truocphamkhac-agilityio / kaggle-vsb-power

:zap: Abandoned but interesting attempt

Kaggle VSB power line fault detection

My solution is very standard and consists in manually extracting features before feeding them to LightGBM. This worked quite well at first and I managed to reach the top 10 of the competition. However RNNs seem to be the right way to go, but I'm not very interested in deep learning. Usually I don't upload Kaggle solutions that didn't do well, but I'm making an exception for this one as I'm quite satisfied with the feature extraction pipeline I put in place. If you want to run the code make sure you are using Python 3 and have installed the dependencies listed in the requirements.txt file.

Splitting the signals

>>> python scripts/split_signals.py

The data provided by the competition is stored in an HDF5 file. Reading from the HDF5 file is anything but fast. My idea was to first split the signals into separate numpy files using the numpy.load method. Loading the signals using numpy.save then takes something in the range of microseconds. This is extremely important because throughout the competition the data will be loaded in memory many times.

Aligning the signals

>>> python scripts/find_signal_origins.py

Although each signal represents one period of an electrical sine wave, they don't all start at the same time. I decided to align them so that they all started from 0 and started by going upwards. This could be useful as some features could be based on a particular region of the signal. I didn't really exploit this as I gave up the competition when RNNs arrived. To align the signals I used a simple method which starts by searching for the two points where the signal crosses 0. Because there is a lot of noise I used a k-means clustering scheme with k = 2 to approximate the two positions. I then decided which of the two crossings was the one that I wanted by looking left and right from both crossings.

Extracting features

>>> python scripts/extract_solo_features.py

I won't go into detail about which features I extracted as I'm sure some people did better and will talk about it when the competition is over. The only thing I want to mention is how I extracted the features. As mentioned above the signals were split into separate .npy file using numpy and were stored in the data directory. I then simply looped over the files and extracted the features in parallel using a ThreadPoolExecutor from the concurrent.futures module from Python's standard library. The trick is that before computing the features I first loaded the ones that had already been computed so that I had didn't recompute them unnecessarily. This is definitely not rocket science but I thought the code to be quite concise and rather readable so I deemed it worthy of being shared online.

Cross-validation folds

>>> python scripts/make_folds.py

I like generating CV folds before doing the machine learning. I save these as a JSON file called folds.json in the oof directory and load them during the machine learning phase. This is practical because you can share the folds with others and use them with multiple models. I made sure that the folds didn't "leak" by putting signals from the same measurement in the training folds as well as the validation folds.

Machine learning

I started by trying to learn the labels of each signals individually. I then converted the problem to a multi-class classification problem by joining the signals of each measurement together. Because each signal has a binary label and there were 3 signals per label this resulted in an 2^3 = 8 class problem. What's more by permuting the 3 signals I was able to augment the multi-label dataset by a factor of 3! = 6. The code is available in the Solution.ipynb notebook. I didn't comment it but it should be readable. In the end this will produce a submission in the submissions directory and out-of-fold predictions in the oof directory.