Use of discrete choice models with recommender systems

Abstract

Recommender systems, also known as personalization systems, are a popular technique for reducing information overload and finding items that are of interest to the user. Increasingly, people are turning to these systems to help them find the information that is most valuable to them. A variety of techniques have been proposed for performing recommendation, including content-based, collaborative, knowledge-based and other techniques. All of the known recommendation techniques have strengths and weaknesses, and many researchers have chosen to combine techniques in different ways. In this dissertation, we investigate the use of discrete choice models as a radically new technique for giving personalized recommendations. Discrete choice modeling allows the integration of item and user specific data as well as contextual information that may be crucial in some applications. By giving a general multidimensional model that depends on a range of inputs, discrete choice subsumes other techniques used in the literature. We present a software package that allows the adaptation of generalized discrete choice models to the recommendation task. Using a generalized framework that integrates recent advances and extensions of discrete choice allows the estimation of complex models that give a realistic representation of the behavior inherent in the choice process, and consequently a better understanding of behavior and improvements in predictions. Statistical learning, an important part of personalization, is realized using Bayesian procedures to update the model as more observations are collected.

(cont.) As a test bed for investigating the effectiveness of this approach, we explore the application of discrete choice as a solution to the problem of recommending academic courses to students. The goal is to facilitate the course selection task by recommending subjects that would satisfy students' personal preferences and suit their abilities and interests. A generalized mixed logit model is used to analyze survey and course evaluation data. The resulting model identifies factors that make an academic subject "recommendable". It is used as the backbone for the recommender system application. The dissertation finally presents the software architecture of this system to highlight the software package's adaptability and extensibility to other applications.