Umeå University's logo

Recommendation systems are primarily used in e-commerce and retail to guide the user in a vast space of available items by providing personalized recommendations that fit the user's interests and need. Numerous types of recommendation systems have been introduced over the years. The most recent development in the field is the sequential recommendation system. Sequential recommenders account for the order in which the user has interacted with items to infer the user's intent, allowing them to provide recommendations accordingly. The data analytic company Siftlab AB has already developed such a recommendation system; however, its application has been limited to transaction data(data depicting only purchases). As a result, the model cannot take advantage of the predictive values of different event types. This thesis introduces a weighted multi-type technique that allows Siftlab's recommendation model to leverage page views alongside purchases in data from an interior design store. We also developed tools and techniques, such as correlation and angle separation analysis, to enhance our examination of user-item behavior. Our research findings indicate that including page view events in training hurts recall, while their inclusion in the prediction stage yields slight improvements. We discovered a rapid decline in correlation between purchases and page views as we considered page views occurring relatively further back in time. Performing a time-based correlation analysis, it became evident that there is a robust time dependency between purchases and page views. Utilizing this time dependency, we enforced a time-dependent threshold on the page views we included in the prediction stage to eliminate irrelevant page view events, further enhancing the model's predictions. We also captured seasonalities phenomena distinctive for an interior design store. Although the result of this work might only be valid for a single data set, we anticipate our work to be the first step in the right direction since the technique we introduce here can be effortlessly adapted to analyze other event types in other data, thus uncovering patterns that can further elevate the model's performance.