Source-linked AI summary
Asymmetric Tri-training for Debiasing Missing-Not-At-Random Explicit Feedback
Yuta Saito
TL;DR
Real-world recommender ratings are often missing-not-at-random, and propensity-based methods face estimation and variance problems. The paper proposes model-agnostic meta-learning based on asymmetric tri-training; experiments report robustness to those issues.
Problem
Real-world rating feedback is often missing-not-at-random, while propensity-based methods are sensitive to propensity-estimation error and high variance.
Source: §1 · pp. 1–2Method
The method uses two predictors to generate pseudo-ratings for training a third predictor and minimizes a propensity-independent upper bound.
Source: §1 · p. 2, §4.2 · p. 6Results
Takeaways & Limitations
The results support upper-bound minimization as an alternative approach to MNAR explicit-feedback recommendation.
Source: §1 · p. 2, §6 · p. 9The NB (true) propensity estimator requires MCAR explicit feedback to estimate the prior rating distribution and cannot be used in most real-world problems.
Source: §5.1.2 · p. 7
Abstract
from arXiv · showhide
In most real-world recommender systems, the observed rating data are subject to selection bias, and the data are thus missing-not-at-random. Developing a method to facilitate the learning of a recommender with biased feedback is one of the most challenging problems, as it is widely known that naive approaches under selection bias often lead to suboptimal results. A well-established solution for the problem is using propensity scoring techniques. The propensity score is the probability of each data being observed, and unbiased performance estimation is possible by weighting each data by the inverse of its propensity. However, the performance of the propensity-based unbiased estimation approach is often affected by choice of the propensity estimation model or the high variance problem. To overcome these limitations, we propose a model-agnostic meta-learning method inspired by the asymmetric tri-training framework for unsupervised domain adaptation. The proposed method utilizes two predictors to generate data with reliable pseudo-ratings and another predictor to make the final predictions. In a theoretical analysis, a propensity-independent upper bound of the true performance metric is derived, and it is demonstrated that the proposed method can minimize this bound. We conduct comprehensive experiments using public real-world datasets. The results suggest that the previous propensity-based methods are largely affected by the choice of propensity models and the variance problem caused by the inverse propensity weighting. Moreover, we show that the proposed meta-learning method is robust to these issues and can facilitate in developing effective recommendations from biased explicit feedback.
1 INTRODUCTION
Real-world rating feedback is often MNAR, while propensity-based debiasing is vulnerable to propensity-estimation error and high variance. The paper proposes model-agnostic meta-learning and reports improved, more robust recommendation quality.
1 INTRODUCTION
- Real-world selection bias makes observed ratings unrepresentative, because exposure depends on past recommendation policies and users’ rating choices.These mechanisms make much real-world feedback missing-not-at-random (MNAR).Source: §1 · p. 1
- Propensity-based methods weight ratings by inverse observation probabilities, but their effectiveness depends on accurate propensity estimates and can suffer from high variance.Model misspecification and diverse item popularity or user activity make these problems difficult to avoid.Source: §1 · p. 2
- The proposed method improves recommendation quality and remains robust when propensity scores are difficult to estimate.Experiments also find prior propensity-based methods affected by propensity-model choice and estimator variance.Source: §1 · p. 2
- The paper proposes model-agnostic meta-learning that minimizes a propensity-independent upper bound of the ideal loss for selection-biased recommendation.The introduction presents this as the first such method for recommender systems.Source: §1 · p. 2
2 RELATED WORK
Related work covers propensity-based debiasing, off-policy methods, and asymmetric tri-training for domain adaptation. It identifies a gap: upper-bound minimization had not been applied to MNAR recommendation.
2.1 Propensity-based Recommendation
- For MNAR explicit feedback, IPS and doubly robust estimators use propensity weighting, but their performance depends on propensity and imputation-model accuracy.Accurate propensity estimation is difficult when analysts cannot control the missing mechanism.Source: §2.1 · p. 2
2.2 Off-policy Evaluation and Learning
- DACPOL optimizes a propensity-independent lower bound and empirically outperforms POEM when past propensities are unknown [2]; upper-bound minimization had not been proposed for MNAR recommendation.The stated gap reflects the difficulty of propensity estimation in real-world recommenders.Source: §2.2 · p. 3
2.3 Unsupervised Domain Adaptation
- Asymmetric tri-training for domain adaptation is interpreted as minimizing H-divergence, and this paper extends upper-bound minimization to MNAR recommendation.The paper says this framework had shown effectiveness in unsupervised domain adaptation.Source: §2.3 · p. 3
3 PRELIMINARIES
This section defines the ideal rating-prediction loss and contrasts estimators for missing-not-at-random feedback. The naive estimator is biased under MNAR, while inverse propensity scoring depends on estimated observation probabilities and can suffer from bias and high variance.
