Use a tree-based model to identify outliers and continue training the model using new data
Use unsupervised One-Class Support Vector Machine to detect outliers