A Weighted Hybrid of TF-IDF and Winnowing for Content-Based Scientific Article Recommendation with Citation-Based Ground Truth

Abstract

Purpose: The rapid growth of digital scientific publications makes it increasingly difficult for researchers to identify relevant literature. This study develops a content-based filtering recommendation system for scientific articles that fuses two complementary similarity channels: TF-IDF with Cosine Similarity for global semantic weighting and Winnowing with Jaccard Similarity for local structural similarity.

Methods: Query and corpus articles pass through a preprocessing pipeline of case folding, filtering, tokenizing, stopword removal, and Sastrawi stemming. The corpus contains 2,059 scholarly articles from the host institution. A systematic grid search over K-gram, window size, and TF-IDF/Winnowing weight combinations was evaluated using Precision@K, Recall@K, F1-Score@K, MAP, and NDCG, with citation-based ground truth derived from the harvested metadata.

Result: The optimal configuration was K-gram = 9, window size = 6, with 60% TF-IDF and 40% Winnowing, reaching a peak aggregate score of 0.2609 and MAP@20 of 0.3091. The hybrid approach outperformed single-method baselines, with positive N-gain at MAP@5, MAP@10, and MAP@20.

Novelty: The work contributes a four-method hybrid recommender that combines TF-IDF-Cosine and Winnowing-Jaccard similarity with grid-search-optimized weights, validated on a local institutional corpus using citation-based ground truth as an objective and reproducible proxy for academic relevance.