ScatterAD: Temporal-Topological Scattering Mechanism for Time Series Anomaly Detection

ScatterAD leverages representation scattering as an inductive signal to jointly model temporal and topological patterns for effective multivariate time series anomaly detection.

Key Features

1. Dual Encoder Architecture:

   • Topological Encoder: Captures graph-structured scattered features
   • Temporal Encoder: Constrains excessive scattering by minimizing mean squared error between adjacent time steps

2. Contrastive Fusion Mechanism:

   • Ensures learned temporal and topological representations are complementary
   • Enhances cross-view consistency by maximizing conditional mutual information between temporal and topological views
   • Strengthens the discriminative power of representations

3. Theoretical Guarantees:

   • Proves that maximizing conditional mutual information can enhance cross-view consistency
   • Contributes to learning more discriminative representations

Model Architecture

Project Structure

Code/
├── data_factory/     # Data loading and processing modules
├── model/           # Model definitions
│   └── ScatterAD.py # Core model implementation
├── metrics/         # Evaluation metrics
├── cpt/            # Model checkpoint directory
├── img/            # Image resources
├── scripts/        # Helper scripts
├── main.py         # Main program entry
├── solver.py       # Training and testing solver
└── requirements.txt # Project dependencies

Getting Started

Download data. You can obtain all benchmarks from dcdetector:Google Cloud.

Requirements

• Python 3.8+
• PyTorch 1.11.0
• PyTorch Geometric 2.1.0
• Other dependencies listed in requirements.txt

Installation

Install dependencies:

pip install -r requirements.txt

Usage

Training a Model

python main.py --mode train \
    --dataset MSL \
    --data_path ../AnomalyDataset/MSL \
    --lr 0.0001 \
    --num_epochs 2 \
    --batch_size 128 \
    --win_size 110 \
    --input_c 55 \
    --output_c 55 \
    --d_model 512 \
    --e_layers 2

Testing a Model

python main.py --mode test \
    --dataset MSL \
    --data_path ../AnomalyDataset/MSL \
    --model_save_path cpt

Training and Testing All Models

bash ./scipts/all.sh

Key Parameters

• --mode: Running mode, options: 'train', 'test', or 'all'
• --dataset: Dataset name
• --data_path: Dataset path
• --lr: Learning rate
• --num_epochs: Number of training epochs
• --batch_size: Batch size
• --win_size: Time window size
• --input_c: Input feature dimension
• --output_c: Output feature dimension
• --d_model: Model dimension
• --e_layers: Number of encoder layers
• --gpu: GPU device ID
• --anormly_ratio: Anomaly ratio threshold

Evaluation Metrics

The model employs 12 comprehensive evaluation metrics:

Label-based Metrics

1. Affiliated-Precision (Aff-P)
2. Affiliated-Recall (Aff-R)
3. Affiliated-F1 (Aff-F)
4. Point-Adjusted-Precision (PA-P)
5. Point-Adjusted-Recall (PA-R)
6. Point-Adjusted-F1 (PA-F)

Score-based Metrics

7. Area Under Precision-Recall Curve (AUC-PR)
8. Area Under ROC Curve (AUC-ROC)
9. Range Area Under Precision-Recall Curve (Range-AUC-PR, R-A-P)
10. Range Area Under ROC Curve (Range-AUC-ROC, R-A-R)
11. Volume Under the Surface-ROC (V-ROC)
12. Volume Under the Surface-PR (V-PR)

Logs

Training and testing logs will be saved in the ../logs/ directory.

License

See the LICENSE file for details.