Research on Prediction Methods Based on the KAN-LSTM Hybrid Model

Abstract

This study investigates a prediction method based on the KAN-LSTM hybrid model, aiming to achieve classification, prediction, and impact assessment of complex systems. Firstly, this study uses the K-means clustering algorithm to classify multi-feature datasets, constructs a feature matrix by selecting numerical indicators, and determines the optimal number of clusters using the Calinski-Harabasz index to systematically reveal the differentiated performance characteristics of different sample groups in time series. Secondly, to address the need for non-linear trend prediction, the KAN-LSTM hybrid model is constructed, integrating binary features, numerical features, and time-series data. The LSTM network captures the dynamic trends of the sequence, while the KAN algorithm calculates sample similarity and performs weighted averaging. Additionally, the Ada-Boost ensemble strategy is employed to optimise the prediction output. Finally, the difference-in-differences (DID) model and regression discontinuity design (RDD) model are employed to conduct causal inference analysis on the impact effects of intervention events. Through the integration of multiple methods, the study forms a complete analytical chain of ‘data classification - trend prediction - impact assessment,’ providing methodological references for modelling complex systems.