A ranking framework and evaluation for diversity-based retrieval

There has been growing momentum in building information retrieval (IR) systems that consider both relevance and diversity of retrieved information, which together improve the usefulness of search results as perceived by users. Some users may genuinely require a set of multiple results to satisfy their information need as there is no single result that completely fulfils the need. Others may be uncertain about their information need and they may submit ambiguous or broad (faceted) queries, either intentionally or unintentionally. A sensible approach to tackle these problems is to diversify search results to address all possible senses underlying those queries or all possible answers satisfying the information need. In this thesis, we explore three aspects of diversity-based document retrieval: 1) recommender systems, 2) retrieval algorithms, and 3) evaluation measures. This first goal of this thesis is to provide an understanding of the need for diversity in search results from the users’ perspective. We develop an interactive recommender system for the purpose of a user study. Designed to facilitate users engaged in exploratory search, the system is featured with content-based browsing, aspectual interfaces, and diverse recommendations. While the diverse recommendations allow users to discover more and different aspects of a search topic, the aspectual interfaces allow users to manage and structure their own search process and results regarding aspects found during browsing. The recommendation feature mines implicit relevance feedback information extracted from a user’s browsing trails and diversifies recommended results with respect to document contents. The result of our user-centred experiment shows that result diversity is needed in realistic retrieval scenarios. Next, we propose a new ranking framework for promoting diversity in a ranked list. We combine two distinct result diversification patterns; this leads to a general framework that enables the development of a variety of ranking algorithms for diversifying documents. To validate our proposal and to gain more insights into approaches for diversifying documents, we empirically compare our integration framework against a common ranking approach (i.e. the probability ranking principle) as well as several diversity-based ranking strategies. These include maximal marginal relevance, modern portfolio theory, and sub-topic-aware diversification based on sub-topic modelling techniques, e.g. clustering, latent Dirichlet allocation, and probabilistic latent semantic analysis. Our findings show that the two diversification patterns can be employed together to improve the effectiveness of ranking diversification. Furthermore, we find that the effectiveness of our framework mainly depends on the effectiveness of the underlying sub-topic modelling techniques. Finally, we examine evaluation measures for diversity retrieval. We analytically identify an issue affecting the de-facto standard measure, novelty-biased discounted cumulative gain (α-nDCG). This issue prevents the measure from behaving as desired, i.e. assessing the effectiveness of systems that provide complete coverage of sub-topics by avoiding excessive redundancy. We show that this issue is of importance as it highly affects the evaluation of retrieval systems, specifically by overrating top-ranked systems that repeatedly retrieve redundant information. To overcome this issue, we derive a theoretically sound solution by defining a safe threshold on a query-basis. We examine the impact of arbitrary settings of the α-nDCG parameter. We evaluate the intuitiveness and reliability of α-nDCG when using our proposed setting on both real and synthetic rankings. We demonstrate that the diversity of document rankings can be intuitively measured by employing the safe threshold. Moreover, our proposal does not harm, but instead increases the reliability of the measure in terms of discriminative power, stability, and sensitivity.