|
|
Probabilistic Models for Recommender Systems
Ahmadli, Aydin ; Vomlelová, Marta (advisor) ; Peška, Ladislav (referee)
Recommender systems are software tools and techniques providing recommendations to users based on their needs. Today, popular e-commerce sites widely use recommender systems to recommend product items, articles, books, music, etc. In this thesis, we discuss various probabilistic models for recommender systems, and put the most focus on implementation of hybrid and interpretable probabilistic content-based collaborative filtering model, called Collaborative Topic model for Poisson distributed ratings (CTMP) augmented with Bernoulli randomness for Online Maximum a Posteriori Estimation (BOPE). Resulting model outperforms the previously existing models significantly with its main competency being in commercial product recommendations. It is a fast, scalable, and efficient in ill-posed cases, including short text and sparse data. The model is trained and tested on well-known MovieLens 20M and NETFLIX datasets, and empirical evaluations such as recall, precision, sparsity and topic interpretations are promising.
|
guest ::
