2018.WACV.Anomaly explanation using metadata
paper
main idea
Anomaly detection is first performed using Principal Component Analysis (PCA). After the identification of anomalies, tags are generated for each picture in the data set; every tag is a word describing the picture, and these tags constitute its metadata. Then, the tags corresponding to the greatest number of anomalies are identified and returned as global explanations of anomalies. The identification of important tags, importance with regard to anomaly detection, is made using algorithms like PRIM (Patient Rule Induction Method) whose objective is to find regions in high-dimensional input space with large values of a real output variable. This explanation can be used with any anomaly detection algorithm; it is therefore a model-agnostic method. It is an explanation by feature importance since the features space has just changed from the space of pixels of the images to the space of metadata, but ultimately the most relevant features/metadata are returned.