Comparison of basic recommendation methods for Czech news articles

The common contemporary recommendation methods were assessed for recommending relevant news articles in the Czech language. We used a TF-IDF vectorization with cosine similarity for content-based recommendations and singular value decomposition for collaborative filtering. The simplest model was the popularity model based on the vote counts of other users. While the best days of expert systems seem to be gone in the era of more sophisticated neural network models, we also wanted to investigate whether expert systems can still show some benefits in providing relatively simple and explainable methods for an ensemble of recommendations. We also used a Bayesian-inspired approach for the hybrid method ensemble. This work aimed to test the fundamental recommendation techniques on smaller datasets with non-English language using three predictors: user interactions by like button, click-throughs (and article text itself). Despite the recent popularity of the neural network models, we found that simple models outperformed the models with neural networks on our dataset.