Why • Installation • Benchmarks • License
This repository is a code submission for the Master's course 'Machine Learning' at the University of Hamburg.
It builds upon a kaggle challenge published in 2021. The goal is to predict drought scores for the 6 next weeks using meteorological, soil and drought data from the last 180 days.
More specifically, it explores the usage of MiniROCKET-transformed features and different classification and regression techniques and performs analyses on feature and history importance.
Droughts are becoming more severe and more frequent.
Even though the prediction of droughts does not prevent them or make them less harmful, it allows countries, cities or individuals to adjust and take preventive measures when needed.
The prediction of droughts may also allow us to better understand influencing factors of droughts.
32 GB of RAM are recommended if you plan to run this notebook again yourself. If you don't have enought RAM, creating a swap file might be a possible solution.
First, you have to download the data from https://www.kaggle.com/cdminix/us-drought-meteorological-data.
Simply place the four files test_timeseries.csv, train_timeseries.csv, validation_timeseries.csv and soil_data.csv directly into the main directory.
Next, run the jupyter notebook by navigating into the directory in your terminal and typing
jupyter notebook
That's it! 🎉 You are ready to go. All relevant requirements will be installed with pip in the notebook itself.
Executing all cells of the notebook again will take several hours.
Javascript widgets are used to display colored progress bars. If you don't have javascript widgets enabled in Jupyter and encounter the error widget javascript not detected. it may not be installed or enabled properly., try running the following commands before starting the notebook:
sudo jupyter nbextension enable --py --sys-prefix widgetsnbextension
jupyter nbextension enable --py widgetsnbextension
The following performance benchmarks were either given or achieved by our own experiments:
| Model | Macro F1 Mean | MAE Mean |
|---|---|---|
| Random guessing | 0.108 | 2.244 |
| Majority class | 0.133 | 0.578 |
| Random guessing (stratified) | 0.164 | 1.046 |
| Ridge regression (MiniROCKET features) | 0.444 | 0.372 |
| Ridge regression (default features) | 0.579 | 0.255 |
| LSTM (@MiniXC) | 0.639 | 0.277 |
Status: August 2021
An up-to-date benchmark list of all handed in models can be found in the original kaggle task: https://www.kaggle.com/cdminix/us-drought-meteorological-data/tasks?taskId=3422.
The confusion matrices of the best-performing regressors per feature type developed in this notebook lare shown below.
Confusion matrix of ridge regression using MiniROCKET-transformed features:
Confusion matrix of ridge regression using default features:
This repository has been published under the MIT license.