CSCW '16, FEBRUARY 27–MARCH2, 2016, SAN FRANCISCO, CA, USA

Style in the Age of Instagram

Predicting Success within the Fashion Industry using Social Media
Jaehyuk Park Giovanni Luca Ciampaglia Emilio Ferrara
Indiana University Indiana University Indiana University
Bloomington, IN, USA Bloomington, IN, USA Bloomington, IN, USA
[email protected] [email protected] [email protected]

ABSTRACT Modeling has in fact a very special meaning in fashion, a

Fashion is a multi-billion dollar industry with social and eco- multi-billion dollar industry with strong social and economi-
nomic implications worldwide. To gain popularity, brands cal implications worldwide [24]. Models play the main role
want to be represented by the top popular models. As new in advertisements and runways, which are, historically, the
faces are selected using stringent (and often criticized) aes- main ways brands communicate with their customers. They
thetic criteria, a priori predictions are made difficult by infor- contribute to frame consumer experience and promote con-
mation cascades and other fundamental trend-setting mecha- sumption, as their attractiveness becomes associated to the
nisms. However, the increasing usage of social media within brands they work for [46]. This is especially true for luxury
and without the industry may be affecting this traditional sys- goods, whose aesthetic value is more important than practi-
tem. We therefore seek to understand the ingredients of suc- cal usage. Only those who appeal the aesthetic sensibility of
cess of fashion models in the age of Instagram. Combin- fashion designers stand chances to become popular [10].
ing data from a comprehensive online fashion database and
Ethnographic studies show in fact that casting directors con-
the popular mobile image-sharing platform, we apply a ma-
sider both objective physical characteristics — such as body
chine learning framework to predict the tenure of a cohort of
size, height — and subjective considerations — the reputa-
new faces for the 2015 Spring / Summer season throughout
tion of the agency representing the model — to be important
the subsequent 2015-16 Fall / Winter season. Our framework
decision-making criteria for casting [33, 19, 32]. However,
successfully predicts most of the new popular models who
the same studies also uncover how information cascades are
appeared in 2015. In particular, we find that a strong social
a critical part of what makes a model successful. Similarly to
media presence may be more important than being under con-
the careers of scientists [34], fashion models also benefit from
tract with a top agency, or than the aesthetic standards sought
strong cumulative advantage effects, by which small differ-
after by the industry.
ences in prestige between competing individuals get ampli-
Author Keywords fied, for example by means of word of mouth [39, 42, 19].
Social Media; Fashion Industry; Science of Success The job market for fashion castings has strong seasonal com-
ponents, revolving around week-long trend-setting industry
ACM Classification Keywords events (“Fashion Weeks”), during which a dense calendar of
Human-centered computing: Collaborative and social com- shows is organized in various locations in a major city. The
puting—Social media; Information systems: World Wide four most prominent Fashion Weeks worldwide take place
Web—Social networks twice a year and, as of 2015, are hosted in the cities of New
York, London, Paris, and Milan. These events facilitate net-
INTRODUCTION working and information sharing, and are seen as a crucial
The success of cultural artifacts is characterized by inher- part of the process by which the fashion industry collectively
ent unpredictability and inequality [40], posing fundamental decides what will be the new trends and the next top models.
problems for a proper understanding of markets based on the
production of cultural goods. Fashion, and fashion model- Social media and mobile image-sharing platforms, Instagram
ing in particular, are typical examples [2, 32]. When trying in particular, are revolutionizing the fashion industry world-
to cast a model for the upcoming seasons, a casting director wide, as interest toward new trends, designers, and products
is faced with a seemingly impossible task: predicting whom, increasingly unfolds online [25, 7, 26]. This has obvious im-
out of the hundreds of new faces she may see at the go-see plications on the job of fashion models too. Traditionally,
calls, will become the top model of the next season. models were not meant to interact directly with their cus-
tomers [33]. It is instead now customary for spectators to use
Instagram to upload photos or videos during runways events.
Permission to make digital or hard copies of all or part of this work for personal or This, in turn, has been argued to influence the way fashion de-
classroom use is granted without fee provided that copies are not made or distributed signers design, shoot, and showcase their runways, especially
for profit or commercial advantage and that copies bear this notice and the full cita-
tion on the first page. Copyrights for components of this work owned by others than for the case of luxury brands [41]. Famous designers such as
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re- Tommy Hilfiger and Kenneth Cole have been reported to take
publish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from [email protected].
advantage of Instagram for customer engagement [14].
CSCW’16, February 27–March 02, 2016, San Francisco, CA, USA.
Copyright © 2016 ACM. ISBN 978-1-4503-3592-8/16/02. . . $15.00
DOI: http://dx.doi.org/10.1145/2818048.2820065

