Statistical features usually work well. You can try Gray Level Co-occurrence Matrix (GLCM), Histogram of Oriented Gradient (HOG), or Grey Level Run Length Matrix (GLRLM).
The values in meas(xdata) represent some kind of measured data. SVM does not know or care how the measurements were taken or what they actuallyare. As far as SVM is concerned, it does not matter whether some parts are (say) temperature and other parts are (say) x-ray pixel intensity, and other parts are (say) acceleration along a particular axis.
"I mean why a set (x,y) is classified as BENIGN for example."
It justis. SVM does not care about cause and effect; at most it cares about correlation. If, for example, one of the readings were I2Sb2F11 concentration and another of the readings were H2S03, and high readings of I2Sb2F11 occurred in the samples marked class Tumor Grade 1, then SVM does not care whether I2Sb2F11causes Tumor, or Tumorcauses I2Sb2F11, or whether the H2S03 was reacting with the probe leads and leading to contamination that the healthy cells rejected but the Tumor cells were not able to flush.
The data you input for SVM does not in itself have an explanatory power: you just take a bunch of measurements of things thatmight be relevant,somehow, and you have some kind of external judgement about the class of the training samples, and SVM tries to figure out what the importantcorrelations are. Oncecorrelations are identified, then the researchers can go back and write up another grant to try to modelcausation.
Best Answer