This project implements a process to perform anomaly correction of categorical data using two methods. One method uses the gradients of a classifier with respect to the inputs to transform a data input into a non-anomalous point. The other method leverages the distribution of the data learned by a diffusion model to correct data into a "healthy form". We also combined the two approaches into a pipeline to enhance the process. The first method produces a mask of the categories that could be modified using the distribution of the data learned by the diffusion model and then produces multiple versions of possible corrected instances.