64
CSCW '16, FEBRUARY 27–MARCH2, 2016, SAN FRANCISCO, CA, USA

As social media become for fashion models a far more im-

portant showcase than magazines and billboards, we wonder
if popularity on such platforms can be used as a proxy to pre-
dict success, and seek to answer these research questions:
RQ 1. Given data about measurable physical and profes-
sional characteristics of a models, can we predict whether
she will be casted for the upcoming fashion season?
RQ 2. Does the addition of relevant signals of social media
activity improve the predictability of success of models?
We tackle these questions using a quantitative approach, with
a mix of exploratory statistics and machine learning experi-
ments. To rule out possible explanations in terms of cumula-
tive advantage [42], we focus on data about a group of new-
comers, who just started their career in the fashion world. As
a simple and reasonable measure of success, we employ the
number of catwalks a model walked.
The rest of the paper is organized as follows. We give a brief
overview of the work related to predicting success in cultural
markets, and fashion in particular, in the next section. We
then describe the two datasets used in the study: the “new
faces” section of the Fashion Model Directory (FMD) web-
site1 and Instagram.2 We then present the main results of this
study. We start with a descriptive analysis of the FMD data,
and estimate the degree of association between the tenure of a
model and a number of standard industry metrics using a re-
gression framework. Finally, we describe the machine learn-
ing approach we employed, and the evaluation metrics used Figure 1: Example of a FMD profile page. Image © FMD –
to assess the quality of the predictions of our statistical frame- The Fashion Model Directory
work. The paper concludes describing the broader relevance
of our findings to the emerging “Science of Success” research
field, and potential future directions. sharing have received attention from the research community,
such as depression and online behavior [1, 43], situated usage
RELATED WORK in museums [21, 44], or organization of information through
In this study we are looking at predicting popularity of fash- tags [37, 27, 18].
ion models using data from social media activity. The study
of trends in cultural markets has a long-standing tradition [5], METHODS
and various researchers have exploited social data from a Fashion Model Directory
wide range of backgrounds. For example, Asur and Huber- We collected data from the Fashion Model Directory (FMD)
man [4, 3] used Twitter data to predict the box-office per- website, one of the largest fashion databases of professional
formance of newly-released blockbuster movies, and later female fashion models [45]. Figure 1 shows the profile of one
Mestyán et al. [35] improved these results using data from of the fashion models on FMD as an example. As can be seen
Wikipedia. Ferrara et al. [12, 13, 23] studied the emergence from the figure, FMD profiles, similarly to a résumé, provide
of information trends in social media settings, and the rise to a mix of biographical, physical, and professional experience
popularity of Instagram users in photography contexts [11]. information, notably casting agencies and walked runways.
In contrast, in the context of fashion, it is worth noting While the database claims to include profiles for over 10,000
that practical application of trend-detection technologies has models, our analysis focuses only on its recent additions,
started to become possible only in recent years, thanks to which are listed under the category “New Faces”. We col-
the increasing availability of online user-generated data about lected our dataset in December 2014, finding N = 431 new
fashion apparel trends [30, 29, 31, 20]. faces for the 2015 Spring / Summer (S / S) season.
Instagram was the platform selected in our study. It is a mo- For each model the FMD data thus consists of the follow-
bile image-sharing service that specializes in instant commu- ing attributes: name, hair color, eye color, height, hip size,
nication of trends and visual information in general [11], in dress size, waist size, shoes size, list of agencies, national-
a way similar to Twitter, Pinterest, and Flickr. Besides the ity, and details about all runways the model walked on (year,
detection of trends, a wide range of aspects related to image season, and city). We discarded the data about hair and eye
1
http://www.fashionmodeldirectory.com color, as the color coding was not reliable enough to allow for
2
http://instagram.com a meaningful characterization of these features. For similar

65
 SESSION: MODELING SOCIAL MEDIA

Top agency Non-top agency No agency Total

3
10
with Instagram 214 33 6 253
w/o Instagram 115 54 9 178
2
Total 329 87 15 431
Frequency

