🧠 Sensor-Based Occupancy Prediction

This project investigates how well we can detect room occupancy from wireless sensor network data, specifically under noisy real-world conditions.
The dataset comes from the Intel Berkeley Research Lab, featuring readings of temperature, humidity, light, and voltage from 54 deployed sensors.

The Datset is available on Kaggle as well as UCI ML Datasets:Kaggle Dataset (divyansh22)

---

💡 Core Concept

The core idea is to evaluate model robustness and sensitivity to noise in sensor data. Two parallel tracks were used:

• Random Forest Classifier (Supervised Learning)
  Trained on a version of the data where 10% bit-flip noise was added to a key feature. This tests whether the model can still generalize well, instead of memorizing patterns.

• K-Means Clustering (Unsupervised Learning)
  Trained on clean, original features to assess how well natural groupings align with actual occupancy. This helps understand the structure of the data without labels.

This contrast shows how supervised and unsupervised learning deal differently with signal and noise.

---

⚖️ Tradeoff Design Philosophy: Prioritizing Safety Over Raw Accuracy while training K-Means Clustering model

• This project was developed with a deliberate and thoughtful trade-off strategy: to prioritize detecting real occupancy at all costs — even if that means tolerating false alarms.
• This is especially relevant for applications where a missed detection (False Negative) could have severe consequences, such as:
  • 🔥 Fire or hazard detection systems
  • 👴 Elderly care and fall monitoring
  • 🏢 Unauthorized access or motion sensing in secure areas

It's better to be alerted unnecessarily than to ignore a real threat.

🎯 Why Accuracy Was Not the Primary Goal

• The original dataset is heavily imbalanced toward the "not occupied" class.

• Had the model simply tried to maximize accuracy, it would have:

• Predicted “not occupied” almost everywhere

  • Achieved around 0.80 accuracy, but completely failed to detect real occupancy (high FN)
  • Yielded a low ROC-AUC (~0.43) due to poor class separation

This intentional bias toward recall and responsiveness produced:

Metric	Value / Behavior	Interpretation
✅ Recall	High	Catches nearly all occupancy cases
⚠️ Precision	Moderate to Low	More false alarms, acceptable in safety context
❌ Accuracy	~0.36	Drops due to many FPs, but that’s expected and acceptable
✅ ROC-AUC	~0.58	Better than random (~0.50), shows improved class separation
✅ False Negatives	Very Low	Critical success — system rarely misses true occupancy

The higher ROC-AUC (~0.58) is not despite the trade-off — it's because of it. The model actually learned to distinguish between classes better by focusing on what matters.

🧠 My Intution

This trade-off reflects how machine learning should behave in real deployment, especially in high-stakes domains:

• Accuracy and ROC-AUC are context-sensitive — they shouldn't be blindly optimized.
• Safety-critical ML must be designed to handle uncertainty, imbalance, and consequence.
• False Positives waste time, but False Negatives cost lives.
• This project demonstrates not just technical modeling, but responsible ML system design — with awareness of domain priorities, risks, and ethical consequences.

---

📊 Results

🔁 Random Forest Classifier (Supervised with 10% noise)

Metric	Value
Accuracy	~0.79
ROC-AUC	~0.78

🔘 K-Means Clustering (Unsupervised)

Metric	Value
Accuracy (Post-labeling)	~0.36
ROC-AUC	~0.58
Silhouette Score	~0.85

---

📁 File Structure

Intel_Sensors/
│
├── Intel_Sensors.ipynb    🔹 Jupyter notebook containing entire ML Workflow
├── intel_sensors.py       🔹 Python File
├── LICENSE                🔹 MIT License
└── README.md              🔹 This file !!

👤 Author

Anuj Kulkarni - aka - steam-bell-92