essentially:

combining with cawpe is easily motivated:
	use alpha to define the degree to which you trust classifiers that claim to be experts from the train data
	sum their probabilities, weighted by alpha
	that's the ensembles output. question is, are we michael gove?
	

but also, building a decision tree on the probabilistic output of the base classifiers is easily motivated too: 
	informally, individual nodes on the tree would take the form of: "when this classifier predicts this class above/below this level of confidence, it splits the classes well"
	and then this concept is recursively applied until leaf nodes are formed OR factors such as max depth are reached AND/OR leaf nodes are pruned back etcetc, depends on the particular tree implementation
	the general characteristic of decision trees are low bias high variance. a lot of the datasets (and, where CAWPE seems to have a relative advantage...) are relatively small. 
	
	PART1: 
	could break down similar to cawpe vs svm: larger datasets the gap disappears? more cases to form more reliable splits from
		conversely, cawpe strictly as a combination scheme itself combines the case classifiers' output no better with increasing dataset size (obviously, since it isn't a learned combination scheme)
		if we take the assumption that the general expected improvement in the base classifiers' performance (and potentially, probabilistic output in particular) from more data is equally leveraged by cawpe and stacking methods
		then stacking with e.g decision trees will naturally gain relative to cawpe. learned combination vs unlearned => more data equals better for learning
		BUT, while it's not the main point of the paper, more data => longer train times etc etc. 
		
	PART2: 
	general experimental idea to confirm this then: 
		take some big dataset, train base classifiers, get prediction info as we already have done
		dataset has m cases
		with the base classifiers always fully trained on all m cases, i.e. their own predictive performance does not change, train the stacking classifier on m/10, 2m/10, 3m/10 .... m stratified randomly sampled metainfo cases
		compare test performance to cawpe at each stage
		cawpe's performance would be static, the stacking classifier's performance would start lower and a) continually increase all the way up to m or b) increase and then taper off at some proportion of m, dataset dependent
		summarise this on e.g 6 datasets, in a figure with 6 subplots, prop of m on x, test acc on y, flat line at cawpe performance on each of 6 datasets and stackign classifier acc as prop of train cases supplied increases
		
		IF IT TAPERS OFF: perhaps cawpe is simply better? (than this particualr stackign classifier, i.e trees) 
		IF IT DOESN'T: the implication is that if more train data were 'magically' found that the stacking algorithm could make use of (transfer learning?...) then accuracy would eventually overtake cawpe. 
		
		
		
		
	