10

1 Table 1: New faces of the 2015 S / S season under contract

10
with at least one top agency and their presence on Instagram.
0
10
Instagram
0 10 20 30 40 50 60
We collected data about the social media presence of our
FMD new faces using Instagram. We found Ni = 253 In-
Walked Runways
stagram accounts (59%), accounting for Wi = 1181 runway
walks (4.67 walks per model on average). Models under
Figure 2: Distribution of walked runways by new faces of the contract with a top agency are more represented on Insta-
S / S prior to the Fashion Week season. gram (65%) than those with a non-top agency (41%), and no
agency at all (40%), see Table 1.
Using the media endpoint of the Instagram API, we collected
reasons, we also discarded nationality information. All body
all media posted by any FMD new face in the three-month
sizes (height, hips, dress, waist, shoes) were converted in the
period before September 4th, 2014, the beginning of the New
metric system. For shoe size we used the Paris point units sys-
York Fashion Week. Metadata of each posted media include
tem. Furthermore, the data were cleaned, substituting, in two
the number of likes and comments, as well as the the metadata
cases where the data about shoe and waist size were missing,
of the first 125 likes of each post (e.g., time stamp of the like,
the missing values with the group average. Before running
name of the liking user, etc.). We then computed the mini-
regressions, any non-categorical variable was also centered
mum, maximum, median, and average number of likes and
around the mean and standardized.
comments of all posts uploaded by each new face, as well as
To account only for a homogeneous set of runway expe- the number of posts during the period, for the three months
riences, we considered only runways occurred during the before and after the fashion week events. Similarly to the
fashion weeks for the 2015 S / S season in New York, Lon- FMD data, all variables were standardized before using them
don, Paris, and Milan. These occurred during the period in regressions.
of September 2014 and were the most recent major fashion
weeks at the time of our data collection (December 2014). Sentiment Analysis
We found that, collectively, the new faces in our sample had We supplement the analysis of social media activity with sen-
performed W = 1402 runway walks (3.25 walks per model timent analysis. To do so, we selected only comments writ-
on average) for B = 313 distinct branded runway events. ten in English. Language was detected using a simple Naive
Bayes classifier [36]. We extracted the comments on posts
Finally, we annotated each agency to reflect its reputation in uploaded before the Fashion Week season, and calculated the
the fashion industry. We use a simple binary classification average sentiment score of each model using Vader, a state-
system, by which high-prestige agencies are assigned to the of-the-art, rule-based algorithm [22]. We included only those
“top agency” category. Since FMD does not provide this in- models who received at least one comment written in English
formation, we retrieved a list of all top agencies from another to any of their posts, finding a subset of N s = 198 models,
online fashion database, Models.com,3 which collects experts who account for W s = 1052 runway walks (5.31 walks per
knowledge to determine the reputation and influence of cast- model on average). Vader is designed to deal with social me-
ing agencies. As a result, 329 out of 431 models were hired dia data, as it is based on a manually-defined vocabulary that
by at a least one top agency, 87 models were not hired by any encodes grammatical and syntactical conventions common to
top agency, and 15 had no agency information in their profile online documents. It is capable of capturing sentiment in-
(which could indicate that either the model was not using any tensity with an accuracy of 84%, which outperforms other
agency, or that the information was simply missing from the algorithms as well as individual human raters.
database; see Table 1).
The majority of new models did not perform even a single Predictive classification of success
runway during the four fashion weeks in September 2014; To forecast success within the new faces cohort, we per-
only around 24% of models (102 out of 431) performed at formed a binary classification exercise. Since most models
least one runway, with a small minority having participated in the cohort did not walk any runway, to avoid further class
to several runways (see Figure 2). This is not surprising, and imbalance we consider two classes: models with zero walks
further suggests that there exists a strong popularity bias in (unpopular) and models with one or more walks (popular).
castings, in line with previous work [19]. To learn the predictive score distributions, we applied three
widely-used machine learning algorithms, based on ensemble
methods and boosting: Decision Tree (DT) (baseline), Ran-
3
http://www.models.com. dom Forest (RF) [6], and AdaBoost (AB) [16]. We used the

66
CSCW '16, FEBRUARY 27–MARCH2, 2016, SAN FRANCISCO, CA, USA

Height (cm) Hips (cm) Dress Waist (cm) Shoes Height Hips Dress Waist Shoes
Mean 177.48 87.98 33.36 60.49 39.44 Height 1.00
Std. Dev 2.49 2.33 1.36 2.17 1.19 Hips 0.01 1.00
Min. 167.00 80.00 30.00 53.50 36.00 Dress −0.06 0.25 1.00
Median 178.00 88.00 33.00 60.00 39.00 Waist 0.02 0.58 0.28 1.00
Max. 183.00 104.00 40.00 77.00 43.00 Shoes 0.36 −0.01 0.00 0.07 1.00

