Research map / Probabilistic and scientific modeling / Treatment effects
Be Aware of the Neighborhood Effect: Modeling Selection Bias under Interference
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.
Formulates neighborhood effect in recommendation as an interference problem in causal inference and proposes kernel-smoothed N-IPS and N-DR estimators of a neighborhood-aware ideal loss for unbiased debiased learning.
Problem
Selection bias in recommender systems is usually debiased under the no-interference (SUTVA) assumption, but a user-item pair's feedback can depend on exposures of other user-item pairs (neighborhood effect), making existing IPS/DR estimators biased.
Approach
Introduce a learnable/low-dimensional neighborhood treatment representation g and a neighborhood-aware ideal loss integrated over g; estimate it with kernel-smoothed IPS and DR estimators whose propensity is modeled over the joint treatment (exposure, g), with theoretical bias, variance, optimal bandwidth and generalization bounds.
Key results
- On semi-synthetic ML-100K, N-MRDR reduces relative error to 0.0622 on ONE vs 0.2578 for MRDR, and N-IPS to 0.2383 vs 0.4766 for IPS.
- On KuaiRec, N-IPS with Gaussian kernel and LR propensity reaches AUC 0.796 and NDCG@50 0.585, versus 0.752 and 0.554 for IPS.
- Proposed estimators show lower relative error than corresponding baselines across all six semi-synthetic prediction matrices, with the advantage holding as neighborhood-effect strength increases.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use neighborhood IPS instead of IPS | Ideal-loss estimation on ML-100K · ONE prediction matrix | RE: 0.4766 ± 0.0060 -> 0.2383 ± 0.0066 | Helped |
| Use neighborhood IPS instead of IPS | Ideal-loss estimation on ML-100K · THREE prediction matrix | RE: 0.5501 ± 0.0056 -> 0.2670 ± 0.0069 | Helped |
| Use neighborhood IPS instead of IPS | Ideal-loss estimation on ML-100K · FOUR prediction matrix | RE: 0.5731 ± 0.0057 -> 0.2829 ± 0.0062 | Helped |
| Use neighborhood IPS instead of IPS | Ideal-loss estimation on ML-100K · ROTATE prediction matrix | RE: 0.1434 ± 0.0040 -> 0.0417 ± 0.0043 | Helped |
| Use neighborhood IPS instead of IPS | Ideal-loss estimation on ML-100K · SKEW prediction matrix | RE: 0.1969 ± 0.0046 -> 0.1024 ± 0.0051 | Helped |
| Use neighborhood IPS instead of IPS | Ideal-loss estimation on ML-100K · CRS prediction matrix | RE: 0.1885 ± 0.0028 -> 0.0966 ± 0.0029 | Helped |
| Use N-DR-JL [NB, Gaussian] instead of SDR | MF on Coat · Coat | MSE: 0.210 -> 0.204 | Helped |
| Use N-IPS [LR, Gaussian] instead of DAMF | MF on Coat · Coat | N@5: 0.643 -> 0.678 | Helped |
| Use N-MRDR-JL [NB, Epanechnikov] instead of SNIPS | MF on KuaiRec · KuaiRec | AUC: 0.788 -> 0.802 | Helped |
Surprising findings
- Existing debiasing methods are unbiased for the neighborhood-aware ideal loss only when the neighborhood representation is independent of the individual exposure conditional on features.
Limitations
- The hypothesis space G of the treatment representation g relies on prior knowledge and is hard to choose in practice.
- Theoretical results rely on kernel-smoothing regularity conditions and bounded propensities/losses.
- Neighborhood definitions in real-world experiments are hand-picked among three choices.
Experimental setup
Datasets: MovieLens 100K Coat Yahoo! R3 KuaiRec
Benchmarks: ML-100K semi-synthetic Coat Yahoo! R3 KuaiRec
Metrics: relative error MSE AUC NDCG@5 NDCG@50
Baselines: Naive IPS SNIPS DR DR-JL MRDR-JL CVIB DIB ASIPS DAMF DR-BIAS DR-MSE
Builds on: inverse propensity scoring doubly robust estimation MRDR SNIPS
Closest papers by meaning
- Recommendations as Treatments: Debiasing Learning and Evaluation (ICML 2016 · similarity 0.89)
- Alleviating Dual Biases in Recommendation (Student Abstract) (AAAI 2025 · similarity 0.85)
- Debiased Collaborative Filtering with Kernel-Based Causal Balancing (ICLR 2024 · similarity 0.85)
- Addressing Hidden Confounding with Heterogeneous Observational Datasets for Recommendation (NeurIPS 2024 · similarity 0.84)
- Relaxing the Accurate Imputation Assumption in Doubly Robust Learning for Debiased Collaborative Filtering (ICML 2024 · similarity 0.83)
- Fine-Grained Dynamic Framework for Bias-Variance Joint Optimization on Data Missing Not at Random (NeurIPS 2024 · similarity 0.83)
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.