PENGEMBANGAN WEBSITE REKOMENDASI JUDUL DAN DRAF ARTIKEL KEGIATAN KOMUNIO PALEMBANG MENGGUNAKAN TF-IDF, NAIVE BAYES DAN COSINE SIMILARITY

DEVELOPMENT OF A WEB-BASED RECOMMENDATION SYSTEM FOR ACTIVITY ARTICLE TITLES AND DRAFTS AT KOMUNIO PALEMBANG USING TF-IDF, NAIVE BAYES, AND COSINE SIMILARITY

Abstract

KOMUNIO Palembang telah memiliki arsip artikel digital, tetapi sistem publikasi yang berjalan belum menyediakan dukungan editorial untuk mengubah deskripsi kegiatan baru menjadi prediksi kategori, artikel rujukan, rekomendasi judul, dan draf awal dalam satu alur. Penelitian ini mengembangkan dan mengevaluasi pipeline editorial berbasis TF-IDF, Multinomial Naive Bayes, dan Cosine Similarity pada 2.991 artikel yang dipetakan ke empat kategori. Data dibagi 80:20 dengan random state 42 dan oversampling diterapkan hanya pada data latih. Model final memperoleh accuracy 83,97%, macro precision 81,97%, macro recall 82,11%, dan macro F1-score 81,95%. Performa antarkelas tidak seragam: Liturgi mencapai F1-score 99,6%, sedangkan Berita dan Kegiatan Gereja hanya 57,0%. Sebanyak 51 dari 114 artikel Berita dan Kegiatan Gereja (44,7%) diprediksi sebagai Orang Muda dan Anak Remaja, yang menunjukkan tumpang tindih kosakata dan batas label. Dibanding majority-class baseline, accuracy meningkat dari 38,06% menjadi 83,97%. Pada pengujian rekomendasi, Cosine Similarity tertinggi berada sekitar 22–26%, sehingga skor ditafsirkan sebagai kedekatan leksikal relatif untuk pemeringkatan rujukan, bukan kesamaan semantik mutlak. Kontribusi penelitian terletak pada integrasi klasifikasi dan pencarian rujukan ringan untuk bantuan editorial berbahasa Indonesia, disertai analisis kesalahan kelas dan batas interpretasi rekomendasi.

KOMUNIO Palembang maintains a digital article archive, but the existing publication system does not provide an integrated editorial workflow that transforms a new activity description into category prediction, reference-article retrieval, title recommendation, and an initial draft. This study develops and evaluates an editorial pipeline based on TF-IDF, Multinomial Naive Bayes, and Cosine Similarity using 2,991 articles mapped into four categories. The data were split 80:20 with a random state of 42, and oversampling was applied only to the training data. The final model achieved 83.97% accuracy, 81.97% macro precision, 82.11% macro recall, and 81.95% macro F1-score. Class performance was heterogeneous: Liturgy reached an F1-score of 99.6%, whereas Church News and Activities reached only 57.0%. Fifty-one of 114 Church News and Activities articles (44.7%) were predicted as Youth and Children, indicating lexical overlap and non-exclusive class boundaries. Compared with a majority-class baseline, accuracy increased from 38.06% to 83.97%. The highest Cosine Similarity scores in the recommendation test were approximately 22–26%; therefore, the scores are interpreted as relative lexical proximity for ranking references rather than absolute semantic equivalence. The contribution is an integrated, lightweight Indonesian editorial-assistance workflow combining classification and content retrieval with class-level error analysis and a conservative interpretation of recommendation scores.