Table 13.

Depressed users most strongly misclassified in each variation of the temporal experiment. Lexical properties of those users’ posts are provided.

		One depression user per control user (1:1)	One depression user per 3 control users (1:3)
Time span		Last 12 weeks	Last 12 weeks
Classifier		BERTa LMb	BERT LM
User		d52	d52
Control probability		0.869	0.935
Sum of post lengths in words		1225	1225
Topic		england belgium hamster time team	england belgium hamster time team
Chief TF-IDFc features		thank team player help time goal cage post second start	thank team player help time goal cage post second start
Depressed vocabulary counts
	people	0	0
	know	1	1
	thing	1	1
	feel	0	0
	time	4	4
	woman	0	0
	go	0	0
	want	2	2
	life	0	0
	relationship	0	0
Control vocabulary counts
	game	2	2
	trade	0	0
	key	0	0
	team	4	4
	play	0	0
	player	1	1
	shiny	0	0
	hatch	0	0
	thank	2	2
	add	1	1

^aBERT: Bidirectional Encoder Representations From Transformers.

^bLM: language model.

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