This a repository for the working paper: "Machine Learning for Clinical Risk: Wavelet Reconstruction Networks for Marked Point Processes". The paper is available at: wrnppl. Please cite with the bib.
Wavelet reconstruction networks are a neural network that generalizes Hawkes processes. In essence, it combines the relative timing and values of features with reduction functions to estimate a step-wise approximation to the hazard (the rate) of an outcome of interest. It captures the timing with 1-d wavelet reconstructions and timing-and-value with 2-d reconstructions.
Input: a four-column CSV (possibly train/tune/test), and a bunch of arguments, e.g., indicating the target event and settings of the model. Output: rate predictions for the target event and a neural network model