An enhancement of item-based collaborative filtering algorithm applied in book recommendations

ABSTRACT: This study focused on enhancing the Item-Based Collaborative Filtering Algorithm tom improve similarly calculation and recommendation generation. The study addressed key challenges of the algorithm: (1) cold start for new items which restricts item similarly, (2) susceptibility to data sparsity and skewed distribution which compromises item representations, and (3) failure to consider the weight of the interaction date and the interaction progress in the calculation. To address these challenges, the following enhancements were implemented: (1) temporary boost for new items to establish item similarity, (2) virtual interactions based on item attributes association and matrix factorization to represent less interacted items and highly interacted items on the same level, and (3) item interaction monitoring progress and time-based weight – depending on the interaction’s aging days. Testing was applied in book recommendation with sequences of 10 users to 20 items, random interactions, item genres, authors, series, dates the item were added to the system, interaction progress, and date of interaction, showed that the enhanced algorithm addressed the key challenges. (1) Temporary boost for new items with similarity scores averaging 49.17% higher (0.270 ≠ 0.050 vs. 0.181 ± 0.019), establishing item similarity for new items, (2) Virtual interactions and matrix factorization resulted in a 2.89% decrease in interaction data (0.369 ≠0.041 vs. 0.380 ±0.067) and a 21.42% reduction in similarity scores (0.202 ≠0.041 vs. 0.257 ±0.017), improving item representations, (3) The consideration of interaction date and progress decreased interaction data by 42.89% (0.217 ≠ 0.050 vs 0.380 ± 0.067), demonstrating the varying weights of the interaction date and the interaction progress. The Enhanced Item-Based Collaborative Filtering Algorithm provides a more reliable assessment of item interaction, enabling a better generation of item recommendations.