Analisis Efektivitas Pembobotan TF-IDF dan Metode Cosine Similarity Untuk Deteksi Plagiarisme

Rido Anugrah (Universitas Pamulang, Banten)
Atang Susila (Universitas Pamulang, Banten)

TIN: TERAPAN INFORMATIKA NUSANTARA
Vol 7 No 5 (2026): October 2026

Plagiarism in academic documents undermines scientific integrity and significantly degrades the quality of education. This study aims to analyze the effectiveness of combining Term Frequency-Inverse Document Frequency (TF-IDF) statistical weighting with the Cosine Similarity distance metric for detecting the degree of global semantic similarity at the document level. The computational experiments were conducted using the Python programming language, supported by the Natural Language Toolkit (NLTK) and Scikit-Learn libraries. The test dataset consists of an internationally standardized secondary dataset from the PAN at CLEF 2011 corpus (a subset of the PAN-PC-2011 corpus), comprising 500 pairs of suspected (query) documents and their corresponding source documents. As the test dataset contains only positive pairs (instances of plagiarism) and no negative pairs (non-plagiarized instances), the binary classification evaluation focuses on True Positives (TP), False Negatives (FN), positive-class accuracy, and recall. The primary contribution of this research is the provision of empirical evidence regarding an optimal threshold (2%) and a comprehensive analysis of the impact of "vector dilution" in full-length documents. The highest recall achieved in the tested scenarios was 94% (470 TP out of 500 pairs) at a 2% threshold. Recall decreased as the threshold increased, reaching 0% at thresholds of 80% and 90%. These results indicate that the threshold significantly influences detection sensitivity for positive instances; however, it does not yet reflect the system's ability to correctly reject non-plagiarized documents. Consequently, this method is best positioned as an initial screening tool that requires further evaluation using both positive and negative pairs.