Vol. 1 No. 01 (2025): Empirical Research

Abstract

We propose a novel Hybrid Temporal-Logic Guided Feature Engineering framework for interpretable AI in social psychology, which integrates structured psychological theories with adaptive reinforcement learning to bridge the gap between behavioral science and machine learning. The core innovation lies in the Temporal-Logic Guided Feature Engineering Module (TL-FEM), which replaces conventional feature extraction by translating psychological constructs into temporal logic rules, thereby generating interpretable behavioral features that align with domain-specific principles. These features are then processed by a Temporal-Logic Constrained Reinforcement Learning Agent (TL-RL), which optimizes intervention strategies while adhering to psychological constraints, ensuring both efficacy and theoretical validity. The hybrid architecture uniquely combines interpretable feature engineering with adaptive learning, enabling dynamic alignment with behavioral insights without sacrificing transparency. Moreover, the framework incorporates state-of-the-art implementations, such as CUDA-accelerated association rule mining and doubao policy representation, to achieve scalable and efficient performance. The proposed method addresses critical challenges in AI-driven behavioral modeling, such as the lack of psychological grounding and opacity in decision-making, while providing actionable explanations for interventions. This work advances the field of interpretable AI by demonstrating how temporal logic and reinforcement learning can be synergistically applied to social psychology, offering a principled approach to designing ethical and adaptive systems for behavioral prediction and influence.

Keywords:

Hybrid Temporal-Logic Guided Feature Engineering;Interpretable AI;Social Psychology;Temporal-Logic Constrained Reinforcement Learning;Behavioral Modeling