Lots of printlines from Math.java when running Lasso

[Xyclade]

Lots of printlines from Math.java when running Lasso

[haifengl]

Sure, you can remove them. At least the first one. But I am little worried about this case. This function ( iterative biconjugate gradient method) usually should converge pretty fast and thus not much print out. Please check if your LASSO model works well for test data. Thanks!

[Xyclade]

The Lasso model is not working well for the test data, but that is the actual goal this time. I'm writing an example of how it can go wrong, thank you for the heads up though! 👍

[haifengl]

Interesting. It doesn't work well because LASSO is not a good fit for the problem? Or the parameter settings (e.g. regularization factor)?

[Xyclade]

It's because there is too few datapoints and no actual statistical relation within the data.

As John Tukey once said:

The data may not contain the answer. The combination of some data and an aching desire for an answer does not ensure that a reasonable answer can be extracted from a given body of data.

This is the case in my example, to make people aware that Machine Learning cannot perform miracles on every single dataset.

[haifengl]

Is your data size less than the dimensionality?

[Xyclade]

yes, its 100 datapoints (with rank 1 to 100) with 27000 + features, this can never go well :) But since it's an example for my blog rather than an actual dataset that I want to use for predictions, I still worked it out to show what goes wrong when you do these kind of things.

In the end result the trained LASSO model should predict a rank value for a new datapoint, but it predicts worse than just predicting the average (50). This makes it a perfect example for what can go wrong if you have no clue what you are doing 👍

[haifengl]

That is the true reason. It is known as small sample size problem. I have a paper (http://lectures.molgen.mpg.de/networkanalysis13/LDA_cancer_classif.pdf) to deal with it.

[Xyclade]

Cool tnx! I'll read that and see if I can incorporate it if that's ok with you?

[haifengl]

Try FLD in SMILE for your case. I don't remember if I implemented this algorithm there. Thanks!