Problem

We were challenged to predict the phase of an earthquake—a multiclass classification task—within a strict 5-hour time limit. The results had to be submitted to Kaggle for leaderboard ranking and also presented live to a panel of judges.

Resolve and Result

We broke the process down into three key stages:

1. Data Cleaning & Transformation

• Applied labeling and One-Hot Encoding to categorical features
• Split into train, validation, and test sets
• Scaling using Standard Scaler
• Handled class imbalance with SMOTE oversampling

2. Exploratory Data Analysis (EDA)

• Used a heatmap to explore correlation between sensor values
• Discovered distribution of earthquake phases pretty different
• No missing values

3. Data Modeling

• Due to limited time and using only a standard laptop, we prioritized fast experimentation
• Performed hyperparameter tuning with Randomized Search
• Tried several models:
  • Decision Tree
  • Random Forest
  • XGBoost (best-performing)

🏆 Results

• Our XGBoost model achieved the highest accuracy of 90%
• We secured 1st place on the Kaggle leaderboard for the competition!

Repository

• Github Code
• Kaggle