Beyond Hybrids: Superior Recommendation System via Probabilistic Similarity Measures and Model Simplification

Sparsity and scalability remain recurring problems in Memory-Based (MB) Collaborative Filtering (CF).Hybrid Memory-Based (HMB) approaches were originally introduced as a way to combine User-Based (UB) and Item-Based (IB) models, with the expectation that the weaknesses of one could be offset by the other.In this work, we revisit that assumption within the HMB framework by evaluating three probabilistic similarity measures: Mutual Information (MI), Adjusted Mutual Information (AMI), and Kullback-Leibler divergence (KL).Across experiments on the MovieLens 100K and 1M datasets, we find that the optimal fusion parameter in many cases falls toward boundary values (0.0 or 1.0).In practice, this means that the hybrid setup simplifies into a standard MB model rather than benefiting from a mixed UB-IB contribution.MI-based variants obtain the strongest results, with NDCG@20 scores of 0.0366 and 0.1356 on MovieLens 100K and 1M, respectively.These values are higher than those of the baseline methods, i.e., HMB_PCC (0.0209, 0.0609), HMB_BC (0.0191, 0.0535), and the UB_Siamese model (0.0167, 0.0371), with statistically significant differences at 𝑝 < 0.05.These findings suggest that when a stronger, non-linear similarity measure such as MI is used, the benefit of linear hybrid fusion becomes limited.Rather than relying on increasingly complex hybrid architectures, our results indicate that improving the expressiveness of the underlying similarity measure can already lead to a simpler model that is both more accurate and computationally more efficient.