Deteksi Anomali Penjualan Menggunakan Algoritma Isolation Forest di PT Multi Nabati Bitung

Abstract

Sales transaction activities in large companies generate substantial amounts of multidimensional data that require effective monitoring to identify unusual transaction patterns. Manual supervision becomes less efficient when transaction volumes increase and historical data do not contain labels indicating normal or anomalous transactions. This study aims to apply an unsupervised machine learning approach using Isolation Forest to detect anomalies in sales transaction data at PT Multi Nabati Bitung and to implement the detection results in a web-based dashboard. The research used historical sales transaction data obtained from the company's database or Enterprise Resource Planning system in Excel/CSV format. The research stages consisted of data collection, preprocessing, exploratory analysis and visualization, feature construction, data normalization, Isolation Forest modeling, qualitative evaluation, and web-based implementation using Flask. The model used six transaction features, namely quantity sold, unit price, total sales, month, day of week, and average price. StandardScaler was applied before model training, while the Isolation Forest model used 100 isolation trees and a contamination value of 0.05. From 1,000 historical transactions, the model identified 50 transactions or 5% as anomalies and 950 transactions or 95% as normal. The detected anomalies generally showed transaction characteristics that differed significantly from the majority pattern, including unusually high sales volume, prices outside the common range, and unusual combinations of transaction attributes. The results indicate that Isolation Forest can support automated preliminary screening of sales transactions. Further research should use verified labeled data to enable quantitative evaluation using classification metrics.

GS Cited Analysis