Using Topic Models in Content-Based News Recommender Systems

We study content-based recommendation of Finnish news in a system with a very small group of users. We compare three standard methods, Naïve Bayes (NB), K-Nearest Neighbor (kNN) Regression and Regulairized Linear Regression in a novel online simulation setting and in a coldstart simulation. We also apply Latent Dirichlet Allocation (LDA) on the large corpus of news and compare the learned features to those found by Singular Value Decomposition (SVD). Our results indicate that Naïve Bayes is the worst of the three models. K-Nearest Neighbor performs consistently well across input features. Regularized Linear Regression performs generally worse than kNN, but reaches similar performance as kNN with some features. Regularized Linear Regression gains statistically significant improvements over the word-features with LDA both on the full data set and in the cold-start simulation. In the cold-start simulation we find that LDA gives statistically significant improvements for all the methods.