A comparison of different types of recommender systems

In this thesis three different types of reccommender systems were compared: baseline predictor, collaborative filtering, content-based recommeder. We looked at what recommender systems are and what they are good for. All three methods were addressed and also some others. Pros anc cons of collaborative filtering and content-based recommenders were addressed. The evaluation of recommender systems was addressed. We took a closer look at collaborative filtering. User-based kNN recommendation was introduced. The term neigbourhood was introduced. The alghoritms for similarity calculation and predicted rating calculation were introduced. We took a closer look at content based recommender. A high level architecture of content-based systems was introduced. Two functions for product similarity calculation were introduced. We took a closer look at evaluation of the recommender systems. The leave-one-out method was introduced and also two measures for error calculation MAE and RMSE. The data was acquired from "Freebase" service. The values of attributes were changed and some of the attributes were eliminated. Colrec script was adapted and used for analyzing the user-based collaborative filtering. The optimization of recommendation technique was performed by using the threshold. Content-based technique was optimized by using different finctions for calculating the similarity between products. Every attribute was weighted with a different weight. The alghorithms for calculating the similarity between products and predicting ratings were introduced. Content-based technique and baseline predictor were implemented by using Python programming language. For ratings storage the two-dimensional array was used, which is included in numpy library