C2ST-Bandit for collaborative multi-category retail dynamic pricing with causal contrastive learning and contextual bandit

In multi-category retail settings, demand interactions across product categories make it difficult for isolated single-category pricing strategies to achieve system-level optimization. Existing dynamic pricing methods based on contextual bandit often model cross-category dependence primarily through statistical associations while overlooking potential causal transmission structures, making the resulting decisions susceptible to confounding effects. This paper proposes C2ST-Bandit, a collaborative dynamic-pricing framework that integrates classifier two-sample testing (C2ST), causal contrastive learning, and contextual bandit decision-making. C2ST is used to detect and quantify changes in demand distributions associated with observed cross-category price variation. The resulting causal contrastive representation is incorporated as an augmented context feature into a Linear Upper Confidence Bound (LinUCB)-based pricing policy to support collaborative multi-category optimization. The framework contains three main components: a C2ST-based cross-category distribution-testing mechanism, a causal contrastive encoder that maps high-dimensional heterogeneous demand signals into a low-dimensional causal embedding, and a causal-aware contextual bandit decision layer for which a cumulative-regret upper bound is derived under a sparse causal-graph assumption. Experiments on two real retail datasets show that C2ST-Bandit reduces cumulative regret by 18.6% and 21.4%, respectively, relative to Independent LinUCB, improves convergence speed, and achieves an F1 score of 0.91 ± 0.017 for causal-edge identification. Ablation experiments demonstrate the complementary contributions of the main modules, while sensitivity analysis confirms the robustness of the framework to key hyperparameter settings. Clinical trial number Not applicable.