A Dynamic Framework for Causal User Profiling and Treatment Segmentation via Uplift Modeling in Internet Lending

The growth of internet lending has created a need for decision frameworks based on models that are both personalized and causally interpretable. Conventional uplift models detect treatment responsiveness without recognizing user heterogeneity, the temporal consistency of user behavior, or the upstream design choices that carry important causal implications. This paper proposes an integrated and reproducible Causal User Profiling (CUP) framework that combines causal inference, uplift modeling, and response-based segmentation within a single pipeline. CUP realizes treatment-effect heterogeneity through a four-type response taxonomy (Persuadable, Sure Thing, Lost Cause, Do-Not-Disturb) and embeds it in a multi-stage pipeline involving hybrid feature selection (Information Value (IV), Causal Forest importance, Population Stability Index (PSI) stability, and Stepwise refinement), stratified clustering with a “C2 replacement strategy,” and meta-learning via both the X-Learner and the Doubly Robust (DR) Learner using Logistic Regression (LR). A component-wise ablation analysis finds that feature selection increases AUUC by 25–30%, C2 clustering by 10–12%, and the DR-Learner + LR by another 5–8%. Overall, the integrated CUP framework yields 45–50% higher AUUC than the baseline (“all features + no clustering + standard learner”) while retaining behaviorally coherent and temporally stable insights. Methodologically, we provide: (i) an end-to-end causal user profiling framework that interoperates profiling, causal estimation, clustering, and uplift evaluation; (ii) a behaviorally and causally consistent response segmentation mechanism grounded in the potential-outcomes model; and (iii) a reproducible experimental design that quantifies pipeline-level uplift gains through systematic ablation. Applied to large-scale internet-lending data, CUP reveals opportunities for treatment-aware personalization, enabling financial institutions to target Persuadables, support Sure Things, and avoid disturbing Do-Not-Disturbs based on causal evidence.