Beyond Simple Relevance: Balancing Heterogeneous Criteria in Information Retrieval Applications
Abstract
Information retrieval (IR) is an important research area that studies how to find the most useful information to satisfy users' needs. Typical IR applications include search engine, e-commerce, social network, etc. Among various evaluation criteria, relevance is often the most important one, i.e. the retrieved information should accurately match the users' intentions. However, in many real world scenarios, relevance is not the only factor under consideration. For example, a recommender system may wish to generate accurate recommendations within affordable time limits; or, a search engine may wish to reduce redundant web pages present to users, even if all of them are relevant. In these cases, an extra criterion (e.g. efficiency, diversity) has to be considered beyond relevance. How to make such a good balance between the traditional relevance criterion and the other criteria becomes an important research task. In this dissertation, we study how to balance heterogeneous criteria in real world IR applications. By the term “heterogeneous”, we mean criteria that are vastly different from relevance, which implies that they cannot be perfectly satisfied by merely optimizing relevance. We first cast the problem into a multi-objective learning framework, and discuss how multi-objective learning can be mathematically formulated. Then we present four case studies, which cover two IR scenarios (web search and recommendation) and three distinct criteria beyond relevance (efficiency, intrinsic diversity and profit). Specifically, the four cases are: efficiency in recommendation, intrinsic diversity in web search, efficiency in web search and profit in recommendation. We elaborate how a multi-objective learning problem is set up to suit each case, and present experiments to show the effectiveness of our proposed algorithms.
Degree
Ph.D.
Advisors
Clifton, Purdue University.
Subject Area
Computer science
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.