Preserving Accuracy while Increasing Novelty: Rank-Aware Evaluation of a Locally Fine-Tuned Hybrid Recommender

Abstract

Balancing accuracy and novelty remains a fundamental challenge in modern recommender systems. We present a novelty-aware hybrid recommender that linearly combines ItemKNN, content-based similarity, popularity signals, and lightweight SVD factors. We further introduce a user-specific novelty term derived from popularity and recency to encourage discovery. To avoid overfitting and maintain interpretability, we adopt local fine-tuning around a near-optimal trade-off rather than global re-optimization. First, we conduct a before and after evaluation using rank-aware metrics (NDCG@10, MRR@10, Precision@10, Recall@10). Then, we measure list-level properties such as novelty with respect to popularity, intra-list diversity by genre dissimilarity, and catalog coverage. Finally, we present a separate experiment using Maximal Marginal Relevance (MMR) re-ranking applied to ItemKNN to situate our contributions within classical diversification. On MovieLens-100K dataset, the locally fine-tuned hybrid preserves ranking accuracy while improving novelty and coverage: relative to the best trade-off baseline, NDCG@10 differs by less than 0.2% absolute on validation and test, while novelty increases modestly and coverage rises by approximately 1.7% on the test split. The MMR variant increases intra-list diversity by approximately 2-3 percentage points on validation with only a slight reduction in NDCG, illustrating a well-controlled accuracy diversity trade-off that complements the hybrid approach. These findings show that carefully designed novelty terms and restrained local fine-tuning can yield measurable gains in novelty and coverage without sacrificing ranking quality, while classical diversification provides an additional, orthogonal mechanism to further increase intra-list diversity. We provide transparent, metrics-based evaluation and reporting of novelty, diversity, and catalog coverage.