Running summarizing tools on a new article

Abstract: Log anomaly detection is an essential component in maintaining the reliability of industrial and production network systems. While many unsupervised LSTM-based methods have been evaluated on open datasets, their behavior on logs collected from actual network systems remains unclear. In this paper, we conduct a comparative evaluation of five unsupervised log anomaly detection models—DeepLog, DeepLog with Word2Vec embeddings, DeepLog with BERT embeddings, LogAnomaly, and LogAnomaly with BERT embeddings—using both public datasets (HDFS and OpenStack) and three proprietary datasets collected from actual network systems. The results show that detection performance depends more strongly on dataset characteristics than on architectural differences or embedding options.