Verification of Generalizability in Software Log Anomaly Detection Models

With recent rapid technological advances, the automatic analysis of software logs has received particular attention. Currently, there is much research on use Deep Learning in field log anomaly detection, and they have reported high accuracy more than 0.9 f1-score. On other hand, are reports that it not been used development. We conducted a generalized evaluation against representative models for detection to elucidate cause this problem. Five were subject: four (two supervised two unsupervised) our proposed Neocortical Algorithm (supervised). commonly Blue Gene/L supercomputer log(BGL) dataset. The learning curves cross-validation showed tendency toward overfitting all models. In addition, survey frequency patterns confirmed need diverse dataset, as many series specific logs.