A Study on Data Reconstruction and Model Parameter Optimization for Implementation of Anomaly Detection System Based on User Behavior Analysis

In this paper, we implemented an anomaly detection system based on user behavior analysis to effectively detect user anomalies in a specific domain. For this purpose, we performed EDA on User Behavior Data, defined feature factors for user behavior features in feature engineering, and proposed a method for combining feature factors. We performed preprocessing and vectorization on the session data to define user behavior patterns as 'Session' and provide them as input to the model. The vectorized session data was pre-trained on BERT Model Architecture using only normal session data. We performed an anomaly detection performance evaluation after fine-tuning using normal and abnormal session data. As a result of the performance evaluation, BERT-Medium-uncased model performed well with an Accuracy of 0.9630 and an F1-Score of 0.9628, and the overall performance was balanced. As a result, we confirmed that by utilizing EDA and feature engineering for data, we can effectively perform pre-training and fine-tuning and implement a high-performance anomaly detection system.