C-SOUL: A Principled Framework for Preference Expansion via Causal Deconfounding and Offline Reinforcement Learning

C-SOUL (Causal System for Open-ended User Learning) is a recommendation framework that replaces engagement optimisation with a mathematically grounded objective: measurable user growth. It combines causal deconfounding via Inverse Propensity Scoring, probabilistic user state modelling via a β-VAE with LSTM backbone, and offline reinforcement learning via a Counterfactual Decision Transformer. A novel Flourishing Reward function directly penalises preference stagnation, incentivising the system to expand user taste horizons through content they actually enjoy. Policy superiority is verified pre-deployment using a Doubly Robust estimator. Prototyped on a synthetic dataset mimicking KuaiRec's structure, achieving a DR policy value of 6.071 and 55.1% effective sample size fraction.