Iterative Expectation for Multi Period Information Retrieval

Many Information Retrieval (IR) models make use of offline statistical techniques to score documents for ranking over a single period, rather than use an online, dynamic system that is responsive to users over time. In this paper, we explicitly formulate a general Multi Period Information Retrieval problem, where we consider retrieval as a stochastic yet controllable process. The ranking action during the process continuously controls the retrieval system’s dynamics, and an optimal ranking policy is found in order to maximise the overall users’ satisfaction over the multiple periods as much as possible. Our derivations show interesting properties about how the posterior probability of the documents relevancy evolves from users feedbacks through clicks, and provides a plug-in framework for incorporating different click models. Based on the Multi-Armed Bandit theory, we propose a simple implementation of our framework using a dynamic ranking rule that takes rank bias and exploration of documents into account. We use TREC data to learn a suitable exploration parameter for our model, and then analyse its performance and a number of variants using a search log data set; the experiments suggest an ability to explore document relevance dynamically over time using user feedback in a way that can handle rank bias.