<![CDATA[AbstractThe rapid growth of e-commerce has increased the potential for the emergence of fake reviews that can mislead consumers and reduce the credibility of online purchasing decisions. This study aims to evaluate the performance of several machine learning algorithms in distinguishing fake and genuine reviews, as well as to develop a purchase recommendation model that considers review authenticity. The dataset used consists of 2,644 product reviews from the Lazada platform, which were labeled using a rule-based approach, followed by text preprocessing, normalization, and feature extraction using TF-IDF. The classification methods applied include Support Vector Machine (SVM), Decision Tree, Random Forest, Naive Bayes, and C4.5. The results show that Random Forest and C4.5 achieved the highest accuracy of 99.81%, followed by Decision Tree (99.62%), SVM (98.30%), and Naive Bayes (93.01%). In addition, a purchase recommendation score was developed by combining rating, sentiment, helpfulness, and purchase status to classify products into recommended and not recommended categories. The findings indicate that most reviews identified as fake still result in positive recommendations, which may introduce bias in conventional recommendation systems. Therefore, integrating fake review detection with sentiment analysis and multi-criteria evaluation is essential to improve the reliability of recommendation systems in e-commerce platforms. AbstrakMaraknya perkembangan e-commerce meningkatkan potensi munculnya ulasan palsu yang dapat menyesatkan konsumen dan menurunkan kredibilitas dalam pengambilan keputusan pembelian secara daring. Penelitian ini bertujuan untuk mengevaluasi kinerja beberapa algoritma machine learning dalam membedakan ulasan palsu dan asli, serta mengembangkan model rekomendasi pembelian yang mempertimbangkan keaslian ulasan. Dataset yang digunakan terdiri dari 2.644 ulasan produk pada platform Lazada yang diberi label menggunakan pendekatan rule-based, kemudian melalui tahapan preprocessing teks, normalisasi, dan ekstraksi fitur menggunakan TF-IDF. Metode klasifikasi yang diterapkan meliputi Support Vector Machine, Decision Tree, Random Forest, Naive Bayes, dan C4.5. Hasil pengujian menunjukkan bahwa Random Forest dan C4.5 mencapai akurasi tertinggi sebesar 99,81%, diikuti oleh Decision Tree (99,62%), SVM (98,30%), dan Naive Bayes (93,01%). Selain itu, dikembangkan skor rekomendasi pembelian dengan menggabungkan rating, sentimen, tingkat helpful, dan status pembelian untuk mengelompokkan produk ke dalam kategori direkomendasikan dan tidak direkomendasikan. Temuan menunjukkan bahwa sebagian besar ulasan yang terdeteksi sebagai palsu tetap menghasilkan rekomendasi positif, sehingga berpotensi menimbulkan bias pada sistem rekomendasi konvensional. Oleh karena itu, integrasi deteksi ulasan palsu dengan analisis sentimen serta penilaian multi-kriteria menjadi penting untuk meningkatkan keandalan sistem rekomendasi pada platform e-commerce. ]]>
