Integrating user preferences into distance metrics

Abstract

Many practical scenarios such as search, classification or clustering benefit from better understanding their users, for instance, to deliver more relevant search results. Instead of committing ourselves to a specific field of research, e.g. by generating user profiles to enhance information retrieval, we seek to incorporate user preferences into the distance metric itself which lies at the heart of many algorithms including Information Retrieval and Machine Learning. The two approaches we explore in this paper allow users to directly convey their preferences in an intuitive way. The first approach adheres to the idea that just stating whether two documents are similar or not is more intuitive for a user than, for instance, assigning them to a broad spectrum of topics. The second approach seeks to take into account a user's mental construct of the world being provided with a user-specific concept hierarchy. To evaluate our two approaches, we perform a text classification task. In the classification setting we use the Reuters RCV1 corpus to simulate user preferences. Our results indicate the principal feasibility of these two approaches and encourage further investigations.