Transformer Models Outperform Traditional Algorithms in Log Anomaly Detection

This study evaluates a Transformer-based model for log anomaly detection across multiple datasets, comparing it with simpler baselines like KNN and Decision Trees. Results show that while the model achieves competitive F1 scores, its performance remains stable even without sequential or temporal encoding—suggesting that semantic information within log events is the strongest indicator of anomalies. In fact, adding temporal or positional data slightly reduced accuracy due to information noise. The findings highlight that Transformers can effectively detect anomalies without relying on time-based patterns, challenging conventional assumptions in sequential data modeling.