Table 2: Body size measures of new faces of the 2015 S / S Table 3: Pairwise correlations between body size measures of
season (N = 431). the new faces of the 2015 S / S season (N = 431)

implementations of scikit-learn [38], optimally tuning the pa-

rameters as follows: entropy is used to measure the quality times for each additional standard deviation (2.49 cm, ap-
of the decision splits; all statistical models employ 25 esti- proximately one inch), relative to the group average baseline.
mators; DT and RF have a pruning setting of max-depth to 5, Larger dress, hips, and shoe sizes are all negatively associated
and RF adopts a maximum of 5 features. This framework al- with the chances of walking a runway, while waist size seem
lows us to evaluate the forecasting power of the various sets of to be not associated in either way.
features presented above using standard performance metrics,
such as AUROC (Area Under the Receiver Operating Char- In Table 3 we also report correlations between all body size
acteristic curve), and accuracy scores. We report results ob- regressors, as a simple test for possible source of multi-
tained by averaging one thousands iterations of k-Fold Cross collinearity. We find that dress, waist, and hip size are pair-
Validation (k = 5) in which 80% of data is used to train the wise correlated with each other, as well as shoe size with
statistical models and the remainder 20% is used for predic- height. All correlations appear to be of moderate entity.
tion. The last experiment of this paper, however, represents a Adding the information on agencies (see Model 2), we find
“real” prediction task in which we train the machine learning a strong association between having a top agency and the
models with the previous fashion season (2015 S / S) data, number of walked runways: models with a top agency have,
and use them to predict the upcoming 2015-16 Fall / Winter everything else being equal, nearly ten times higher chances
(F / W) season. (exp (2.29) = 9.87) of walking a runway, than their coun-
terparts represented by non-top agencies. The chances of
RESULTS walking for models without a prestigious agency drop sub-
stantially, as the expected count for the baseline is now only
Descriptive analysis
0.28 walks. This is consistent with previous research [19],
We first analyzed height, hips, dress, waist, and shoe size of and highlights the role of agencies in setting fashion models
the new faces group and assessed whether there is any obvi- trends in the fashion industry.
ous association with the number of walked runways. Fig-
ures 3(a)–(e) show the distribution of body size measure- We then focus on models with an Instagram account (Ni =
ments and, where available (Figures 3(a)–(c)) how these fea- 253) and assess how the average number of posted media,
tures compare to the US female population for the closest received likes, and comments is associated to the count of
age group (20–29 y.o. for height and waist size [17]; 18–25 walked runways. Figures 3(g)–(i) show distribution his-
y.o. for hip size [47]). Here the data have been plotted be- tograms (log10-scaled) of these variables.
fore rescaling. The variables are all distributed within narrow
Model 3 and Model 4 replicate the above findings on the
ranges, following approximately normal distributions. Un-
subset of new faces with an Instagram account. In partic-
surprisingly, even the shortest model in our dataset is much
taller than the US female average. In general, as far as body ular, within this subset the gap between those with a top
measures are concerned, our group of new faces seems to rep- agency and those without is even more marked. Adding the
Instagram-related variables does not change much the asso-
resent a very biased sample of the general female population.
ciation with height, hip size, and agency. Instagram activity
Table 2 reports standard descriptive statistics of the sample.
seem to have mixed associations with runway walks. Ad-
We study the association between body size measures and the ditional posts over the average activity yield a 15% higher
main dependent variable, the number of runways the models chances of walking a runway but, surprisingly, more likes
walked prior to the 2015 S / S fashion week season, using a tend to lower the chances of walking a runway (about 10%
regression framework. less). The average number of Instagram Likes is highly cor-
related with the average number of comments (r = 0.82)
Since our measure of success is based on the count of walked
yet it has a negligible correlation with the number of posts
runways, we used a Poisson regression model to estimate the
(r = 0.15). The number of received comments does not seem
chances of walking a single runway as a function of various
to be correlated with the number of posts (r = 0.00).
regression features. We start by only considering physical at-
tributes as regressors. The results are reported in Table 4 (see The overall picture does not change significantly when we
Model 1). The expected number of walked runways for the look at the sentiment expressed by Instagram users when
baseline new face is 2.25. Height is positively associated to commenting on the media posted by our new faces (Model
increased chances of walking a runway — specifically, 2.27 6, 7, and 8 of Table 4). The sentiment itself appears to be

