LANL-Earthquake-Prediction

Submissions for the Kaggle competition found here: https://www.kaggle.com/c/LANL-Earthquake-Prediction

Competition Description

Given a segment of an acoustic signal, one must predict the 'time_to_failure', or in other words, time before the next laboratory earthquake. The metric used for model evaluation is Mean Absolute Error (MAE).

Note

I will update the repository with the best approach I have been able to come up with yet. The repo is live, and subject to change.

Best Approach thus far:

I used Gradient Boosting after extracting several features from the input acoustic data.

Data Engineering:

1. Split input signal into chunks of 150000 data points. This is done based on the information given by the competition organisers.
2. For each of these segments, several statistical features are calculated (can be viewed in the notebook).
3. The training data is a collection of these segments, 4194 of them, to be precise.
4. The data is then scaled using scikit-learn's StandardScaler() function.

Next, we use the GradientBoostingRegressor and train over the training data. The training is done using K-Fold cross validation, and for each fold, the estimator object (model) is saved. After examining the results, the best estimator is chosen to make the predictions on the testing data. I used the huber loss function as it is less prone to fluctuations and outlier values, and I feel in this case, it gives better approximations.

Kaggle Score for Approach

I got MAE=1.583 on the Public Leaderboard for this approach.