ADA-TGN

ADA-TGN (Temporal Graph Networks with Adaptive Memory Protection) is an anomaly detection and root cause analysis framework for microservice systems. It extends TGN with a Dual-Stage Soft-Gating mechanism that prevents memory contamination by anomalous data.

For full usage, see the ADA-TGN repository.

Files

File	Description
ada_tgn.pt	Unified model — single checkpoint trained on all 60 experiments
{experiment_id}_rep{rep}.pt	Per-experiment models — 60 individual checkpoints, e.g. checkoutservice_cpu_rep1.pt

How to Use

git clone https://github.com/cyb3rlab/ADA-TGN.git && cd ADA-TGN
bash setup_env.sh
bash run_evaluate.sh          # unified model (default)
bash run_evaluate.sh --model_type per_experiment

The script evaluates two tasks across all 60 experiments and prints a summary to stdout:

• Anomaly Detection (AD): Precision, Recall, F1 — using a per-(node, metric-group) adaptive threshold (dSPOT)
• Root Cause Analysis (RCA): AC@1, AC@3, Avg@5 broken down by fault type — services ranked by IQR-normalised reconstruction error

Training Details

Trained on normal period data only (72 windows per experiment, 10-second snapshots). Evaluated on RE2-OB — 60 experiments (5 services × 4 fault types × 3 replications). See norun9/re2ob-pyg for dataset details.

Library / Hardware	Version
Python	3.9+
PyTorch	≥ 2.0
PyTorch Geometric	≥ 2.3
Hardware	NVIDIA A40, CUDA 12.1

License

Apache 2.0