67
 SESSION: MODELING SOCIAL MEDIA

160
140 a b c d
120
100
80
60
40
20
0
Frequency

160 165 170 175 180 185 50 60 70 80 90 100 80 85 90 95 100 105 30 32 34 36 38 40

Height (cm) Waist (cm) Hips (cm) Dress size
160
140 e g h i
120
100
80
60
40
20
0
36 37 38 39 40 41 42 43 0 100 200 300 400 1 2 3 4 5 0.0 0.5 1.0 1.5 2.0 2.5 3.0
Shoe size Instagram posts Inst. Avg. Likes (log10) Inst. Avg. Comments + 1 (log10)
Figure 3: Distribution of body size measures and Instagram activity of new faces of the 2015 S / S season (N = 431). Dashed
lines, where shown, indicate averages computed on the closest matching age group in the US female population.

positively associated to better chances of walking a runway Naive Bayes, etc.): none yielded AUROCs or accuracy above
(23% more), together with the overall number of comments. 60%.
Including the information about agency and all Instagram- To test the actual forecasting power of our framework based
related variables yields better statistical models for both the on the classifiers trained on the 2015 S / S data, we attempt
overall sample of new faces and those with an Instagram ac- to predict the popularity labels for the next season, the 2015-
count, as shown by both the Akaike (AIC) and Bayesian In- 16 F / W Fashion Week, that is, on a completely separate test
formation criterion (BIC) scores. This indicates that all vari- set. To do so, we manually collected a new and more re-
ables potentially provide useful predictive signals. In the next cent dataset (May 2015) containing the new faces of the lat-
section we describe the results of the prediction tasks in de- est fashion season, and up-to-date information about runways
tail. performed during the 2015-16 F / W Fashion Week (February
12–March 11, 2015). We found 15 such new face profiles.
This set is roughly balanced (8 fashion models ran at least
Forecasting success in fashion one top walk, 7 did not appear in any of the four main events),
For each classification algorithm (DT, RF, and AB) we and each profile links to an Instagram account, allowing us to
learned three distinct predictive models: (i) with only employ all predictive features (BODY + AGENCY + INSTA).
body size measures (height, hips, dress, waist, and shoes)
(BODY); (ii) with physical attributes and the binary infor- Social media features for the validation test set were built us-
mation about whether the fashion model has a top agencies ing the meta-data of media posted in the three months be-
or not (BODY + AGENCY); and, (iii) with body size measures, fore the season only (November 12, 2014 to February 11,
agency information, and Instagram-related signals —number 2015). The results for the best predictive model, Random
of posts, average number of likes and comments received— Forest (RF), along with the true popularity labels (became
(BODY + AGENCY + INSTA). For the latter statistical model, we popular), are shown in Table 6. Random Forest scores an
restrict the training data to use only the media posted in the AUROC performance above 81%: impressively, RF is able to
three months before the fashion week. As shown in Table 5, correctly predict 6 out of 8 fashion models who became popu-
and consistently with results from the previous section, when lar during the 2015-16 F / W, using training data from the past
trained on 2015 S / S runway walks data, social media fea- season only. Random Forest also successfully identified 6 of
tures improve accuracy of the statistical model. According to the 7 fashion models who did not perform in any top event.
t-tests, all improvements are statistically significant. We also The confusion matrix in Fig. 4 summarizes these results.
tried other statistical models [38] (SVM, Logistic Regression,

