Dynamic identification and intervention of consumer low-carbon behavior in sustainable textile and apparel consumption driven by deep reinforcement learning

Abstract

Promoting low-carbon consumer behavior is essential for climate mitigation, especially in textile and apparel consumption where purchase, use, reuse, and disposal decisions jointly shape carbon and resource impacts. Existing intelligent intervention systems are often constrained by behavioral complexity, biased feedback, and the intention-action gap. To address these limitations, this study proposes Dynamic Causal Disentanglement with Reinforcement Learning (DCDRL), a deep reinforcement learning framework for dynamically identifying consumers’ intrinsic low-carbon motivation and separating it from extrinsic behavioral biases such as conformity and salience effects. By combining causal graph modeling, disentangled representation learning, conservative offline reinforcement learning, and dynamic negative sampling, DCDRL learns debias-aware intervention policies for sustainable textile/apparel and other low-carbon consumption scenarios. The corrected experiments separately evaluate Smart Energy (SE-1M) and Sustainable Choice-Apparel (SC-A) datasets and show that DCDRL outperforms sequential recommendation, deep learning, and offline reinforcement learning baselines under a consistent full-catalog evaluation protocol.