Research map / Learning theory and algorithms / Kernels and regression theory
Fine-Grained Dynamic Framework for Bias-Variance Joint Optimization on Data Missing Not at Random
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proves limits of regularizers and unbiased estimators for MNAR debiasing, then proposes a dynamic per-user-item estimator (D-IPS/D-DR) that adaptively picks an exponent alpha to jointly optimize bias and variance, yielding bounded variance and generalization bounds.
Problem
Debiased estimators for missing-not-at-random (MNAR) recommendation data (IPS, DR, regularizers) have unbounded variance and generalization bounds when propensities approach zero, and bias-variance trade-offs are handled statically.
Approach
Replace static IPS/DR weights with a dynamic weight 1/f^alpha(p_hat) where alpha in [0,1] is optimized per user-item pair by minimizing w1*bias^2-type term plus w2*variance term, giving a closed-form optimal alpha and bounded variance/generalization bound.
Key results
- On Yahoo! R3, D-DR reaches AUC 0.6999 vs 0.6863 for DR (about 2% gain) and NDCG@5 gain of 4.53% over DR on Coat/Yahoo gain table.
- On Coat, D-DR improves AUC by 3.53% over DR and D-IPS improves AUC by 3.16% over IPS.
- Dynamic estimators outperform static IPS/DR variants on most metrics, but gains for SNIPS and MRDR-JL variance reduction are not obvious.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply dynamic estimator weighting to DR | DR with MF · COAT | AUC: 0.7538±0.0032 -> 0.7804±0.0023 | Helped |
| apply dynamic estimator weighting to DR | DR with MF · COAT | NDCG@5: 0.6425±0.0096 -> 0.6671±0.0051 | Helped |
| apply dynamic estimator weighting to DR | DR with MF · Yahoo! R3 | AUC: 0.6863±0.0013 -> 0.6999±0.0026 | Helped |
| apply dynamic estimator weighting to DR | DR with MF · Yahoo! R3 | NDCG@5: 0.6738±0.0033 -> 0.7043±0.0042 | Helped |
| use log propensity mapping instead of power mapping (ablation) | D-DR with MF · COAT | AUC: 0.7710±0.0014 -> 0.7804±0.0023 | Helped |
| use log propensity mapping instead of sine mapping (ablation) | D-DR with MF · COAT | AUC: 0.7763±0.0021 -> 0.7804±0.0023 | Helped |
| use log propensity mapping instead of tanh mapping (ablation) | D-DR with MF · COAT | AUC: 0.7792±0.0019 -> 0.7804±0.0023 | Helped |
| use log propensity mapping instead of power mapping (ablation) | D-DR with MF · COAT | NDCG@5: 0.6384±0.0047 -> 0.6671±0.0051 | Helped |
| use log propensity mapping instead of sine mapping (ablation) | D-DR with MF · COAT | NDCG@5: 0.6516±0.0052 -> 0.6671±0.0051 | Helped |
| use log propensity mapping instead of tanh mapping (ablation) | D-DR with MF · COAT | NDCG@5: 0.6608±0.0053 -> 0.6671±0.0051 | Helped |
| increase the bias-variance weight ratio (ablation) | dynamic estimators · COAT and Yahoo! R3 | prediction performance | Mixed |
Surprising findings
- Higher unbiasedness does not directly translate to better performance; the best performance and lowest variance occur at intermediate weight ratios rather than the least-biased setting.
- Dynamic framework gives little variance or performance benefit for SNIPS and MRDR-JL, which already reduce variance by sacrificing unbiasedness.
Limitations
- Search for optimal weights w1, w2 and the function f remains open, as the authors state.
- Theoretical guarantees assume accurate propensities (p_hat = p) for the optimal alpha; gains are modest on some settings and metrics.
- Experiments limited to two small public datasets with matrix factorization as the base model.
Experimental setup
Datasets: COAT Yahoo! R3
Benchmarks: COAT Yahoo! R3
Metrics: AUC NDCG@5
Baselines: naive MF IPS SNIPS IPS-AT CVIB IPS-V2 DR DR-JL MRDR-JL Stable DR Stable MRDR TDR-CL
Builds on: inverse propensity scoring doubly robust estimator Stable DR MRDR TDR
Closest papers by meaning
- Unveiling Extraneous Sampling Bias with Data Missing-Not-At-Random (NeurIPS 2025 · similarity 0.88)
- The Bias-Variance Tradeoff in Data-Driven Optimization: A Local Misspecification Perspective (NeurIPS 2025 · similarity 0.85)
- Addressing Hidden Confounding with Heterogeneous Observational Datasets for Recommendation (NeurIPS 2024 · similarity 0.85)
- TranSUN: A Preemptive Paradigm to Eradicate Retransformation Bias Intrinsically from Regression Models in Recommender Systems (NeurIPS 2025 · similarity 0.84)
- Relaxing the Accurate Imputation Assumption in Doubly Robust Learning for Debiased Collaborative Filtering (ICML 2024 · similarity 0.84)
- Debiased Collaborative Filtering with Kernel-Based Causal Balancing (ICLR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.