Table 9.

Depressed users most strongly misclassified in each variation of the preemptive depression identification experiment^a.

		One depression user per control user (1:1)	One depression user per 3 control users (1:3)	One depression user per 5 control users (1:5)	One depression user per 10 control users (1:10)
Classifier		MentalBERT LMb	MentalBERT LM	MentalRoBERTa LM	SVMc using word embeddings
User		d13	d38	d13	d57
Control probability		0.93	0.94	0.99	0.98
Sum of post lengths in words		1696	1888	1696	55,897
Topic		news hawaii time dead blue	sir-geo welcomed invite leave warlock	news hawaii time dead blue	good time people years problem
Chief TF-IDFd features		love minnesota diablo time man bud zoidberg like month hawaii	sir geo welcome invite warlock leave titan psn run need	love minnesota diablo time man bud zoidberg like month hawaii	good know use make time thank link want try like
Depressed vocabulary counts
	people	1	1	1	64
	know	6	0	6	93
	thing	3	0	3	35
	feel	2	2	2	10
	time	5	8	5	99
	woman	1	0	1	7
	go	3	0	3	54
	want	3	1	3	71
	life	2	0	2	28
	relationship	0	0	0	2
Control vocabulary counts
	game	0	1	0	9
	trade	0	0	0	2
	key	0	0	0	4
	team	2	3	2	4
	play	0	1	0	35
	player	0	0	0	8
	shiny	0	0	0	0
	hatch	0	0	0	0
	thank	1	1	0	15
	add	0	2	0	14

^aLexical properties of those users’ posts are provided.

^bLM: language model.

^cSVM: support vector machine.

^dTF-IDF: term frequency–inverse document frequency.