Recommender systems have been utilized in several e-commerce applications. There are three types of recommender systems: content based filtering, collaborative filtering, and hybrid recommender systems. In this paper, two types of collaborative filtering techniques are evaluated using the Movielens dataset, which contains 1 million ratings. These two types are matrix factorization and user based collaborative filtering with cosine similarity function. The evaluation of the two types is based on the Root Mean Square Error (RMSE) of the complete dataset and different partitions of the complete dataset. The partitions are determined by age, genre, or date of rating. For both types, the results show that the RMSE of the complete dataset is less than that of each partition. Also, in this thesis, we introduce a new hybrid technique which integrates age, genre, and date into the definition of cosine similarity function. The new technique is evaluated using two Movielens datasets of different sizes: 100,000 ratings and 1 million ratings. For both datasets, the evaluation results show that the RMSE of the new hybrid technique is less than that of the user based collaborative filtering with traditional cosine function. For the dataset containing 100,000 ratings, the evaluation results show that the RMSE of the new technique is lower than that of matrix factorization for small training sets and higher for large training sets.
Paper
Full text
Implementation and Evaluation of Movie Recommender Systems Using Collaborative Filtering
Semantic Scholar · Computer Science · 2021
Abstract
Recommender systems have been utilized in several e-commerce applications. There are three types of recommender systems: content based filtering, collaborative filtering, and hybrid recommender systems. In this paper, two types of collaborative filtering techniques are evaluated using the Movielens dataset, which contains 1 million ratings. These two types are matrix factorization and user based collaborative filtering with cosine similarity function. The evaluation of the two types is based on the Root Mean Square Error (RMSE) of the complete dataset and different partitions of the complete dataset. The partitions are determined by age, genre, or date of rating. For both types, the results show that the RMSE of the complete dataset is less than that of each partition. Also, in this thesis, we introduce a new hybrid technique which integrates age, genre, and date into the definition of cosine similarity function. The new technique is evaluated using two Movielens datasets of different sizes: 100,000 ratings and 1 million ratings. For both datasets, the evaluation results show that the RMSE of the new hybrid technique is less than that of the user based collaborative filtering with traditional cosine function. For the dataset containing 100,000 ratings, the evaluation results show that the RMSE of the new technique is lower than that of matrix factorization for small training sets and higher for large training sets.