Adaptive Threshold Hybrid LSTM-Transformer — Anomaly Detection

An interactive research web platform showcasing a novel hybrid LSTM-Transformer architecture with adaptive thresholds for multivariate time series anomaly detection in logistics networks.

---

📊 Model Performance

Metric	Value
F1 Score	0.90 (20% better than LSTM baseline)
Precision	0.91
Recall	0.89
AUC-ROC	0.95
Inference Latency	3.2 ms/sample
Model Size	2.3 MB
Training Time	250 seconds on RTX 3090

---

🎯 Key Features

• Hybrid LSTM-Transformer: Combines sequential modeling (LSTM) with long-range dependencies (Transformer)
• Adaptive Threshold Mechanism: Learned per-sample decision thresholds (+5.6% F1 improvement)
• Interactive Demo: Upload CSV data and visualize anomaly detection in real time
• Benchmark Comparisons: Side-by-side evaluation against LSTM, Transformer, and classical baselines
• Research Paper: Full methodology available to read and download

---

📁 Project Structure

anomaly-detection/
├── app/
│   ├── page.tsx               # Home / landing page
│   ├── demo/page.tsx          # Interactive anomaly detection demo
│   ├── benchmarks/page.tsx    # Model comparison & ablation study
│   ├── research/page.tsx      # Research paper viewer + PDF download
│   ├── about/page.tsx         # About the project
│   ├── layout.tsx             # Root layout
│   └── globals.css            # Global styles
│
├── components/
│   ├── theme-provider.tsx     # Dark/light theme
│   └── ui/                    # shadcn/ui component library
│
├── public/
│   └── research_paper_AD_Anomaly_Detection.pdf   # Downloadable research paper
│
├── lib/utils.ts               # Utility functions
├── package.json               # Dependencies
├── next.config.mjs            # Next.js config
├── tailwind.config / postcss  # Styling config
└── tsconfig.json              # TypeScript config

---

🚀 Running Locally

# Clone the repository
git clone https://github.com/your-username/anomaly-detection
cd anomaly-detection

# Install dependencies
pnpm install

# Start the dev server
pnpm dev

Open http://localhost:3000 in your browser.

---

🌐 Deploying to Vercel

1. Push this repository to GitHub
2. Go to vercel.com and import the repo
3. Vercel auto-detects Next.js — no configuration needed
4. Click Deploy

---

📈 Benchmark Results

F1 Score Ranking:
1. HybridLSTMTransformer (Adaptive): 0.90 ⭐
2. TransformerAutoencoder:           0.82
3. LSTMAutoencoder:                  0.79
4. SimpleAutoencoder:                0.72
5. IsolationForest:                  0.70

Ablation Study — F1 impact if component removed:
- Adaptive Threshold:   -5.6% (0.90 → 0.85)
- Transformer Branch:   -6.7% (0.90 → 0.84)
- LSTM Branch:          -8.9% (0.90 → 0.82)
- Entire Hybrid Design: -20.0% (0.90 → 0.72)

---

📧 Contact

• Email: harshvardhanmagar0@gmail.com
• GitHub Issues: Report bugs or request features

---

📊 Citation

@article{magar2026hybrid,
  title={Adaptive Threshold Hybrid LSTM-Transformer for Multivariate Time Series Anomaly Detection in Logistics Networks},
  author={Harshvardhan Santosh Magar},
  year={2026}
}

---

Last Updated: June 2026 | Version: 1.0.0