[논문 리뷰] The Deconfounded Recommender: A Causal Inference Approach to Recommendation

Zhaoran Wang, Dawen Liang외 2명|arXiv (Cornell University)|2018. 08. 20.

Advanced Causal Inference Techniques참고 문헌 6인용 수 43

한 줄 요약

이 논문은 deconfounded recommender를 도입합니다. 두 단계의 인과적 접근법으로 exposure modeling via Poisson factorization을 통해 substitute confounders를 추정하고 debias ratings를 수행하여 추천에서의 미관측 교란에 대한 강건성을 향상시킵니다.

ABSTRACT

The goal of recommendation is to show users items that they will like. Though usually framed as a prediction, the spirit of recommendation is to answer an interventional question---for each user and movie, what would the rating be if we "forced" the user to watch the movie? To this end, we develop a causal approach to recommendation, one where watching a movie is a "treatment" and a user's rating is an "outcome." The problem is there may be unobserved confounders, variables that affect both which movies the users watch and how they rate them; unobserved confounders impede causal predictions with observational data. To solve this problem, we develop the deconfounded recommender, a way to use classical recommendation models for causal recommendation. Following Wang & Blei [23], the deconfounded recommender involves two probabilistic models. The first models which movies the users watch; it provides a substitute for the unobserved confounders. The second one models how each user rates each movie; it employs the substitute to help account for confounders. This two-stage approach removes bias due to confounding. It improves recommendation and enjoys stable performance against interventions on test sets.

연구 동기 및 목표

• Motivate recommendation as a causal inference problem with interventions (watching a movie) and outcomes (ratings).
• Develop a two-model framework to address unobserved confounding in observational data.
• Leverage exposure data to infer substitute confounders and adjust the ratings model accordingly.

제안 방법

• Model exposure with Poisson factorization to obtain per-user latent factors as substitute confounders.
• Reconstruct an exposure proxy \u001bhat{a}_{ui} from the Poisson factorization posterior.
• Fit an outcome model that includes the substitute confounder to predict ratings when a movie is watched (y_{ui}(1)).
• Compute potential ratings for unseen movies using the fitted outcome model to produce causal recommendations.
• Show that this deconfounding approach reduces bias and yields more robust predictions under confounding.

실험 결과

연구 질문

• RQ1 How can unobserved confounders bias causal interpretation in recommendation systems?
• RQ2 Can a substitute confounder derived from exposure modeling correct ratings models to achieve unbiased causal predictions?
• RQ3 Does the deconfounded recommender improve rating prediction and ranking under confounded observational data?

주요 결과

• The deconfounded recommender yields more accurate rating predictions than classical matrix factorization baselines under varying levels of confounding.
• It outperforms IPW-based causal methods on random test sets in both real and simulated data.
• On Movielens datasets, it achieves lower per-item MSE/MAE than baselines, indicating better universal prediction across items.
• The method remains robust to unobserved confounding and improves NDCG/recall metrics in ranking tasks on random test sets.
• In simulations, the approach shows greater resilience to confounding than classical MF methods and IPW variants.