Personalized Risk Aware Recommender Systems

Abstract

Recommender systems play a vital role in many sectors. Ease of use and availability increase the importance of these systems. E-commerce recommender systems provide recommendation lists based on user preferences and interest. The goal of e-commerce recommender systems from a user viewpoint is saving time and effort, and from a business viewpoint is increasing sales. Providing recommendations based only on user preferences can be risky because users’ acceptance of recommendations may differ based on their situations. It is important to consider risks before providing recommendations to reduce the chance of irritating users with different items in the recommendation lists. Many studies have shown that the user intention to make a purchase can be employed to understand users’ behaviour and provide different recommendations based on their intention. To decrease risks that may negatively affect system goals, assessing risks and controlling them are crucial steps. Risk assessment has two approaches: model driven and data driven. In this thesis, we consider the risk of the user not purchasing before providing a recommendation. The model driven approach is performed to obtain a better understanding behind the prediction. The disadvantage of the model driven approach is that the prediction does not consider a change in user behaviour. In order to address this shortcoming, we propose a data driven approach. Experimentally, we show the proposed data driven model outperforms two state-of-the-art models on two publicly available data sets in regards to AUC of ROC area by 13% and 6%, respectively. Because different users can accept different levels of diversity in their recommendation lists, we also propose a risk assessment algorithm that predicts how many categories users can accept in the recommendation list based on their interaction with the system. Additionally, we propose a recommender system framework which includes the risk assessments of user purchase intention and user interaction with the system, the proposed framework can provide personalized recommendations with increased diversity in the recommendation lists. Experimentally, we show the resulting framework has higher coverage of items when compared with item-to-item collaborative filtering recommender systems and popularity-based recommender systems on two publicly available data sets, reaching 89.9% and 70.2%, respectively.