Since all threshold methods that I implemented so far are quite unintuitive and hard to verify, I spent last week writing tests for most of their parts. I created test-cases based on fixed and random datasets (mainly taken from the papers where the methods were introduced), in which results are compared against MATLAB implementations of the same problem. For most cases, the difference is asserted to be smaller than 10E-15.

Although it takes a lot of time and effort, I believe these automatic tests should come with every more complicated method in SHOGUN. I remember quite some cases where I spent an hour on finding an error that would have been easily caught by a test.

I hope to finish all all statistical tests based on the quadratic MMD this week. I still need to implement the biased version (which can be used along with the threshold methods) and a possibility for choosing the type of statistic to use. I also plan to implement the linear time MMD and corresponding tests (Gaussian approximation of null-distribution). I won’t have Internet access this week since I am in nowhere in Poland — so expect a larger patch at the end of the week.