Emotiv Brainwave SVM-based Login
- Webserver running PHP (PHP 7.0 with lighttdp tested)
- (probably) Re-combile libsvm (svm-predict)
The original data looks something like:
Channel,Theta,Alpha,Lowbeta,Highbeta,Gamma
AF3,621.4072977,794.9707577,167.8989619,431.5566596,462.8267375
T7,588.8385383,1002.008753,143.9251969,503.5703621,611.768041
First we need to prepare the data.
Right now we have 2 users "meteor" and "rumba". Each user has 10 'samples' collected from different experiments. First, remove 2 of$
Combine the remaining samples into libsvm format (column 4 and 5 are high/low beta):
cat * | awk -F',' '{print "1 1:" $4 " 2:" $5}' | grep -v beta > meteor.single.train.libsvm
This code means that the user meteor has a label of "1". If you will be creating a multi-class classifer, other users need to have a different label.
If you get a bunch of empty lines (I did) then you can remove them with this:
cat meteor.single.train.libsvm | grep -v "1 1: 2:" > meteor.single.train.clean.libsvm
Now you should have a very large training set FOR EACH USER.
If each user have been assigned different labels, then you can combine them.
First scale: svm-scale -l -1 -u 1 -s range meteor.single.train.clean.libsvm > meteor.single.train.clean.libsvm.scale
After first, reuse 'range' file: svm-scale -r range rumba.single.train.clean.libsvm > rumba.single.train.clean.libsvm.scale
For multi-class, I used default training settings: svm-train multi.train.clean.libsvm.scale This produces "multi.train.clean.libsvm.scale.model" that can be used for classification of both users.
For one class: svm-train -s 2 -n 0.2 meteor.single.train.clean.libsvm.scale One model is produced for each user.
Test data goes through the same libsvm format and scaling process. Do not combine test data with other data. Make sure you know which test data belongs to which user.
Once test data has been scaled, and the model has been trained: svm-predict meteor.single.test.clean.libsvm.scale rumba.single.train.clean.libsvm.scale.model predict.out Here we are using meteor data, and trying to classify it with rumba's model. It should result in a "-1". If meteors data is used with meteors model it should result in a "1". (One-class)
For multi-class, if meteor has a label of 1 and rumba has a label of 2, meteors data should produce "1" and rumbas data should procuce "2".
The way the data is set up, each line of "predict.out" will be classified. For a quick and dirty solution, I am taking the average of the classified values awk '{s+=$1}END{print NR}' RS=" " predict.out So far, this method has worked very well.
We could also calculate what the error rate was, and cut off at a certain point.