Automated time activity classification based on global positioning system (GPS) tracking data

Table 2 Time activity classification using the random forest models^a

			Indoor (modeled)	Outdoor static (modeled)	Outdoor walking (modeled)	In-vehicle travel (modeled)	Sensitivity^b	Specificity^c	Precision^d
HCTLS random forest model^e	10-fold cross validation^f	Indoor (coded)	21959	6177	1604	260	73.2%	81.8%	64.1%
		Outdoor static (coded)	9777	9749	3596	1830	39.1%	81.7%	42.3%
		Outdoor walking (coded)	1052	1820	9931	477	74.8%	92.9%	62.5%
		In-vehicle travel (coded)	1475	5307	750	21762	74.3%	96.2%	89.4%
	Evaluation against the full UCI dataset	Indoor (coded)	82874	26278	170	314	75.6%	91.6%	98.9%
		Outdoor static (coded)	850	3137	35	214	74.1%	76.2%	10.1%
		Outdoor walking (coded)	101	993	555	244	29.3%	99.8%	67.7%
		In-vehicle travel (coded)	0	515	60	4562	88.8%	99.3%	85.5%
UCI random forest model^g	10-fold cross validation^f	Indoor (coded)	3978	860	120	42	79.6%	99.0%	97.3%
		Outdoor static (coded)	109	3285	471	371	77.5%	90.2%	73.6%
		Outdoor walking (coded)	0	170	1313	410	69.4%	94.4%	62.1%
		In-vehicle travel (coded)	0	146	210	4781	93.1%	92.6%	85.3%
	Evaluation against the full the full HCTLS dataset	Indoor (coded)	153216	54796	128829	1894	45.2%	92.8%	96.9%
		Outdoor static (coded)	3519	6590	13015	1828	26.4%	84.5%	10.0%
		Outdoor walking (coded)	320	725	11840	395	89.2%	63.0%	7.5%
		In-vehicle travel (coded)	999	3464	3550	21281	72.6%	98.9%	83.8%

^aThe results reported here came from random forest models with 10 trees and a maximum depth of 3 for each tree.
^bSensitivity was calculated as true positive estimation/(true positive estimation + false negative estimation).
^cSpecificity was calculated as true negative estimation/(true negative estimation + false positive estimation).
^dPrecision was calculated as true positive estimation/(true positive estimation + false positive estimation).
^eThe model was developed based on all outdoor static, outdoor walking, in-vehicle travel and randomly selected 30,000 indoor points
^fThe reported validation results were the averages from repeated 10-fold cross validation.
^gThe model was developed based on all outdoor static, outdoor walking, in-vehicle travel and randomly selected 5,000 indoor points from the supplemental UCI data.

ISSN: 1476-069X