Understanding and Improving Recommender Systems’ Performance in the Presence of Practical User-, Item-, and Marketing-Oriented Considerations

Abstract

Users, items, and recommendation algorithms represent three critical components of recommender systems. Users interact with items through the help of recommendation algorithms, and recommendations naturally can impact users' consumption decisions. Users' feedback to the system (e.g., in the form of item preference ratings) is then used as training data for future recommendations. Many prior studies on recommender systems make certain common assumptions about their application domains, such as: the feedback provided by the users to the system is unbiased; the items are available in the system (and, thus, can be recommended) for extended periods of time; the main criterion for a successful recommendation is the relevance of an item to a user (e.g., not taking into account the business considerations of the recommendation providers). My dissertation investigates the use and performance of recommender systems in situations when such assumptions do not hold by looking into several practically relevant scenarios with important user-, item-, and marketing-oriented considerations. The first essay studies a user-oriented consideration when users' behaviors are biased because of the systems' recommendation. I show that such biases can substantially impair the system's capability to improve predictive accuracy, recommendation diversity, and users' consumption relevance over time. And intentional recommendation perturbations substantially amplify the negative impact of high bias degrees and cause long-lasting effects on the system performance. The second essay addresses the methodological challenges when there is an item-oriented consideration, i.e., the items to be recommended have short life length. I propose a deep-learning algorithm and show it has better prediction accuracy overall and on cold-start items than that of the state-of-the-art benchmarks. The third essay focuses on a marketing-oriented requirement when recommending new product categories to existing customers. I develop category-introduction-oriented recommendation methods and estimate the causal economic impact of different strategies for new category recommendations. The findings of my dissertation advance understanding of the impact of the various user, item, and marketing considerations and have numerous practical implications for system developers.