TABLE 2.

Prediction performance (Accuracy and AUC) of the STL, MTL-moods, and MTL-user methods. Bolded entries represent significant improvements over the STL model, indicating that multitasking for personalization is by far the most effective approach.

	Classifier	Mood	Stress	Health
Baseline	Majority class	50.4%, .500	50.7%, .500	54.4%, .500
	LSSVM	60.2%, .603	58.1%, .581	62.3%, .614
STL	LR	56.9%, .569	59.4%, .594	55.4%, .544
	NN	60.5%, .606	60.1%, .600	65.9%, .648
	NN (all feats)	65.8%, .658	67.9%, .678	59.0%, .591
	MTMKL	59.4%, .594	58.8%, .587	62.0%, .610
MTL - moods	HBLR	58.3%, .583	57.8%, .578	55.1%, .551
	MTL-NN	60.2%, .602	60.1%, .600	65.3%, .643
	MTL-NN (all feats)	67.0%, .670	68.2%, .682	63.0%, .623
	MTMKL	78.7%, .787	77.6%, .776	78.7%, .786
MTL - people	HBLR	72.0%, .720	73.4%, .734	76.1%, .760
	MTL-NN	77.6%, .776	78.6%, .785	79.7%, .792
	MTL-NN (all feats)	78.4%, .784	81.5%, .815	82.2%, .818