Recommender Systems; Contextual Multi-Armed Bandit Algorithms for the purpose of targeted advertisement within e-commerce

Abstract

The topic of Recommender Systems is and have been a hot topic the last century as the market for e-commerce keeps extending. Basic techniques for recommending popular items are used on more or less all e-commerce platforms. Many e-commerce based platforms use simple techniques such as "people who bought this also bought that", while others have very complex recommender systems for customised recommendations depending on users pro les. Regardless of what techniques that are being used, companies want to make their customers happy as well as increasing their own profit. Because of this there is a constant demand for smart systems using up-to-date algorithms and techniques to achieve relevant advertisements. This thesis focuses on evaluating the performance of Contextual Multi-Armed Bandit Algorithms in a, for the specific algorithm, not yet fully explored use-area of recommender systems, namely the area of garmentbased e-commerce. The evaluation consists in measuring the performance mostly in terms of successful recommendations, while discussing satisfaction-level of customers. In addition to this we decided to experiment with di erent privacy-preserving techniques to see how it affects the performed recommendations. This kind of evaluation is, to our knowledge, absent in current literature, which is why we decided to pursue the idea. The evaluation is carried out through use of self-implemented algorithms. Using the frameworks for machine-learning and implementing recommender systems, Apache Mahout and LensKit, the algorithms used in this thesis are implemented by ourselves in java. The implemented algorithms turned out to be better than what was expected initially, managing to predict purchase-behavior of some users with a probability of over 21%. Through observing the results of the implemented application we made it possible to identify new possible use-areas for Multi-Armed Bandit Algorithms within the topic of Recommender Systems.