A Deep Reinforcement Learning-Based Approach to Query-Free Interactive Target Item Retrieval

We consider the task of query-free interactive target item retrieval. In this task, a user has a concept or category of items in mind and the retrieval system has to find the right item that falls within the category. Early in a session, the degree of uncertainty about the category of items of interest to the user is high. Hence, it may be more efficient to explicitly ask users about their preference than to use a traditional recommender system (RS) approach that displays very similar items that have the highest estimate of being relevant. We propose a deep reinforcement learning-based approach for relevance feedback interactions between user and system. We introduce an actor-critic framework to iteratively select sets of items based on real-time relevance feedback from users and their purchase history, thereby maximizing satisfaction with the entire session. We compare our proposal with state-of-the-art relevance feedback methods as well as RSs; it leads to increased user satisfaction within a session, independent of the way in which we measure user satisfaction and of the number of items displayed on the result page.