Matrix factorization models are becoming increasingly popular in the field of collaborative filtering recommender systems. Recent developments in this area of research use a penalization method, such as the L2 penalty, to restrict overfitting and reduce sparseness. We propose an alternative way of regularizing matrix factorization for recommender systems, i.e., the elastic net. A compromise between the L1 and L2 penalties, the elastic net can be implemented in any coefficient estimation scenario. We evaluate the performance of our model on real-world data, namely the MovieLens 100K dataset. Comparison with two more restrictive models shows that our proposed regularization provides superior accuracy in predictions, as measured by the mean absolute error. Moreover, prediction errors for individual users occur less often, and we are able to accurately predict 95.02% of the ratings with an error of at most two points from the real ratings, given on a scale from 1 to 5. Finally, sensitivity analysis shows the stability of the proposed solution.

Additional Metadata
Keywords Elastic net, Movielens, Recommender systems, Regularization
Persistent URL dx.doi.org/10.1145/3341105.3373847, hdl.handle.net/1765/126166
Conference 35th Annual ACM Symposium on Applied Computing, SAC 2020
Citation
Mitroi, B. (Bianca), & Frasincar, F. (2020). An elastic net regularized matrix factorization technique for recommender systems. In Proceedings of the ACM Symposium on Applied Computing (pp. 2184–2192). doi:10.1145/3341105.3373847