Research map / Probabilistic and scientific modeling / Treatment effects
Debiased Collaborative Filtering with Kernel-Based Causal Balancing
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.
Proposes kernel-based causal balancing (adaptive kernel balancing, AKB) in an RKHS for propensity learning in debiased collaborative filtering, applied to IPS and DR estimators, with generalization bounds.
Problem
Propensity-based debiasing for collaborative filtering learns propensities via cross-entropy or with a few manually chosen balancing functions, which fails to satisfy the full causal balancing constraint and leads to insufficient balancing and biased predictions.
Approach
Approximate balancing functions in an RKHS built from Gaussian/exponential kernels (universal, representer theorem), fit the prediction error with kernel functions, and adaptively balance the J kernel functions with largest coefficients via a convex entropy-regularized objective, plugged into IPS (KBIPS) and DR (KBDR) estimators.
Key results
- On Coat, AKBDR-Exp reaches AUC 0.745 versus 0.735 for the best DR baseline (SDR), with significance marked at p<=0.05.
- On Product, AKBDR-Gau achieves AUC 0.782 and NDCG@20 0.503 versus 0.773 and 0.491 for the best baseline (SDR).
- Random kernel balancing (RKB) performs worst among kernel variants, while adaptive kernel balancing (AKB) performs best, attributed to balancing the most important kernel functions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use adaptive rather than random kernel-function selection (ablation) | Kernel-balancing IPS and DR methods · Coat, Music, Product | Debiasing performance | Helped |
| Use adaptive rather than worst-case kernel balancing (ablation) | Kernel-balancing methods · Coat, Music, Product | Debiasing performance | Helped |
| Increase the number of balanced functions J (ablation) | Kernel-balancing methods except WKB · Product | Debiasing performance | Helped |
| Increase J for worst-case kernel balancing (ablation) | WKB · Product | Debiasing performance | No clear effect |
| Use kernel balancing rather than moment balancing (ablation) | Kernel-balancing methods · Product | Debiasing performance | Helped |
| Vary balancing regularization strength γ (ablation) | AKB methods · Music, Product | Debiasing performance | Mixed |
Surprising findings
- Worst-case kernel balancing (WKB) performance does not change with J, unlike AKB and moment balancing, which improve monotonically with J.
Limitations
- Balancing with only finite samples cannot truly balance infinite-order moments; the method approximates functions with a finite set of kernel functions.
- Generalization bounds assume bounded kernel, bounded weights, and Lipschitz loss, and the universal-approximation assumption on the error function is only approximate with error epsilon.
Experimental setup
Datasets: Coat Music Product
Benchmarks: Coat Music Product
Metrics: AUC NDCG@K F1@K
Baselines: IPS SNIPS ASIPS IPS-V2 DR DR-JL MRDR DR-BIAS DR-MSE MR TDR TDR-JL
Builds on: inverse propensity scoring doubly robust estimation covariate balancing propensity score matrix factorization
Closest papers by meaning
- Relaxing the Accurate Imputation Assumption in Doubly Robust Learning for Debiased Collaborative Filtering (ICML 2024 · similarity 0.87)
- Alleviating Dual Biases in Recommendation (Student Abstract) (AAAI 2025 · similarity 0.86)
- Be Aware of the Neighborhood Effect: Modeling Selection Bias under Interference (ICLR 2024 · similarity 0.85)
- Addressing Hidden Confounding with Heterogeneous Observational Datasets for Recommendation (NeurIPS 2024 · similarity 0.84)
- From Pairwise to Ranking: Climbing the Ladder to Ideal Collaborative Filtering with Pseudo-Ranking (AAAI 2025 · similarity 0.84)
- Fine-Grained Dynamic Framework for Bias-Variance Joint Optimization on Data Missing Not at Random (NeurIPS 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.