Improving co-SVD for cold-start recommendations using sparsity reduction

Recommender systems are highly dependent on the users' or items' historical data. The completeness of the data determines the performance of the models, especially for models based on the collaborative filtering (CF) technique. Under cold-start situations, where there are limited relevant historical data on the users' or items' information, producing accurate recommendations are challenging. We propose the use of the implicit Alternating Least Square (iALS) method to predict users' preferences and impute it into the matrix co-factorization algorithm, co-SVD. The proposed approach aims to alleviate the cold-start problem that is most evident in large datasets that have high sparsity. In addition, we included the results for two cold-start situations, cold-start user and cold-start item (long-tail), using our hybrid co-SVD with artificial ratings imputation. The F1 score of the top-5 recommendations generated by the proposed approach improved from 25.08% to 30.09% under the cold-start user situation. With the long-tail item situation, the proposed approach improved from 20.8% to 23.19%. The proposed approach is method-agnostic, and other CF-based models can benefit from this imputation method.