-A - Interpretable classification method and system based on rule learning and neural network architecture searching

Abstract

An interpretable classification method and system based on rule learning and neural network architecture searching belong to the field of artificial intelligence and machine learning. The method comprises the following steps of performing discretization and binarization processing on continuous features of original data, performing neural rule learning on data of an input layer, generating a rule set through conjunctive extraction operation, performing neural network architecture searching on a model structure by using reinforcement learning, and finding an optimal scheme of model architecture processing. The system comprises a feature preprocessing module, a rule learning and generating module and a neural network architecture searching module, wherein the feature preprocessing module can realize the steps of the method. Experimental results show that the method has strong competitiveness in the aspects of prediction accuracy and interpretability, can generate highly interpretable logic rules, and simultaneously keeps excellent prediction performance.

Inventors

• WANG YUANGANG
• Cao Hejia
• WEI LIJING
• JIANG XUETING
• HE XINCHENG
• XUE JIAYU
• SHEN BOWEN

Assignees

• 大连民族大学

Dates

Publication Date

20260512

Application Date

20260213

Claims (5)

1. 1. The system for interpretable classification based on rule learning and neural network architecture search is characterized by comprising a feature preprocessing module, a rule learning and generating module and a neural network architecture search module: The feature preprocessing module is used for discretizing the continuous feature value by using a continuous feature discretizing method based on CACC and converting the continuous feature value into binary features by adopting single-heat coding; the rule learning and generating module processes input layer data by adopting neural rule learning data, converts trained MLLP into a discrete CRS structure through weight binarization and combines logic conjunctions and disjunctions to construct a rule set with interpretability; and the neural network architecture searching module is used for restraining the joint layer-extraction layer alternating structure based on the neural architecture searching framework of reinforcement learning, generating a candidate architecture by adopting RNN, and automatically optimizing the network layer number and the node number by taking verification precision as a reward signal.
2. 2. The method of classifying a system according to claim 1, comprising the steps of: s1, discretizing a continuous characteristic value by using a continuous characteristic discretization method based on CACC, and converting the discretized continuous characteristic and original category characteristic into a binary characteristic vector by adopting independent thermal coding; s2, processing input layer data by adopting neural rule learning data, and constructing a rule set with interpretability through logic conjunctions and disjunctions; S3, adopting a neural framework search framework based on reinforcement learning, restraining a conjunctive layer-disjunctive layer alternating structure, adopting RNN to generate a candidate framework, and automatically optimizing the number of network layers and the number of nodes by taking verification precision as a reward signal.
3. 3. The classification method according to claim 2, wherein in the step S1, the continuous feature values are discretized by using a CACC-based continuous feature discretization method, specifically, for each continuous feature First, the minimum value is determined And maximum value Collecting its unique value and arranging them to form a set Calculating midpoints of neighboring values generates a candidate boundary set Initializing a discrete scheme = ; Calculating CACC values of the new scheme by iteratively inserting candidate boundaries: ; ; Wherein, the For the number of categories to be considered, For the number of intervals of the number of intervals, Is the first Class III The number of samples in the interval, ns represents the number of samples, N p,+ represents the number of samples belonging to the p-th class, N +,q represents the number of samples with the value of attribute A falling in the q-th interval, and the scheme of maximizing CACC is selected and updated And then performing single-heat coding conversion on the discretized continuous features and the original discrete features into binary features.
4. 4. The classification method according to claim 2, wherein in step S2, the neural rule learning model is constructed by using the multi-layer logic sensor MLLP as a continuous value version of the concept rule set CRS, and simulating the logical conjunctions and disjunctions by the activation function: ; Wherein, is provided with And Respectively represent corresponding CRS nodes in MLLP And Is a neuron of (a) a neuron of (b); representation MLLP of the first embodiment A weight matrix of layers, wherein E [0, 1]; output must be limited to the [0,1] range when simulating extraction and conjunctive operations; thus, use is made of Function clipping weights, followed by training MLLP models: ; Wherein the method comprises the steps of Is the mean square error of the signal, Representation of The term of the regularization, The model of MLLP is represented as such, Representing all weights of the model.
5. 5. The classification method according to claim 2, wherein the step S3 finds an optimal configuration including the number of network layers and the number of nodes per layer using a neural architecture search framework based on reinforcement learning, so that the resulting MLLP obtains an optimal classification performance while maintaining model interpretability; And adopting RNN as a controller to generate framework description of candidate MLLP models, updating controller parameters by taking verification precision as a reward signal, restricting the framework to follow a conjunctive layer-disjunctive layer alternating mode, a weight form and network depth and width in a search space, introducing distributed parallel training, stopping in advance and parameter sharing strategies, and converting the optimal framework into an interpretable regularization classifier.

