21. Variants'have'got'increasingly'complex'over'2me'
statistically significant differences between the edit distance
momentum of variants of banned tags and those of the
advisory tags (p=.35). We conjecture the pro-ED
community adopts increasing lexical variance in their tags
to avoid Instagram’s moderation of tags through bans or
in size, “thighgap” five times, and “thinspo” more than two
times. This increased activity shows that the pro-ED
community continues to thrive even though overall
participation dropped on some tags.
Figure 2. Changes in Levenshtein’s edit distance with emergence of newer lexical variations over time – shown for “anorexia”,
“eatingdisorder” and “thighgap” tag chains. Each data point in the scatter plot corresponds to the edit distance of a particular
variant at a certain point in time.
(Chancellor,'Pater,'Clear,'Gilbert,'De'Choudhury,'CSCW'2016)'
Tag$chain$ Max.$edit$dist.$ Mean$edit$dist.$ Momentum$
ana' 5' 2.556' ±1.257' 1.281'
anorexia' 7' 1.939' ±1.043' 1.285'
anorexianervosa' 8' 1.629' ±1.096' 1.192'
bonespo' 6' 2.5' ±1.708' 1.367'
bulimia' 7' 1.755' ±0.980' 1.203'
ea2ngdisorder' 5' 1.629' ±0.778' 1.156'
secretsociety123' 6' 2.255' ±1.239' 1.321'
skinny' 4' 1.722' ±0.870' 1.221'
thighgap' 5' 2.084' ±1.006' 1.218'
thin' 3' 1.889' ±0.737' 1.188'
thinspira2on' 6' 2.307' ±1.318' 1.396'
thinspo' 9' 3.125' ±1.952' 1.383'
Mean'momentum' 1.3'
22. Modera2on'was'followed'by'increased'social'
engagement'
Next, we examined volume of unique users associated with
the root tags and their variants as well as the Jaccard
similarit) overlap of users between the two (Table 4). In
general, there are some tag chains where there is
considerable overlap of users between the root tags and
adopters of their variants (e.g, “bulimia,” “secretsociety”).
However, most tag chains have little overlap (e.g., “ana,”
adopt these variations to overcome moderation restrictions
enforced by Instagram. It also implies that adoption of
lexical variation in tag usage might be an intrinsic
individual characteristic; that is, the users likely to embrace
this strategy are perhaps a small fraction of those who use
the root tags. Alternatively, it may also indicate the
propensity of a certain segment of the pro-ED community
to adopt the lexical variations in their content sharing,
perhaps to avoid discoverability more broadly, build and
whereas upward arrows indicate an increase.
Figure 3. Proportion of weekly posts for six root tags and their corresponding three most frequent variants over time. The
vertical grey lines indicate time when Instagram publicly reported change in its community policies (Apr 2012).
(Chancellor,'Pater,'Clear,'Gilbert,'De'Choudhury,'CSCW'2016)'
28. Post Post tags Raw Rounded Annotated
fly, dragonfly, prints, fashion, sketch, illustration, ink, masonry, textiles 1.02 1 1
rainbow, mexicancat, burrito, tablet, rainbowcat, rainbows, wallpaper, cat, live-
wallpapers, burritocat, livewallpaper, mexican, catwallpaper
1.15 1 1
starve, bodydismorphia, scars, ana, selfharmmm, bulimia, dead, bones, suicide,
triggerwarning, cuts, razorblade, depression, anorexia, razor, sue, cut, mia, tw,
blade, fat, hungry, purge, ugly, cutting, deb, blood
2.87 3 3
disappointment, pain, gayteen, selfharmmm, relapsed, selfinjury, selfhate, slit,
cutting, fat, cuts, anorexic, depression, razor, gaykik, death, blood, depressed,
ugly, blithe, imsorry, anxiety, triggering, cutter
2.77 3 3
beautiful, suicidal, ana, bulimic, bulimia, fat, sad, tumblr, cut, blood, cuts, love,
ew, hate, anorexic, depression, anorexia, funny, suicide, ugly, mia, helpme,
depressed, cutting, killme, thin, skinny, happy
2.82 3 3
tumblr, depressed, quote, scars, ehtilb, dark, selfharmmm, blithe, depression 2.01 2 3
koolaid, lazy, girl, thinspoo, sugarfree, exercise 1.67 2 3
Table 2: Example posts with tags, raw MIS (Raw) generated using our method, discretized MIS rating (Rounded) and MIS per
agreed upon annotations from four researchers (Annotated). All post images are blurred to avoid disclosure of user information.
