Enhancing Multivariate Time Series Anomaly Detection through Causal Disentanglement with Graph Neural Networks

This repo includes a complete framework for multivariate anomaly detection, using a model that is heavily inspired by CDRL4AD(Causally Disentangled Representation Learning for Anomaly Detection).
For contact, DataScienceLabs

❗Key Features

1. Causal Discovery:
   • Utilizes an attention mechanism to identify and learn time-lagged causal relationships within multivariate time series data.
   • Constructs a Directed Acyclic Graph (DAG) to represent these causal relationships, enhancing the model's understanding of how past data influences future outcomes.
2. Causally Disentangled Representation (CDR):
   • Employs a multi-head decoder variational autoencoder (VAE) to create causally disentangled representations.
   • Ensures each latent variable aligns with predefined causal relationships, allowing for sophisticated data representation and better interpretability.
3. Node and Edge Correlation Representation (NECR):
   • Encodes variable correlations at both node and edge levels using a dual-level Graph Attention Network (GAT).
   • Learns graph structures by connecting nodes with similar patterns, improving the model’s ability to detect anomalies through structural insights.
4. Temporal Dependency Representation (TDR):
   • Captures sequential relationships within the data, accounting for dependencies over time.
   • Aggregates features from past observations to model temporal dependencies, enhancing the detection of anomalies with temporal context.
5. Root Cause Analysis:
   • Provides clear insights into the root causes of detected anomalies by analyzing variable contributions.
   • Identifies top-K root cause variables, facilitating efficient diagnosis and problem-solving in real-time scenarios.
6. Model Interpretability:
   • Offers t-SNE visualization of causally disentangled representations, illustrating distinct clusters of variable representations.
   • Enhances model transparency by explicitly encoding causal influences and their significance in the representation.
7. Robust Performance:
   • Demonstrates high precision, recall, and F1 scores across diverse datasets, showcasing the model's robustness to different data characteristics and distributions.
   • Includes comprehensive ablation studies to validate the importance of each model component.

🖥️ Getting Started

To clone the repo:

git clone https://github.com/datascience-labs/cdrl4ad.git && cd cdrl4ad

Install dependencies (virtual env is recommended):

pip install torch==2.0.0 torchvision==0.15.1 torchaudio==2.0.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

To train:

python train.py --dataset <dataset> --epochs <epoch>

Data

We used the SWaT, HAI, SMD, and PSM datasets. The originals of each dataset can be requested through the links below.

Dataset	Source
SWaT	iTrust
HAI	HAI
SMD	OmniAnomaly
PSM	RANSynCoders

Cite

TBA