When building a recommender system, we typically work with a highly sparse matrix of user-item interactions. Traditional collaborative filtering algorithms operate under a convenient but flawed assumption: that the missing data is Missing At Random (MAR). In other words, they assume that whether a user rates a movie is completely independent of how much they will actually like it.
In reality, user feedback is Missing Not At Random (MNAR). As Benjamin M. Marlin et al. demonstrated in Recommender Systems, Missing Data and Statistical Model Estimation, users selectively interact with items based on their latent preferences and external algorithmic recommendations. You do not randomly pick a movie to watch; you pick one you expect to enjoy. This self-selection creates a severe selection bias or positivity bias. If a model treats unobserved items as random missing values, it will systematically overfit to the positive skew of the observed data, leading to poor recommendations across the broader item catalog.
Jointly Modeling Selection and Preference
One way to tackle selection bias is to explicitly model why a data point is missing. Instead of treating the missingness mechanism as a nuisance, we can treat it as a valuable source of information.
In Probabilistic Matrix Factorization with Non-random Missing Data, José Miguel Hernández-Lobato et al. introduced a joint probabilistic framework that couples the rating prediction with an observation process. The model does not just learn a user's preference for an item; it also learns the probability that the user will choose to rate it. Because unobserved items are statistically more likely to be items the user would dislike, this joint modeling allows the system to leverage the absence of an interaction as a negative signal. Similarly, Shohei Ohsawa et al. developed Gated Probabilistic Matrix Factorization: Learning Users' Attention from Missing Values, which treats the reason for consumption as a dynamic filter that determines which latent user preferences are active during an interaction.
Recommendations as Treatments: Propensity Scoring
Another paradigm shifts the problem from generative modeling to causal inference. In Recommendations as Treatments: Debiasing Learning and Evaluation, Tobias Schnabel et al. proposed treating the exposure of an item to a user as a "treatment" in a clinical trial.
If we can estimate the propensity score—the probability $P(O_{ui} = 1)$ that a user $u$ is exposed to item $i$—we can use Inverse Propensity Scoring (IPS) to re-weight the observed data. The intuition is straightforward: if a user interacts with a niche item that had a very low probability of being recommended, that interaction represents a large, silent segment of the population. By weighting the loss function by the inverse of the propensity score ($1 / P(O_{ui}=1)$), we mathematically reconstruct an unbiased estimate of the true population performance.
However, IPS has a major vulnerability: variance. If an item's propensity score is extremely small, its inverse becomes massive, causing a single interaction to wildly swing the model's gradients. Furthermore, as Yuta Saito et al. point out in Towards Resolving Propensity Contradiction in Offline Recommender Learning, standard IPS suffers from a "propensity contradiction" because estimating accurate propensities often requires the very unbiased data we lack in the first place.
The Safety Net of Doubly Robust Learning
To mitigate the high variance of propensity scoring and the model-misspecification risks of pure imputation, researchers turned to Doubly Robust (DR) learning. First introduced to this domain by Xiaojie Wang et al. in Doubly Robust Joint Learning for Recommendation on Data Missing Not at Random, DR estimators combine two distinct components:
- An Inverse Propensity Score model to re-weight observed entries.
- An Error Imputation Model (EIM) that predicts "pseudo-labels" or imputed errors for the unobserved entries.
The magic of the Doubly Robust approach lies in its mathematical redundancy. The estimator is structured such that the overall prediction error remains unbiased if either the propensity model is accurate or the error imputation model is accurate.
To see why this works intuitively, consider the two failure modes:
- If the propensity model is completely wrong: The imputation model acts as a backup, filling in the missing entries with reasonable pseudo-labels so the model does not overfit to the biased observed sample.
- If the imputation model is completely wrong: The propensity-weighted observed entries act as a corrective force, canceling out the bias introduced by the bad imputations.
Because it is highly unlikely that both models fail in ways that perfectly compound, DR learning provides a remarkably stable foundation for training on MNAR data.
Refining and Stabilizing the Estimators
Recent research has focused on making these doubly robust frameworks robust to real-world edge cases. For instance, Haoxuan Li et al. introduced StableDR: Stabilized Doubly Robust Learning for Recommendation on Data Missing Not at Random to bound extreme propensity weights and prevent variance explosions.
In Relaxing the Accurate Imputation Assumption in Doubly Robust Learning for Debiased Collaborative Filtering, the authors demonstrated that DR estimators can be designed to tolerate systematic, user- or item-specific biases within the imputation model itself, freeing practitioners from needing perfect pseudo-labels. To combat the issue of overconfident internal models, WONBIN KWEON et al. proposed the Doubly Calibrated Estimator for Recommendation on Data Missing Not At Random, which incorporates calibration experts to ensure that both propensity scores and imputed errors match empirical realities.
By moving away from the fiction of random missingness and embracing the causal and generative structures of user choice, modern collaborative filtering can reliably learn true user preferences from highly biased, real-world data.
Key Foundational Work
Recommender Systems, Missing Data and Statistical Model Estimation (2011) — contributed 2 concept elements to this direction.
Online learning for collaborative filtering (2012) — contributed 1 concept elements to this direction.