Research: Correcting Exposure Bias in Recommender Systems with Implicit User Feedback U...

Research question: How can causal inference techniques be leveraged to correct exposure bias in recommender systems trained on implicit feedback data?

Recommender systems have become integral to online platforms, where implicit user feedback (such as clicks, views, or purchases) often serves as the primary signal for learning user preferences. However, implicit feedback is inherently biased because users are only exposed to a subset of items, leading to exposure bias that can distort model training and evaluation.

The existing literature on recommender systems largely focuses on improving predictive accuracy but frequently overlooks the challenges posed by exposure bias in implicit feedback settings. While some recent works propose methods for bias correction, there is limited research on systematic application of causal inference techniques to address exposure bias, particularly in large-scale, real-world datasets.

This project will conduct a rigorous literature review, formulate hypotheses, and design experiments employing causal inference methods (e.g., inverse propensity scoring, counterfactual estimation) to correct exposure bias in recommender systems. Public datasets such as MovieLens and Amazon Reviews will be used to evaluate the effectiveness of these corrections compared to standard baselines.

The findings are expected to contribute to reproducible analytics and fairer recommendations, improving both algorithmic performance and user experience. This research addresses a critical gap in making recommender systems more robust and equitable in the presence of implicit feedback biases.