Session Search by Direct Policy Learning

Abstract

This paper proposes a novel retrieval model for session search. Through gradient descent, the model finds optimal policies for the best search engine actions from what is observed in the user and search engine interactions. The proposed framework applies direct policy learning to session search such that it greatly reduce the model complexity than prior work. It is also a flexible design, which includes a wide range of features describing the rich interactions in session search. The framework is shown to be highly effective evaluated on the recent TREC Session Tracks. As part of the efforts to bring reinforcement learning to information retrieval, this paper makes a novel contribution in theoretical modeling for session search.