In a perfect world, a recommendation engine would know how every user feels about every item in its catalog. In reality, we only see a tiny sliver of that matrix. Most items are never seen by most users, and the data we do have is systematically skewed. Users don't rate movies at random; they rate movies they chose to watch, often because they already expected to like them.
This phenomenon is known as selection bias, and it means our data is Missing Not At Random (MNAR). If we train models on this biased feedback without correction, we create a self-fulfilling prophecy: the system recommends what is popular or previously seen, the user interacts with those suggestions, and the model "learns" that its initial biases were correct.
The Feedback Loop and the Missing Data Problem
The fundamental challenge in recommendation is that the "ground truth" is hidden behind a veil of exposure. As Thorsten Joachims and Adith Swaminathan argue in Recommendations as Treatments, we should view a recommendation not as a simple prediction, but as a medical treatment. Just as a doctor only sees the outcome of a drug for patients who actually took it, a recommender only sees a "rating" for items it chose to display.
Early work by Benjamin M. Marlin in Modeling User Rating Profiles For Collaborative Filtering highlighted that ignoring this missingness mechanism leads to significant inaccuracies. If a user hates horror movies, they simply won't watch them. A naive model might interpret the absence of horror movie ratings as "neutral" or "unknown," rather than "strong dislike," leading to poor suggestions.
Furthermore, Carlos Riquelme and Sven Schmit demonstrate in Human Interaction with Recommendation Systems that these systems create a feedback loop. The algorithm's past choices dictate the future training data, potentially trapping users in "filter bubbles" where they are never exposed to diverse content.
Counterfactual Reasoning and Propensity Scoring
To break this loop, researchers borrow tools from causal inference. The goal is to perform counterfactual evaluation: estimating how a user would have reacted to an item they were never shown.
The most common tool for this is Inverse Propensity Scoring (IPS). The intuition is simple: if an item had a very low probability of being shown to a user (a low "propensity"), but the user interacted with it anyway, that interaction is highly informative and should be given more weight in the training process. Tobias Schnabel and colleagues in Recommendations as Treatments: Debiasing Learning and Evaluation showed that by re-weighting the observed data by the inverse of the propensity score, we can create an unbiased estimator of the system's true performance.
However, IPS has a major weakness: if propensities are very small, the weights become massive, leading to high variance and unstable models. Yuta Saito addresses this in Relevance Matrix Factorization by introducing clipping techniques to balance bias and variance.
Doubly Robust and Multiple Robust Estimators
To achieve more stability, researchers have moved toward Doubly Robust (DR) estimators. These methods combine two different approaches:
- An Error Imputation Model that tries to predict the missing ratings directly.
- A Propensity Model that accounts for the probability of an item being seen.
As Xiaojie Wang and Jianzhong Qi explain in Doubly Robust Joint Learning for Recommendation on Data Missing Not At Random, the "double" refers to the fact that the estimator remains unbiased if either the imputation model or the propensity model is correct. This provides a safety net against model misspecification. Recent advancements like TDR-CL by Peng Wu and Multiple Robust Learning by Haoxuan Li extend this further, using multiple candidate models to "cancel out" biases even more effectively.
Addressing the Propensity Contradiction
A persistent thorn in this research is the "propensity contradiction": to debias a system using IPS, you need to know the true propensities, but those propensities are often unknown or biased themselves. Yuta Saito tackles this in Towards Resolving Propensity Contradiction in Offline Recommender Learning by proposing Domain Adversarial Matrix Factorization (DAMF). Instead of trying to estimate propensities, DAMF uses adversarial learning to minimize an error bound that is independent of propensity, allowing for debiasing even when the exposure mechanism is a complete mystery.
Beyond Exposure: Deconfounding and Latent Factors
Sometimes, the bias isn't just about what the system chose to show. There might be "unobserved confounders"—hidden factors like a user's social circle or a sudden marketing blitz—that influence both what a user sees and how they rate it.
Dawen Liang and David Blei introduced The Deconfounded Recommender to handle this. By using latent variable models to capture these hidden influences, they can "subtract" the effect of the confounder, revealing the user's true, unadulterated preference. Similarly, Ruichu Cai and Fengzhu Wu developed REST, which reconstructs latent exposure strategies to account for external factors like holiday promotions that skew interaction data.
By treating recommendation as a causal inference problem rather than a simple pattern-matching task, this field is moving toward systems that don't just echo our past choices, but truly understand our underlying preferences.