Research map / Probabilistic and scientific modeling / Variational and generative sampling
DR-VAE: Debiased and Representation-enhanced Variational Autoencoder for Collaborative Recommendation
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 DR-VAE, a VAE-based collaborative filtering framework that adds an inverse-propensity debiasing estimator for exposure bias and a continuous-normalizing-flow regularizer to mitigate posterior collapse.
Problem
VAE-based collaborative filtering suffers from exposure bias (unobserved interactions are not necessarily negative) and posterior collapse (KL term forces posteriors toward an oversimplified prior), which degrade recommendation quality.
Approach
Rewrites the VAE reconstruction loss as a propensity-weighted unbiased estimator tailored to exponential-family likelihoods (Poisson, Multinomial, Gaussian, Bernoulli), and replaces the standard KL-to-fixed-prior regularizer with a KL to a continuous-normalizing-flow-defined posterior so the prior can be complex.
Key results
- On semi-synthetic ML-100K (Top@20), DR-VAE reaches Recall 0.1796 and NDCG 0.1407, beating the best baseline S-DRO (0.1746 / 0.1362).
- On KuaiRec (Top@20), DR-VAE achieves Recall 0.2060 and HR 0.9951 versus best baseline InvCF Recall 0.1928.
- On real-world Amazon Toys (Top@20), R-VAE improves Recall over BiVAE from 0.0890 to 0.1045 (about 17% relative gain).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the Debiasing Estimator and Flow-based Representation Enhancement to VAE | VAE · ML-100K Top@20 | Recall: 0.1746* -> 0.1796 | Helped |
| add the Debiasing Estimator and Flow-based Representation Enhancement to VAE | VAE · KuaiRec Top@20 | Recall: 0.1928* -> 0.2060 | Helped |
| add the Flow-based Representation Enhancement module (ablation) | VAE-based recommendation · Amazon Toys | Recall@20: 0.0647 -> 0.1045 | Helped |
| add the Flow-based Representation Enhancement module (ablation) | VAE-based recommendation · ModCloth | Recall@50: 0.4535 -> 0.5569 | Helped |
| add the Debiasing Estimator (ablation) | VAE · ML-100K recommendation rate across item-exposure groups | standard deviation of recommendation rate | Helped |
| add the Flow-based Representation Enhancement module (ablation) | VAE · four datasets | average latent-variable variance | Helped |
| set regularization weight β to 0.01 (ablation) | DR-VAE · ML-100K and KuaiRec | NDCG@20 | Helped |
| set propensity-score exponent η to 0.3 (ablation) | DR-VAE · ML-100K and KuaiRec | NDCG@20 | Helped |
Limitations
- Propensity scores are estimated with a simple heuristic (item-frequency power with hyperparameter eta) rather than a learned model.
- Debiasing is evaluated only on semi-synthetic datasets with simulated intervention data.
- Sensitivity to beta and eta requires tuning; experiments use a single RTX 3090 with limited dataset scale.
Experimental setup
Datasets: ML-100K KuaiRec Amazon Toys ModCloth
Benchmarks: ML-100K KuaiRec Amazon Toys ModCloth
Metrics: Recall Precision NDCG HR
Baselines: IPS-MF Rel-MF MACR BC Loss S-DRO InvCF MultVAE RecVAE BiVAE MF BPR
Builds on: MultVAE BiVAE Inverse propensity scoring (Saito et al. 2020) Continuous normalizing flow (Chen et al. 2018) FFJORD
Closest papers by meaning
- On the Mechanisms of Collaborative Learning in VAE Recommenders (ICLR 2026 · similarity 0.92)
- Personalized Federated Collaborative Filtering: A Variational AutoEncoder Approach (AAAI 2025 · similarity 0.88)
- FACE: A General Framework for Mapping Collaborative Filtering Embeddings into LLM Tokens (NeurIPS 2025 · similarity 0.86)
- Preference Diffusion for Recommendation (ICLR 2025 · similarity 0.85)
- Understanding and Improving Adversarial Collaborative Filtering for Robust Recommendation (NeurIPS 2024 · similarity 0.85)
- Recommendations with Sparse Comparison Data: Provably Fast Convergence for Nonconvex Matrix Factorization (ICML 2025 · similarity 0.85)
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.