Efficient, safe and adaptive learning from user interactions

User interactions occur naturally in modern Information Retrieval systems and can provide implicit feedback for a retrieval system. A big challenge in using user interactions is that they are both noisy and biased. In this dissertation we identify and propose solutions for three major challenges in unbiased learning from user interactions: efficiency, safety and adaptiveness.
First, counterfactual learning can be inefficient due to the high variance introduced by the inverse propensity scores. To address this problem, this dissertation proposes a novel learning algorithm with provably better convergence rate than traditional IPS-weighted SGD.
Second, we find that historical interaction data may be limited and interactions on potentially high quality items may be missing. To tackle this challenge we introduce a safe counterfactual learning algorithm that can periodically deploy its learned model to change what user interactions are gathered.
Third, we look at the adaptiveness of counterfactual learning in situations where user preferences change over time. We propose two new estimators, based on weighting historical interactions, that enjoy reduced bias in the non-stationary setting.
Finally, the thesis looks beyond search-specific interactions to other types of interactions, looking specifically at activity logs in a cloud file storage system. This line of work explores how interaction logs beyond traditional search logs can be used to improve the ranking quality of modern Information Retrieval systems.