LogCDM
A lightweight log anomaly detection method that combines three signals: causal Transformer surprisal, graph attention structural distance, and transition-drift chi-square divergence. It fuses them through sigmoid gating and supports per-edge interpretable alerts and real-time deployment on CPU.
Timeline (last 90 days)
LogCDM uses a small detector with only 8.1-8.8M parameters and completes scoring of a single session on CPU in 5.5-10.5 milliseconds.
LogCDM achieves an F1 score of 0.984 on the HDFS dataset.
LogCDM achieves an F1 score of 0.870 on the chronologically split BGL dataset.
LogCDM achieves an F1 score of 0.850 on the Thunderbird dataset.
Under a unified evaluation protocol, an 8.1B-parameter LLM detector's F1 score differs from LogCDM's by only 0.013-0.12.
On the BGL dataset, LogCDM outperforms all compact baselines.
The core design of LogCDM is to feed three complementary signals (next-event surprisal, graph attention embeddings with reference graph distance, and transition drift statistics) to a sigmoid gate for fusion.
The paper proves that LogCDM's transition drift statistic is exactly the frequency-weighted chi-square divergence between transition kernels.
Under a shared-covariance Gaussian model, LogCDM's sigmoid gate is equivalent to a Bayesian posterior.
Ablation experiments show that the structural signal brings F1 gains of 0.025-0.030 on BGL and Thunderbird, but contributes almost nothing on HDFS.
1 more timeline events
All Facts (11)
LogCDM uses a small detector with only 8.1-8.8M parameters and completes scoring of a single session on CPU in 5.5-10.5 milliseconds.
50%UnverifiedLogCDM achieves an F1 score of 0.984 on the HDFS dataset.
50%UnverifiedLogCDM achieves an F1 score of 0.870 on the chronologically split BGL dataset.
50%UnverifiedLogCDM achieves an F1 score of 0.850 on the Thunderbird dataset.
50%UnverifiedUnder a unified evaluation protocol, an 8.1B-parameter LLM detector's F1 score differs from LogCDM's by only 0.013-0.12.
50%UnverifiedOn the BGL dataset, LogCDM outperforms all compact baselines.
50%UnverifiedThe core design of LogCDM is to feed three complementary signals (next-event surprisal, graph attention embeddings with reference graph distance, and transition drift statistics) to a sigmoid gate for fusion.
50%UnverifiedThe paper proves that LogCDM's transition drift statistic is exactly the frequency-weighted chi-square divergence between transition kernels.
50%UnverifiedUnder a shared-covariance Gaussian model, LogCDM's sigmoid gate is equivalent to a Bayesian posterior.
50%UnverifiedAblation experiments show that the structural signal brings F1 gains of 0.025-0.030 on BGL and Thunderbird, but contributes almost nothing on HDFS.
50%UnverifiedLogCDM's per-edge decomposition lets every alert be traced back to a specific transition edge, thereby providing interpretability.
50%