Top-N Recommendation (Implicit Feedback)

Abstract

The rapid growth of digital platforms has led to a massive increase in user-item interaction data, most of which is recorded as implicit
feedback such as purchase history and reading duration. Unlike explicit ratings, implicit feedback presents fundamental challenges
including extreme sparsity and ambiguity, as non-interaction does not necessarily indicate disinterest. This study develops a Top-N
book recommendation system by implementing Item-Based Collaborative Filtering (IBCF) as a baseline and Alternating Least
Squares (ALS) as the primary model, addressing the challenge of a highly sparse dataset (13,876 users × 123,069 items). The
methodology involves transforming raw interaction data into Compressed Sparse Row (CSR) format, performing systematic
hyperparameter tuning on the ALS model, and implementing rigorous data preprocessing techniques to optimize performance.
The experimental results demonstrate a progressive improvement trajectory: the baseline IBCF model achieved a Mean Average
Precision at 10 (MAP@10) score of 0.0186, the initial ALS model improved to 0.0243, hyperparameter tuning further increased
performance to 0.0257, while the integration of systematic data preprocessing (K-Core filtering) with advanced hyperparameter
tuning ultimately yielded a score of 0.0622 representing a 234% improvement over the baseline. Qualitative analysis reveals that
while IBCF produces monotonous and homogeneous recommendations limited to similar publishers and series, ALS
provides substantially more diverse and exploratory results by capturing latent behavioral patterns across multiple genres and
thematic categories. This study concludes that the optimized ALS model, combining matrix factorization with rigorous data
engineering, is significantly more effective than in handling sparse implicit feedback, delivering superior rankingaccuracy, enhanced recommendation diversity, and improved user discovery experience.