Probabilistic Modeling in Dynamic Information Retrieval

Dynamic modeling is used to design systems that are adaptive to their changing environment and is currently poorly understood in information retrieval systems. Common elements in the information retrieval methodology, such as documents, relevance, users and tasks, are dynamic entities that may evolve over the course of several interactions, which is increasingly captured in search log datasets. Conventional frameworks and models in information retrieval treat these elements as static, or only consider local interactivity, without consideration for the optimisation of all potential interactions. Further to this, advances in information retrieval interface, contextual personalization and ad display demand models that can intelligently react to users over time. This thesis proposes a new area of information retrieval research called Dynamic Information Retrieval. The term dynamics is defined and what it means within the context of information retrieval. Three examples of current areas of research in information retrieval which can be described as dynamic are covered: multi-page search, online learning to rank and session search. A probabilistic model for dynamic information retrieval is introduced and analysed, and applied in practical algorithms throughout. This framework is based on the partially observable Markov decision process model, and solved using dynamic programming and the Bellman equation. Comparisons are made against well-established techniques that show improvements in ranking quality and in particular, document diversification. The limitations of this approach are explored and appropriate approximation techniques are investigated, resulting in the development of an efficient multi-armed bandit based ranking algorithm. Finally, the extraction of dynamic behaviour from search logs is also demonstrated as an application, showing that dynamic information retrieval modeling is an effective and versatile tool in state of the art information retrieval research