UCLA

In this paper, I compared 6 semi-supervised point outlier detection algorithms: LOF, robust PCA, autoencoder, SOM, one-class SVM and isolation forest. In all experiments, I training the models on only normal data points. Then, I use the models to detect the outliers in the testing data sets basing on the fact that if a point is not a normal point, the point is an outlier. I do the experiment on both generated data and real data. I found each of the 6 algorithms has both advantages and disadvantages. I have described them in details and give some suggested solutions to the weak points.