68
CSCW '16, FEBRUARY 27–MARCH2, 2016, SAN FRANCISCO, CA, USA

All new faces w/ Instagram w/ Instagram & Sentiment

(N = 431, W = 1402) (Ni = 253, Wi = 1181) (N s = 198, W s = 1052)
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7 Model 8
Intercept 0.81∗∗∗ −1.27∗∗∗ 1.21∗∗∗ −7.70∗ −7.70∗ 1.32∗∗∗ −7.59 −7.65∗
Height 0.82∗∗∗ 0.76∗∗∗ 0.89∗∗∗ 0.85∗∗∗ 0.88∗∗∗ 0.89∗∗∗ 0.86∗∗∗ 0.90∗∗∗
Dress −0.23∗∗∗ −0.17∗∗∗ −0.13∗∗∗ −0.05 −0.05 −0.16∗∗∗ −0.13∗∗∗ −0.12∗∗∗
Hips −0.35∗∗∗ −0.33∗∗∗ −0.24∗∗∗ −0.28∗∗∗ −0.29∗∗∗ −0.23∗∗∗ −0.33∗∗∗ −0.36∗∗∗
Waist −0.04 −0.00 0.06∗ 0.11∗∗∗ 0.10∗∗∗ 0.10∗∗∗ 0.18∗∗∗ 0.19∗∗∗
Shoes −0.38∗∗∗ −0.37∗∗∗ −0.35∗∗∗ −0.37∗∗∗ −0.36∗∗∗ −0.41∗∗∗ −0.44∗∗∗ −0.45∗∗∗
Has Top Agency 2.29∗∗∗ 9.05∗∗ 9.04∗∗ 9.03∗ 9.07∗∗
Inst. Posts 0.14∗∗∗ 0.08∗∗
Inst. Likes −0.18∗∗∗ −0.17∗∗∗
Inst. Comments 0.24∗∗∗ 0.21∗∗∗
Inst. Sentiment 0.16∗∗∗
AIC 4814.41 4531.28 3339.79 3064.96 3038.96 2734.13 2489.47 2462.77
BIC 1824.81 1545.74 1639.85 1368.52 1353.12 1426.88 1185.46 1171.92

Table 4: Poisson regression results for the new faces of the 2015 S / S season. Dependent variable is the count of runways walked.
Legend: ∗ : p < 0.05; ∗∗ : p < 0.01; ∗∗∗ : p < 0.001.

BODY BODY + AGENCY BODY + AGENCY + INSTA

ACC ROC ACC ROC ACC ROC

Decision Tree 0.596 0.558 0.635(+0.039)∗∗∗ 0.563(+0.005)∗∗∗ 0.694(+0.059)∗∗∗ 0.619(+0.056)∗∗∗

Random Forest 0.643 0.549 0.656(+0.013)∗∗∗ 0.586(+0.037)∗∗∗ 0.733(+0.077)∗∗∗ 0.688(+0.102)∗∗∗
Ada Boost 0.640 0.533 0.636(-0.004)∗∗∗ 0.556(+0.023)∗∗∗ 0.692(+0.056)∗∗∗ 0.640(+0.084)∗∗∗

Table 5: Accuracy (ACC) and Area Under the ROC curve (ROC) values for all classifiers. Increments for the classifier with all
features (BODY + AGENCY + INSTA) are computed over that without Instagram-related features (BODY + AGENCY), which is in turn
computed over the baseline (BODY). All improvements are statistically significant. Random Forest is the model with the best
predictive power, scoring a top accuracy of 73.3% and an AUROC of 68.8%. Legend: ∗ : p < 0.05; ∗∗ : p < 0.01; ∗∗∗ : p < 0.001.

0.86
Random Forest to fail. Fashion Model 10, the only false pos-
Unpopular

itive, on the other hand shows very high social media engage-
0.68
ment levels, yet did not perform in any top runway during the
2015-16 F / W season. The FMD profile pictures of the six
True label

6 7 (86 %) 1 7 (14 %)
models whose success was correctly predicted by our frame-
0.50 work are shown in Figure 5.
To understand which features contribute most to the predic-
Popular

0.32 tive signal we compute feature importance. The contribution

of each feature is here calculated as information gain. The
2 8 (25 %) 6 8 (75 %) results for Random Forest (RF) are shown in Table 7. For
0.14 prediction purposes, we note how the three social media ac-
Unpopular Popular
tivity features contribute as much as having optimal physi-
Predicted label cal attributes and more than being under contract with a top
agency. The top 3 features used by RF are: (i) Instagram
Figure 4: Confusion matrix of the Random Forest perfor- #Likes, (ii) Height, and (iii) #Posts. Similar results hold for
mance for the prediction task (2015-16 F / W Fashion Week). other classifiers.

