Improved probabilistic matrix factorization model for sparse datasets /

Ar, Yılmaz
The amount of information on the World Wide Web has increased significantly owing to advancing web and information technologies. This has made it difficult for users to obtain relevant and useful information thus there is a need for information filtering. Recommender Systems (RS) have emerged as a technique to overcome the problem. Collaborative Filtering (CF) that is one of the widely used RS approaches aims to predict users’ preference concerning an item. The main idea behind CF is the users who agreed in the past will agree in the future. The Probabilistic Matrix Factorization (PMF) is the preferred CF technique in the literature due to its high accuracy and scalability. This thesis demonstrates the importance of the initialization techniques for the user and the item latent vectors in the PMF algorithm with real and synthetic datasets and proposes five different initialization techniques. The suggested approaches produce better results in comparison with the state-of-the-art techniques in particularly very sparse datasets.