(Chancellor,'Lin,'Goodman,'Zerwas,'De'Choudhury,'CSCW'2016)'
Inferred'MIS'of'proOED'posts'
29. 10 20 30 40 50
0
0.05
0.1
0.15
Frac.users(med,high)
Months
10 20 30 40 50
0.7
0.8
0.9
1
Frac.users(low)
med
high
low
s (Table 3), we find that a notable fraction of these
so use the Instagram platform for sharing non-mental
elated content, as indicated by the low level of MIS
rity of the posts.
MIS Rating Post Count Percentage
Low (1) 22,913,989 87.41%
Medium (2) 1,990,031 7.59%
High (3) 1,311,288 5.00%
e 4: Proportion of posts with di↵erent MIS rating.
r a deeper examination of the way the three levels of
ange over time reveals that, while small in the propor-
posts, alarmingly, the relative fraction of users who
o-ED content and show high MIS has been on the rise.
(Chancellor,'Lin,'Goodman,'Zerwas,'De'Choudhury,'CSCW'2016)'
• From'month'18'(Mar'2012)'to'
month'48'(Oct'2014),'both'
medium'and'high'MIS'ra2ng'user'
propor2ons'show'a'steep'increase,'
whereas'low'MIS'ra2ng'shows'
decline'during'the'same'period.''
• At'its'peak,'as'many'as'10%'of'
users'in'our'data'are'inferred'to'
express'high'MIS'ra2ng'
30. o the two best models: w = 7 and w = 10. “Intc” is inter-
ept, and i is the coe cient weight corresponding to the i-th
redictor variable, in other words it is the MIS rating at time
i when the prediction is for MIS rating at time t.
Model Accuracy (%) Precision Recall F1
w = 1 59.89 0.647 0.655 0.651
w = 2 62.76 0.674 0.706 0.690
w = 3 67.98 0.676 0.724 0.699
w = 5 69.32 0.723 0.797 0.758
w = 7 81.89 0.817 0.804 0.810
w = 10 73.99 0.808 0.790 0.799
w = 13 72.26 0.808 0.754 0.780
w = 15 67.42 0.774 0.728 0.751
w = 17 66.32 0.684 0.619 0.650
w = 20 61.50 0.638 0.617 0.627
able 6: Predicting low, medium, high MIS ratings of users in
eldout test set using the regularized multinomial logit model.
his comparison with the null model is statistically signif-
cant after Bonferroni correction for multiple testing (↵ =
.005 since we consider 10 di↵erent models corresponding
o the 10 values of sliding window w). The best model fit
n terms of lowest deviance) is given by the w = 7 model
2
(7, N = 90K) = 915.523 279.801 = 635.722, p < 10 10
),
with best fits (low deviance) for models where w is closer to
w = 7, and decreasing as w goes lower or higher.
Next, in Figure 3 we report the mean weights of the coef-
cients ( ) in the models generated on the bootstrap train-
ng samples. Specifically we report on two models with best
model fits in Table 5. We observe that the weights of the co-
cients monotonically decrease backwards in time from the
2 4 6 8 10
20
30
40
50
60
70
80
Performancemeasure
Prediction horizon (months)
accuracy
F1
(Chancellor,'Lin,'Goodman,'Zerwas,'De'Choudhury,'CSCW'2016)'
• Historical'tagged'content'and'levels'
of'MIS'inferred'from'Instagram'
content'may'predict'the'MIS'level'
in'a'user’s'content'in'the'future.'
• We'are'able'to'forecast'a'user’s'
manifested'MIS'over'2me'up'to'
eight'months'into'the'future'based'
on'their'(low,'medium'or'high)'MIS'
previously'seen'in'their'Instagram'
posts.''