Bias Propagation and Effective Boundaries of Debiasing in Recommender Systems: An Oracle-Based Diagnosis Study

[Objective] Exposure bias causes the interactions observed by recommender systems to diverge systematically from true user preferences, yet two practical questions remain unresolved: how bias propagates across the recommendation pipeline, and under what conditions inverse propensity scoring (IPS) debiasing yields a net benefit. This study aims to measure both precisely. [Methods] We build an oracle-based diagnostic framework on synthetic data with known ground truth and closed-form propensity scores, decompose the pipeline into three stages (recall, ranking, mixed ranking), and evaluate naive, IPS and SNIPS estimators through bias-strength sweeps, propensity dynamic-range sweeps, bootstrap significance tests, a user-side extension, and a propensity-estimation-error extension. [Results] Bias primarily induces ranking reversal rather than simple level shifts: the Spearman correlation between observed and true rankings declines from +0.87 to -0.51 as bias strength increases. An effective boundary emerges at ESS/N ≈ 0.358 below which IPS/SNIPS debiasing turns detrimental; the boundary is stable across scales and robust to propensity estimation error (|Δ| ≤ 0.0023). User-side bias must be modeled at user-item granularity. [Limitations] The numerical boundary 0.358 is specific to our synthetic setup; the framework is static and does not simulate feedback loops; real-data experiments are restricted to transferability checks. [Conclusions] The primary cost of bias is ranking distortion, not level inflation. Practitioners should measure the propensity dynamic range (ESS/N) before applying IPS, and careful measurement should take precedence over further algorithm development.