search

mlandry22 / kaggle

Public Kaggle Code and Info
43 stars 46 forks source link

error in use_tf_kernel.m #4

Open attaullah opened 9 years ago

attaullah commented 9 years ago

in folder https://github.com/mlandry22/kaggle/tree/master/bci_challenge, when executing use_tf_kernel.m in Ocatve 4 on Ubuntu gives error: error: 'bintfd' undefined near line 52 column 23 error: called from use_tf_kernel at line 52 column 21

mlandry22 commented 9 years ago

Ah, sorry about that. Been a while, but not too surprising that the code dump didn't quite capture everything in our environments. I was running it in Octave, too, so it does work if I can get the supporting files posted, though I've migrated systems since then. I will try to check it out, but it might be a big before I can update it properly. You may have noticed that this is a poorly organized collection of code that doesn't really go together except thematic content related to Kaggle competitions ;-) Glad you tried it out. Was your interest in running it specific to the competition itself or signal processing or something else?

mlandry22 commented 9 years ago

A quick look at your contributions and I see Alexandre's code, too. That doesn't quite answer the question of your ultimate goal, but I can comment a little further. The kernel methods we were using had very little impact on our score. I would certainly recommend Alexandre's DSP methods over ours without question. The methods of creating features with less-refined DSP methodology that looks more like ML feature creation than DSP is potentially interesting. And possibly the analysis methods in isolating the fifth-trial events, what to do with it, and maybe the ML methods themselves. But the specific Matlab/Octave code was not integral in our solution. Sorry for causing the issues. I will try to remedy the error. Thanks for mentioning it.

attaullah commented 9 years ago

Actually I am student of MS Computer science. I am investigating different pre-processing steps for improving AUC of classifier on data from https://www.kaggle.com/c/inria-bci-challenge and from leader board, I came across your name. I am keen to know about your solution, so I tried your solution . I will be more than happy if you share your working code or report or any material related to techniques you used in this competition.
Any help in this regard will be much appreciated.

mlandry22 commented 9 years ago

We perhaps wrote over each other, so first thing is that for real DSP feature processing, I would recommend Alexandre's code. Note that he is very much an expert in that domain, and from what I remember, his methods are fairly easy to follow along and reuse. He won other related competitions as well and had great write-ups.

That said, the feature processing I used is shown most directly at line 61 of primary_model.R: https://github.com/mlandry22/kaggle/blob/master/bci_challenge/primary_model.R

There are a few comments in that section, and those comments pertain to my naive, but consistent, way of selecting features. My pattern was one of EDA, really. I plotted a bunch of signals with the entire graph colored red or green based on the output, and then tried to find commonality in those charts based on patterns my eyes were picking up. I would then try to encode that "feature" that my eyes were seeing, and would test the new feature against the entire data set to see if it was decent or not on all samples rather than the few I was focusing on. Each feature was tested individually, and then added to my feature set if the independent AUC was decent. That is why you see AUC comments along with each feature calculation.

The methodology I followed is a decent one for feature generation on any problem. However, it is not a great scalable way to get started with a signal processing task, for which we do have good algorithms tailored to that domain. Good luck to you, and if you can run Alexandre's code, you'll have a stronger feature basis than my own. I'll be glad to answer any questions, if you still have any.