<![CDATA[Abstrak - Penelitian ini membandingkan kinerja Support Vector Machine, Random Forest, dan XGBoost dengan pendekatan rekayasa fitur untuk mengklasifikasikan produk terlaris pada platform Tokopedia. Data diperoleh melalui web scraping terhadap produk elektronik dari enam merek, yaitu Baseus, Fantech, Logitech, Sony, Ugreen, dan Xiaomi, dengan total awal 1.147 baris. Tahap prapemrosesan mencakup validasi tipe data, penanganan nilai hilang dan duplikasi, pembentukan label, penghapusan zona ambigu, transformasi logaritmik, pengodean merek, serta pencegahan kebocoran target. Produk dengan jumlah terjual sekurang-kurangnya 500 diberi label Terlaris, produk dengan jumlah terjual kurang dari 100 diberi label Tidak Terlaris, sedangkan data dengan penjualan 100-499 dikeluarkan. Dataset akhir berjumlah 884 data, terdiri atas 587 produk Terlaris dan 297 produk Tidak Terlaris. Data dibagi secara stratifikasi dengan rasio 70:30 menjadi 618 data latih dan 266 data uji. Rekayasa fitur menghasilkan 24 fitur yang mencakup harga, diskon, rating, kualitas, posisi harga terhadap merek, segmentasi harga, dan interaksi antarfitur. Hasil pengujian menunjukkan bahwa Random Forest dengan fitur dasar dan rekayasa fitur pada ambang keputusan 0,53 memberikan kinerja terbaik, yaitu akurasi 86,84%, precision tertimbang 0,8675, recall tertimbang 0,8684, F1-score tertimbang 0,8656, dan F1-score makro 0,8463. Model tersebut mengungguli XGBoost dan Support Vector Machine. Fitur yang paling dominan adalah rating_kuadrat, Rating, nilai_produk_score, kompetitif_score, rating_x_harga_rel, dan brand_encoded. Temuan ini menunjukkan bahwa rekayasa fitur memberi peningkatan paling besar pada Random Forest dan membantu model menangkap hubungan nonlinier antara kualitas, harga, diskon, dan karakteristik merek. Kata kunci : Rekayasa Fitur; Pembelajaran Mesin; Produk Terlaris; Abstract - This study compares the performance of Support Vector Machine, Random Forest, and XGBoost using a feature engineering approach to classify best-selling products on Tokopedia. The data were collected through web scraping from electronic products of six brands, namely Baseus, Fantech, Logitech, Sony, Ugreen, and Xiaomi, producing 1,147 initial records. Preprocessing included data-type validation, missing value and duplicate handling, label construction, ambiguous-zone exclusion, logarithmic transformation, brand encoding, and target-leakage prevention. Products with at least 500 units sold were labelled Best Selling, products with fewer than 100 units sold were labelled Non-Best Selling, while records with 100-499 units sold were excluded. The final dataset contained 884 records, consisting of 587 Best-Selling and 297 Non-Best-Selling products. A stratified 70:30 split produced 618 training and 266 testing records. Feature engineering generated 24 predictors representing prices, discounts, ratings, quality, brand-relative prices, price tiers, and feature interactions. The results show that Random Forest with basic and engineered features at a decision threshold of 0.53 achieved the best performance, with 86.84% accuracy, 0.8675 weighted precision, 0.8684 weighted recall, 0.8656 weighted F1-score, and 0.8463 macro F1-score. It outperformed XGBoost and Support Vector Machine. The most influential features were rating_squared, Rating, product_value_score, competitiveness_score, rating_by_relative_price, and brand_encoded. These findings demonstrate that feature engineering provides the largest improvement for Random Forest and helps capture nonlinear relationships among product quality, price, discount, and brand characteristics. Keywords: Feature Engineering; Machine Learning; Best-Selling Products;]]>
