Research on a Dynamic Behavior Analysis Framework Based on Random Forest and Markov Chain

Abstract

This paper constructs a comprehensive analytical framework integrating constrained optimization, machine learning and dynamic transition modeling based on multi-season competition data. First, statistical feature analysis and correlation tests are adopted to model the stability of the scoring system. On this basis, a constrained convex optimization model is established to inversely estimate unobservable audience voting ratios, and the Spearman correlation coefficient is used for result verification to enhance the reliability and consistency of estimations. Next, the random forest regression model combined with the SHAP interpretation method is applied to quantitatively analyze key factors affecting scoring and voting results, realizing the importance assessment of multi-dimensional features and interpretable modeling. Furthermore, a Markov chain-based dynamic transition model is developed. With the Softmax probability allocation mechanism and utility function incorporated, this model simulates the transition of audience support across different stages and characterizes the dynamic evolution of popularity in competitive scenarios. Experimental results demonstrate that the proposed model can accurately recover hidden voting distributions and effectively capture result discrepancies and dynamic variations under different rules. Featuring good universality, stability and scalability, this method can serve as a reference for complex behavior prediction and dynamic decision analysis.