MADCluster: Model-Agnostic Anomaly Detection with Self-supervised Clustering Network

MADCluster is a lightweight, plug-in anomaly detection framework for unsupervised multivariate time-series anomaly detection. It is designed to be model-agnostic: you can attach MADCluster on top of various deep anomaly-detection backbones (e.g., reconstruction-based, transformer-based, one-class objectives) by treating the backbone as a Base Embedder that produces temporal representations.

A key challenge in deep one-class anomaly detection is hypersphere collapse, where the network converges to a trivial representation (e.g., near-zero embeddings) and fails to form a meaningful boundary for normality. MADCluster mitigates this issue by single-clustering normal patterns while jointly learning and continuously updating the cluster center, rather than relying on a fixed centroid.

MADCluster consists of three main components:

• Base Embedder: extracts high-dimensional temporal dynamics from input sequences (any backbone can be used).
• Cluster Distance Mapping: pulls embeddings toward a learnable normal center and encourages a compact normal region.
• Sequence-wise Clustering: updates the center online through self-learning, producing stable single-cluster behavior.

To enable effective single clustering ($k=1$) in anomaly detection—where common clustering objectives become degenerate—MADCluster introduces a novel One-directed Adaptive loss, along with a mathematical optimization proof provided in the appendix. This loss trains a one-sided threshold parameter to progressively refine the normal cluster assignment and stabilize centroid learning.

Results. Across four public benchmarks (MSL, SMAP, SMD, PSM), MADCluster consistently improves backbone performance on both point-wise and region-aware metrics (e.g., F1, Affiliation Precision/Recall, Range-AUC, VUS), while remaining computationally lean and easy to integrate.

---

Key Contributions

• Model-agnostic plug-in: works with diverse backbone architectures with minimal modification.
• Prevents hypersphere collapse: dynamic center updates preserve representational expressiveness.
• One-directed Adaptive loss + proof: stable single-cluster learning for one-class anomaly detection.
• Consistent benchmark gains: improves multiple families of baselines on standard datasets.

---

Paper & Supplementary

• Paper (arXiv): https://arxiv.org/abs/2505.16223
• Supplementary Appendix (proofs & extended experiments): docs/MADCluster_Appendix.pdf https://github.com/SYLee1996/MADCluster/blob/main/docs/MADCluster_Appendix.pdf

Note: This repository includes an implementation for running MADCluster across multiple datasets and objectives. See Quick Start below for reproduction.

Project Structure

docs/
└── MADCluster_Appendix.pdf  
datasets/
├── MSL/
├── PSM/
├── SMAP/
└── SMD/
MADCluster/
├── RESULTS/
├── MADCluster_MAIN.py
├── MADCluster_MODEL.py
├── MADCluster_run_all_datasets.py
├── MADCluster_SOLVER.py
└── MADCluster_UTILS.py
README.md

---

Quick Start

1. Set up the Conda environment

conda create -n madcluster python=3.12
conda activate madcluster

2. Install required libraries

pip install pandas numpy torch vus einops

3.1. Run with MADCluster

python MADCluster_run_all_datasets.py --objective one-class --MADCluster

3.2. Run with base model only

python MADCluster_run_all_datasets.py --objective one-class

---

Performance Evaluation Example (MSL)

You can check the MSL dataset performance using the code below:

import pandas as pd

# Replace with your actual result file path
results = pd.read_csv('./RESULTS/<your_result_file>.csv')

print('Results Summary:')
print(f'Precision: {results["precision"].mean():.5f},    Recall: {results["recall"].mean():.5f},     F1: {results["f1_score"].mean():.5f}')
print(f'AU-PR: {results["aupr"].mean():.5f},     AU-ROC: {results["roc_auc"].mean():.5f}')
print(f'R_AUC_ROC: {results["R_AUC_ROC"].mean():.5f},     R_AUC_PR: {results["R_AUC_PR"].mean():.5f}')
print(f'VUS_ROC: {results["VUS_ROC"].mean():.5f},     VUS_PR: {results["VUS_PR"].mean():.5f}')
print(f'Affiliation_Precision: {results["Affiliation_Precision"].mean():.5f},     Affiliation_Recall: {results["Affiliation_Recall"].mean():.5f}')

---

Requirements

• Python 3.12
• pandas
• numpy
• torch
• vus
• einops

---

Dataset Preparation

The following subdirectories are required under the datasets/ folder.

• NASA Datasets — Mars Science Laboratory (MSL) and Soil Moisture Active Passive (SMAP)
  Collected from NASA spacecraft, these datasets contain anomaly information based on incident reports for spacecraft monitoring systems.
  📎 Source

• SMD (Server Machine Dataset)
  Gathered from 28 servers over 10 days, with normal activity observed during the first 5 days and anomalies injected sporadically in the last 5 days.
  📎 Source

• PSM (Pooled Server Metrics)
  Internally collected from multiple application server nodes at eBay with 26 monitored dimensions.
  📎 Source

---

Supplementary Materials

Additional materials that extend the main paper are provided below:

• Mathematical Proofs

  • Analysis of the One-directed Adaptive loss function.

• Extended Experiments

  • Ablation studies (e.g. Multi-cluster ($k>1$) performance analysis and computational efficiency)
  • Image anomaly detection transferability (e.g., MVTec AD)

• 📄 Download Supplementary Appendix (PDF)

These materials are referenced in the paper and are provided for transparency and reproducibility.