Interactive Learning To Rank And Visual Rank Interpretation

Abstract

Many algorithms in the Information Retrieval domain have been developed considering training models using vast amounts of data. The acquisition of this data, however, is time-consuming and requires lots of human effort. Active Learning techniques try to solve this problem by reducing the number of instances needed in the training phase by selecting relevant instances to be labelled. Although such an approach has been proved to be effective, it is still hard to understand how the model is changing after every relevance feedback. As a potential solution, the use of visualizations to help users to understand models is becoming a widespread approach both to understand the overall behaviour of a model and to analyze individual data instances. In this thesis, I explore the utilization of a Learning to Rank algorithm in a relevance feedback scenario and the use of visualizations to understand the reasoning behind the model's ranking decisions.