Description

Interpretable classification method and system based on rule learning and neural network architecture searching Technical Field The invention belongs to the field of artificial intelligence and machine learning, and particularly relates to an interpretable classification method based on rule learning and neural network architecture searching. Technical Field With the rapid development of artificial intelligence technology, deep learning has been widely applied in the field of machine learning by virtue of its strong feature characterization capability, and has achieved breakthrough progress in a plurality of complex tasks. However, in critical fields where reliability and interpretability requirements of model output are extremely high, such as medical health, financial wind control, judicial decisions, etc., the "black box" characteristics of deep neural networks severely limit practical applications. Therefore, how to enhance the interpretability and auditability of the decision process while ensuring the performance of the model has become a core problem to be solved in the current high risk field. The existing interpretability study mainly comprises two paths, namely a post interpretation method, wherein the prediction mechanism is indirectly revealed by analyzing the output behaviors of the existing complex model, and an endogenous interpretable model, such as a decision tree and a rule set model, is used for realizing the transparency of the reasoning process by depending on the logic expression of the structure. The former has the advantage of strong versatility, is almost suitable for any model architecture, and can be deployed without changing the original model structure. However, the interpretation often lacks accuracy and consistency due to the inability to understand the internal mechanisms of the model deeply, and the true decision logic of the model is difficult to reflect. Furthermore, most methods have only local interpretable capabilities, and it is difficult to provide a global transparent view. The latter enhances the transparency of the model by introducing explicit rules, and in recent years, performance is often improved by combining strategies such as ensemble learning, but in a complex task, the original interpretability is difficult to maintain due to the increased stacking layers of the model structure, so that the application range of the model is limited. To establish an effective balance between model performance and interpretability, researchers have proposed neural rule learning methods aimed at fusing logical rule representations with the expressive power of depth models, enabling the models to automatically generate interpretable rule sets to guide decisions. However, the method still faces bottlenecks such as high complexity of rule generation, weak feature interaction modeling capability, dependence on manual design of a model structure and the like, and the potential of the method in practical application is difficult to fully develop. In recent years, neural architecture searching (Neural Architecture Search, NAS) has become an important means to improve the efficiency of model structure design. The NAS automatically explores the combination of the network structure and the super parameters through mechanisms such as reinforcement learning, evolutionary algorithm or gradient optimization, reduces manual intervention, and dynamically optimizes the depth and the width of the network according to task requirements, so that the adaptability and the generalization capability of the model are improved. In view of the above, there is currently no classification method that can simultaneously achieve model performance, interpretability and structural adaptation. Therefore, the interpretable classification method integrating rule learning and NAS is provided, logic rules are automatically learned to represent the characteristics of structured data, and the method is combined with searching to adapt to the data driving network structure of different tasks, so that the classification precision and the model transparency are effectively balanced, the human intervention cost in model development is reduced, and the method has wide application prospect and important research value. Disclosure of Invention The invention provides an interpretable classification method based on rule learning and neural network architecture searching, and aims to solve the problem that prediction performance and model interpretability are difficult to consider in the existing structured data classification method. The invention realizes the synchronous improvement of classification performance and model transparency by introducing the fusion design of the rule learning mechanism and the neural network architecture search technology, has automatic modeling, high interpretability and good generalization capability, and is suitable for scenes with higher requirements on decision reliability, such as medical