Purpose: Build a support vector machine classifier using f_10fps
and grp
outputs from bsoid_gmm.m. This allows future prediction of user defined groups. The algorithm defaults at using the clusters from GMM, but user can also input the merged groups and have a classifier built to their desires. However, we would argue against building a classifier based on human perception of what is the same behavior; instead, build this classifier based on the x number of GMM clusters, and merge the predicted labels later.
function [OF_mdl,CV_amean,CV_asem,acc_fig] = bsoid_mdl(f_10fps,grp,hldout,cv_it,btchsz)
Run bsoid_gmm.md first
-
F_10FPS
Compiled features that were used to cluster, 10fps temporal resolution. -
GRP
Statistically different groups of actions based on data. Output is 10Hz. -
HLDOUT
Percentage of data randomly held out for test. Default is 0.20 (80/20 training/testing). -
CV_IT
Number of times to run cross-validation on. Default is 100. -
BTCHSZ
Batch size for randsampling, make sure the hold out data is <= CV_IT*BTCHSZ. Default is 200. -
IT
The number of random initialization for Gaussian Mixture Models. This attempts to find global optimum, instead of local optimum. Default is 20.
-
OF_MDL
Support Vector Machine Classifier Model. -
CV_AMEAN
Cross-validated accuracy mean. -
CV_ASEM
Cross-validated accuracy standard error fo the mean. -
ACC_FIG
Box plot showing classifier performance with individual data points representing a randomly subsampled test set from the hold out portion.
Run bsoid_svm.md next