Denoising with autoencoders

Autoencoders, initially better known for non-linear data compression (a kind of non-linear PCA) and image denoising, started being recognized as an interesting tool for tabular competitions after Michael Jahrer (https://www.kaggle.com/mjahrer) successfully used them to win the Porto Seguro’s Safe Driver Prediction competition (https://www.kaggle.com/c/porto-seguro-safe-driver-prediction). Porto Seguro was a popular, insurance-based risk analysis competition (more than 5,000 participants) characterized by particularly noisy features.

Michael Jahrer describes how he found a better representation of the numeric data for subsequent neural net supervised learning by using denoising autoencoders (DAEs). A DAE can produce a new dataset with a huge number of features based on the activations of the hidden layers at the center of the network, as well as the activations of the middle layers encoding the information.

In his famous post (https://www...