Hybrid Enhanced Neighbourhood and Latent Factor Model for Cold-Start Recommendations

Abstract

Among different recommendation strategies, collaborative filtering remains a commonly utilized method for generating personalized suggestions. The traditional collaborative algorithms face performance declines due to the sparse rating of data and the item cold-start problem. To overcome these challenges, this paper introduces a novel hybrid model called HCE-KNNCF (Hybrid Cognition-Enabled K-Nearest Neighbor Collaborative Filtering). The proposed model generates predicted rating by a combination of SVD-based matrix factorization and the enhanced KNN model using a weighted hybrid approach. The cognition-based KNN ensures that only relevant neighbors contribute to the rating prediction phase and the SVD-based collaborative approach is employed to model latent user-item relationships, thereby mitigating the effects of data sparsity. Experimental evaluations on the MovieLens 100 K, MovieLens 1 M, and Book-Crossing datasets show that HCE-KNNCF achieves improved prediction accuracy compared with most traditional and hybrid benchmark models. The model achieves the best MAE and RMSE results on the MovieLens 100 K and Book-Crossing datasets, while maintaining competitive performance on MovieLens 1 M. In cold-start scenarios, HCE-KNNCF demonstrate that a small increase in MAE and RMSE, indicating that the proposed approach remains stable when interaction data are limited.