DISCUSSION
Social media are increasingly used as sensors of social collec-
The elements of success in fashion modeling
tive phenomena [4, 13]. Increasingly, the usage of social data,
The analysis of the prediction task provides some interest-
often in conjunction with other data sources, proves crucial to
ing insights: Fashion Models 11 and 13, although having top
be able to represent real-world events, trends, information dif-
agencies, did not rise to popularity; Fashion Model 5, how-
fusion, and social behavior. In this study we were concerned
ever, became popular even without a top agency: both of with understanding whether it is possible to predict fashion
these dynamics have been correctly captured by our predic- models popularity, complementing physical and professional
tion framework. It is worth considering when our framework information with social data.
failed. Fashion Models 2 and 3 represent the two false neg-
atives: they both exhibit low social media activity, a signal Our methodology has of course limitations, and we here re-
highly regarded by our predictor (see below), which induces port few notable ones:

69
 SESSION: MODELING SOCIAL MEDIA

Fashion Height Hips Waist Dress Shoes Instagram Instagram Instagram Has top Became RF Prediction
Model ID Posts Comments Likes agency popular
1 178 86.5 58 33 41.0 59 4 148 True True True
2 178 86.0 60 33 40.0 24 0 32 False True False
3 179 88.0 61 34 39.0 0 0 0 True True False
4 180 89.0 60 34 41.0 52 1 93 True True True
5 175 86.0 58 33 38.0 163 2 70 True False False
6 180 89.0 60 34 41.0 2 7 48 True True True
7 180 90.0 61 34 40.0 10 3 116 True True True
8 178 87.0 61 33 39.0 34 2 90 True True True
9 183 86.0 62 34 41.0 16 2 61 True True True
10 176 87.0 59 33 38.5 17 17 647 True False True
11 177 86.0 60 32 38.5 38 2 51 True False False
12 180 90.0 60 33 40.0 29 1 59 False False False
13 169 88.0 60 33 38.0 49 9 570 True False False
14 179 94.0 65 35 41.0 58 3 52 False False False
15 180 83.0 62 35 43.0 11 15 546 False False False

Table 6: Performance of our predictive models trained on 2015 S / S data, and tested on 15 new fashion models who appeared
in the 2015-16 F / W Fashion Week. Our best classifier, Random Forest, correctly predicts 6 out of 8 positive instances (became
popular), and 6 out of 7 negative ones (not becoming popular) yielding 80% accuracy and an AUROC score of 81.25%.

(a) Fashion Model 1 (b) Fashion Model 4 (c) Fashion Model 6

(d) Fashion Model 7 (e) Fashion Model 8 (f) Fashion Model 9

Figure 5: FMD profiles of the six new faces whose success (having at least one runway during 2015-16 F / W Fashion Week) was
correctly predicted by our framework. All images © FMD – The Fashion Model Directory.

• All brands during the season are equally treated. Runways due to the intrinsic scale of fashion events and to our data
of higher reputation brands, such as Hermès or Chanel, sources, more data in the future will be needed to deter-
should be reflected with higher weights if compared to new mine the general performance of our framework.
and relatively unpopular brands. We plan to incorporate
such prestige in future revisions of our statistical models, • Our study is confined to one single online platform, Insta-
and observe what effects this yields. gram: its peculiar characteristics (e.g., the mobile-oriented
nature) might affect the dynamics of content generation
• Our measure of popularity only takes into account the num- and perceived popularity, as opposed to other platforms
ber of runways walked. This neglects several aspects of with different usage purposes, like information sharing
popularity within the fashion industry, such as appearance (Twitter [28]) or befriending activities (Facebook [9]).
in magazines and social events. We plan to incorporate
further dimensions of success in future work, to determine • Finally, our study is limited to analyze only female fash-
how these additional dimensions play along with the suc- ion, while male modeling is increasingly becoming more
cess measured by runways. mainstream. It will be interesting to see, when data be-
come available, whether our results apply to the male fash-
• Our “real” prediction task is tested on a very small dataset ion modeling market as well.
containing only 15 fashion models appeared during the
2015-16 F / W Fashion Week: although this limitation is

