Analysis code for "When should recommender systems explore? Dynamic user modeling from randomized exposure logs"

Reproducible analysis notebook accompanying the manuscript "When should recommender systems explore? Dynamic user modeling from randomized exposure logs" (submitted to User Modeling and User-Adapted Interaction). A single Jupyter notebook regenerates every table, figure and number in the paper from the public KuaiRand-Pure dataset (Gao et al., 2022; Zenodo record 10439422). The data are not redistributed here. The notebook implements: leakage-safe features computed from each user's observed candidate-pool history (with automated leakage tests), a temporal train/validation/test split, static and dynamic engagement models (logistic regression, LightGBM, GRU benchmark) with isotonic calibration, user-level cluster bootstrap inference, an exploration-receptivity score (ERS), cross-fitted doubly robust (AIPW) estimation of the effect of off-profile exposure using the randomized exposures, off-policy evaluation of exploration policies (IPS, SNIPS, DR), ablations, subgroup analyses and SHAP interpretation. Runs on Kaggle (CPU, about 1 hour in FULL mode) or locally; see README for instructions. Set RUN_MODE = "FULL" to reproduce the paper (the default "FAST" mode is a smoke test on a user sample). The export step writes a SHA-256 manifest of all inputs and outputs.