Our article “Balancing Exploration and Exploitation in Online Learning to Rank for IR” (to appear in Information Retrieval) addresses the question of whether online learning to rank systems need to balance exploration and exploitation.
Intuitively, when learning online from natural user interactions with a search system, a search engine needs to exploit what it has learned so far, in order to fulfill user expectations in search performance as well as possible. At the same time, it needs to explore, to ensure that the collected feedback is as informative as possible for future learning. Often, the results that would be most useful for the current search are not the same as those most useful for learning.
In our article we formalize the above problem as an exploration-exploitation dilemma. We develop two approaches for balancing exploration and exploitation in an online learning to rank for IR setting, one based on a pairwise, one based on a listwise learning approach. We show that, as hypothesized, balancing exploration and exploitation improves online performance in both types of approaches. However, the optimal balance depends on the approach, and on other factors, such as the amount of noise in user feedback.
At CIKM 2011, I presented our paper “A Probabilistic Method for Inferring Preferences from Clicks“. It proposes a new method, called Probabilistic Interleaving, for inferring feedback for comparing rankers using click data.
Interleaved comparison methods have been developed recently to compare two candidate rankings (e.g., two similar versions of the ranking algorithm of a search engine) by observing user behavior (in particular clicks on displayed result documents). This form of evaluation is highly scalable, as click data can be collected at very low cost, as long as the search engine is used. It is also assumed to more accurately capture true user preferences than the most common form of evaluation – manual editorial judgments.
Our proposed method addresses problems of sensitivity and fidelity that we identify in previous methods. It is based on a probabilistic interleaving process, where a result list that combines documents from both candidate rankers is generated in such a way that documents that were ranked highly by either ranker are likely to be displayed towards the top of the interleaved list. This interleaved list is displayed to the user, instead of showing either original ranking. User clicks on documents are then assigned to the original rankers in a way that reflects how highly each ranker initially placed these documents. The result is a comparison method that naturally re-weights clicks according to the distance between the original rankings. It is sensitive to even small changes between rankings and at the same time only infers large differences when the distance between the original rankings is large.
An intriguing direction for follow-up work is to investigate how effective the inferred feedback is beyond evaluation setups, e.g., for online learning to rank.