Descriptive Matrix Factorization for Recommender Systems in Digital Interactive Entertainment – Københavns Universitet

Descriptive Matrix Factorization for Recommender Systems in Digital Interactive Entertainment

You are cordially invited to Anamaria Todor's MSc thesis defense

Recommender systems can be effectively used in the field of digital interactive entertainment and game development to improve the quality of gameplay. In the classical sense, recommenders are already used in the context of games by review websites and distribution platforms to suggest players other games they could try based on what they already like.

However, games themselves rarely employ them. This is firstly due to the game developers' reluctance to take player preferences into account and, secondly, due to the fact that the data models recommender systems rely on take considerable time to compute, making them difficult to use in real-time applications such as video games.

This thesis tries to address the mentioned disadvantage by adapting Simplex Volume Maximization Non-Negative Matrix Factorization (SiVM-NMF), a relatively new, easily scalable and linear time algorithm previously used in the field of archetypal analysis by Thurau, Kersting, Wahabzada, and Bauckhage in 2002, to the field of recommender systems. The algorithm is implemented and evaluated from the scalability, performance, prediction accuracy and interpretability points of view in a structured framework. It is compared with a modified Singular Value Decomposition (SVD) method which has more or less become the standard in model-based recommending techniques since the 2006 Netflix challenge. Empirical findings suggest that while SiVM-NMF scales better than SVD to large datasets, it cannot be used with recommender systems in its current form, most likely because of its geometric interpretation incompatibility with sparse data.

Further research is needed to seamlessly adapt SiVM-NMF to the recommender systems domain, but a functional SiVM-NMF recommender would be an important step closer to actively using online recommender systems in the field of digital interactive entertainment, in order to engage players more closely with games by providing content that constantly challenges them based on their skill level, yet never overwhelms. A SiVM-NMF recommender could be the main factor that differentiates a successful game from a mediocre one, especially in the mobile market space.

Supervisor: Christian Igel, Christian Thurau

Censor: Peter Dolog, AAU