3.1 Problem Formulation
- The ideal loss measures prediction error over the user-item pairs, but missing true ratings make it impossible to calculate directly.The section frames estimating this loss from observed feedback as necessary for effective recommendation under MNAR.Source: §3.1 · p. 3
3.2 Naive Estimator
- Under MNAR, the naive estimator's average loss over observed ratings is biased, unlike under the MCAR assumption.The section therefore motivates an estimator that addresses this bias.Source: §3.2 · p. 3
3.3 Inverse Propensity Score Estimator
- IPS weights observed-rating losses by inverse propensity to estimate ideal loss, with propensity defined as the probability a rating is observed.The estimator is unbiased against ideal loss when its propensity scores are correct.Source: §3.3 · p. 4
- IPS can lose unbiasedness when propensity estimates are misspecified, and large inverse propensities can produce high variance and a loose generalization bound.The bound depends on both bias and variance; propensity-estimation error can enlarge the bias term.Source: §3.3 · p. 4
4 METHOD
Asymmetric tri-training uses two predictors to create pseudo-ratings and trains a third predictor on them. Its analysis derives a propensity-independent upper bound and identifies how the training procedure addresses parts of that bound.
4.1 Meta-learning procedure
- Two predictors accept pseudo-ratings for unlabeled user-item pairs when their predictions are sufficiently similar; the remaining predictor trains on the resulting dataset.The predictors may use different recommendation algorithms, and sampling user-item pairs is optional.Source: §4.1 · p. 4
4.2 Theoretical Analysis
- The proposed method targets a propensity-independent upper bound under a triangle-inequality loss and a bounded-loss assumption.The bound depends on pseudo-rating loss, agreement between the first two predictors, and the second predictor's ideal loss; it can be loose if that predictor performs poorly.Source: §4.2 · pp. 5–6
- Yahoo! R3 has a larger train-test rating-distribution shift than Coat, while both datasets show significantly different distributions.The analysis identifies the KL divergence between training and test rating distributions as the comparison measure.Source: §4.2 · p. 6, Figure 1 · p. 6
- During pseudo-labeling, A1 and A2 are expected to become more similar, while training A3 on generated pseudo-labels minimizes the pseudo-rating-loss term.The method attempts to keep the agreement term small rather than minimize it; experiments report minimizing the sum of those two terms.Source: §4.2 · p. 6
5 EXPERIMENTAL RESULTS
Experiments on real-world rating datasets show that asymmetric tri-training improves robustness to propensity-estimation choice and variance, while minimizing its propensity-independent upper bound tracks improved test performance.
5.1 Experimental Setup
- The experiments use MovieLens 100K, Yahoo! R3, and Coat, whose training ratings are MNAR; Yahoo! R3 and Coat provide MCAR test ratings.For MovieLens 100K, the test set is resampled to give each item a uniform observed probability.Source: §5.1 · p. 6
5.1.2 Compared methods and propensity estimators.
- MF-IPS with AT uses two differently initialized MF-IPS predictors and a third predictor trained with naive loss, making its final training step independent of propensity scores.The true-prior propensity estimator is included only for reference because it requires MCAR explicit feedback, unavailable in most real-world problems.Source: §5.1.2 · p. 7
5.2 Results & Discussions
- MF-IPS without AT worsens as propensity skew increases, whereas AT performs relatively well, especially under large skew, because its final prediction step avoids inverse-propensity weighting.The setup varies min_items to change propensity-score skew; the reported advantage is especially pronounced at smaller min_items.Source: §5.2 · p. 7, §5.1 · p. 7, Figure 2 · p. 7
- Across Yahoo! R3 and Coat, AT stabilizes MF-IPS across propensity estimators, especially on Yahoo! R3; it improves Yahoo! rating prediction and Coat nDCG@3, while Coat rating prediction is similar.Table 1 compares prediction and ranking performance on both datasets.Source: Table 1 · p. 8, §5.2 · p. 8
- The method minimizes the sum of pseudo-label loss and predictor similarity, the propensity-independent upper bound of ideal loss.Figure 3 tracks both terms and their sum during pseudo-labeling.Source: §5.2 · pp. 8–9, Figure 3 · p. 9
- Test-set MSE almost monotonically decreases during pseudo-labeling, supporting upper-bound minimization as an approach to improve prediction accuracy.Figure 4 reports averaged test-set MSE and standard deviations during the pseudo-labeling step.Source: §5.2 · p. 9, Figure 4 · p. 9
6 CONCLUSION
The paper concludes that its meta-learning method minimizes part of a propensity-independent upper bound while remaining robust to propensity misspecification and variance issues. Future work would extend the approach to other domain-adaptation methods and implicit-feedback recommendation.
6 CONCLUSION
- The model-agnostic meta-learning method minimizes part of the propensity-independent upper bound of ideal loss while keeping the bound tight during training.Source: §6 · p. 9
- Experiments show the method is robust to variance issues and the choice of propensity estimation model, which affect previous propensity-based recommendations.Source: §6 · p. 9
- Future work plans to apply other unsupervised domain-adaptation methods and develop a similar approach for implicit feedback, which has not been thoroughly investigated for debiasing.Source: §6 · p. 9