Sensor Data Classification: CEIP-DS-JECRC

This repository contains a high-performance machine learning solution for classifying sensor-based data. The approach utilizes LightGBM with custom windowed features and precision threshold tuning to handle highly imbalanced classes.

---

🚀 Overview

The objective is to predict binary targets from a time-series sensor dataset ($X1$ to $X5$). This solution focuses on capturing temporal dependencies and signal volatility through manual feature extraction.

🛠️ Feature Engineering

To improve the model's predictive power, the following transformations were applied:

• Time Features: Extracted hour, day_of_week, and is_weekend from timestamps.
• Sensor Aggregates: Computed row-wise mean, std, and max across all sensors.
• Windowing & Smoothing: * Rolling Mean & Variance: Captured local trends and volatility.
  • EWMA: Exponentially Weighted Moving Averages for noise reduction.
  • Differencing: Calculated step-to-step changes ($X_t - X_{t-1}$).
• Signal Ratios: Created $X1/X2$ and $X3/X4$ interaction features.

📉 Methodology

• Validation: A 73/27 Chronological Split was used to simulate real-world forecasting and prevent data leakage from future timestamps.
• Scaling: Features were normalized using StandardScaler.
• Model: LGBMClassifier with scale_pos_weight to address the heavy class imbalance.
• Optimization: Custom thresholding was implemented. Instead of the default $0.5$, the model searches for the optimal threshold (found at 0.97) to maximize the F1-Score.

📊 Performance

• Validation F1-Score: ~0.618
• Optimal Threshold: 0.97
• Model Params: 250 estimators, 0.03 learning rate, max depth of 5.

📂 Requirements

import pandas as pd
import numpy as np
from lightgbm import LGBMClassifier
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import f1_score