Efficient Session-based Item-to-item Recommendation

OData support
Simon Gábor
Department of Automation and Applied Informatics

Many new areas of research rose in the past few years in the field of recommendation systems. Still to this day, the most widespread measure for explicit recommenders is RMSE (Root Mean Squared Error). Its biggest mistake is disregarding the order of elements, which has a great importance in real usage, since in practice, it is only possible the show the users a short, but hopefully relevant list.

After realizing this problem, not only new algorithms were created, but ranking-based evaluation (e.g.: nDCG (normalized Discounted Cumulative Gain)), and the usage of contextual elements were also considered. In the case of traditional CF (Collaborative Filtering) recommendation systems, it is assumed that the user already rated, or at least viewed multiple elements in the system. This assumption is not always true. As it is often the case, the visitors of a website may be looking for certain elements without logging in, going through “random paths” (session).

The goal of this thesis is to develop a recommendation system, that can provide recommendations to any user visiting the site. This includes both highly personalized recommendations for already existing active users, and more general recommendations for first time visitors. It also includes giving recommendations for both users, and for items.

The thesis contains an introduction of multiple recommendation models based on the latest results in the area. These models are implemented, and then compared through several simulations, evaluated with measurements that are widely used in the industry. Also examined are the effects of differently scheduled updates on these models.


Please sign in to download the files of this thesis.