70
CSCW '16, FEBRUARY 27–MARCH2, 2016, SAN FRANCISCO, CA, USA

CONCLUSIONS ing her visibility above that of 76% of her competitors.4 Our
The ingredients to career success oftentimes remain myste- analysis aimed at understanding the factors that play a role in
rious. In the fashion industry, style is often credited as that obtaining such popularity, including physical attributes, the
ineffable quality all successful individuals have. The present reputation of casting agencies, and the importance of social
contribution shows how a number of seemingly disconnected media presence and reactions.
characteristics are actually tightly entangled: physical at-
Regarding our exploratory analysis (cfr. Table 4), we find that
tributes are required for inclusion in the modeling profession,
thinner and slender individuals are more likely to walk in run-
but do not suffice. The professional contribution of trend-
ways. Compared to the general population, models are often
setting top agencies play an equally important role. And, as
singled out for their extremely skinny and tall looks. How-
we first show in this paper, in the new era of social networks,
ever, it is interesting that even among themselves, these pref-
online presence helps succeed, as we see by the improvement
erences — towards skinny and tall models — are still signifi-
in the predictive power of our forecasting models.
cantly related to the number of runways they can join. Beauty
We submit a few possible explanation to this observation: is notoriously a hard-to-define quality and, in the case of the
in a world with limited attention [8], information cascades fashion industry, largely a by-product of a collective effort,
and the wisdom of the collectives are precious indicators for rather than an inherent quality [32]. While beyond the scope
casting agencies, promoters, marketeers, agents, recruiters, of the present work, an intriguing question that follows up
and the fashion industry in general. The response of the on- from it is whether Instagram and other social media are in-
line audience plays an increasingly important role in the of- deed changing the traditional notions of beauty.
fline fashion industry world: a rising star in the online world
Research on the fashion industry thus far has been largely
will hardly be ignored, and will probably be noticed by a top
qualitative, relying on methods such as interviews with small
agency, facts that will enhance her likelihood to succeed. In
number of models and casting directors [33, 19]. To the best
other words, buzz on social media is a proxy for the buzz in
of our knowledge, this is first time a large online fashion
the offline world, and this reduces uncertainty on the part of
database has been explored in a quantitative way, together
the industry.
with data from online social activity. As the impact of so-
Yet, it remains interesting that, in the regression models, in- cial media — especially Instagram — becomes significant in
creased activity on social media had only a weak association the fashion industry, predictive methods have the potential to
with heightened success (though on average fashion models leverage collective attention and the wisdom of the broader
with an Instagram account tended also to have done more user population, which reflect some of the popularity of fash-
shows). Perhaps, even these small differences have more ion models, to predict their career success.
chances of getting amplified due to word of mouth and collec-
Fashion modeling is one of the best examples of a cultural
tive attention, so that social media may be just facilitating the
market, like music, art, and literature. In all these markets,
information cascades mentioned before. Lacking data on the
determining quality of cultural products is hard because of
word of mouth among industry professionals, in this work we
inherent uncertainty, and thus market actors must rely on so-
did not investigate actual information cascades, but we be-
cial conventions and buzz as a proxy for success. In the case
lieve that further research is needed to better elucidate this
of fashion models, here we show that the buzz going on social
point.
media (Instagram in this case) is a reliable predictor of early
We also note how fashion modeling exhibits a strong winner- career success. Our results are in line with previous work that
takes-all component. In an industry that seems to be governed shows that social signals have a prominent role in determin-
by such a survival of the fittest mechanism, the difference be- ing success of cultural contents [40, 4], and so we can expect
tween performing a show in a premier venue or not becomes that similar approaches to cultural predictions will work in
crucial: while the majority of new faces will not appear in other markets too. Even scientific production and the stock
any prestigious avenue, having even one single runway in one market are, to some extent, ruled by prestige and buzz [34,
such venue may decree the success of a new model, bring- 15]. Thus we expect that the essence of our findings might
inform cultural producers and scholars well beyond the fash-
ion industry.
Feature Type Importance In conclusion, computer-mediated collectives are increas-
ingly disrupting the way culture is consumed and produced.
Height Physical 0.16
Dress Physical 0.05 Understanding how use of internet communication platforms
Hips Physical 0.09 affects cultural production is just an instance of the study of
Waist Physical 0.10 work in computer-mediated environments and an interesting
Shoes Physical 0.09 challenge for future research.
Has Top Agency Professional 0.05
Instagram Posts Social 0.16
Instagram Comments Social 0.13 ACKNOWLEDGEMENTS
Instagram Likes Social 0.18 The authors would like to thank Bria Carter, Chela Blunt,
and Johnny Villamil for their help with data coding; Mau-
Table 7: Feature importance (Random Forest model) to pre- reen Briggs, Alessandro Flammini, and Filippo Menczer for
dict fashion models’ success. 4
Only 24% of fashion models ran at least one top walk in our dataset.

71
 SESSION: MODELING SOCIAL MEDIA

useful feedback. GLC was supported in part by the Swiss Na- 13. Emilio Ferrara, Onur Varol, Filippo Menczer, and
tional Science Foundation (fellowship no. 142353) and the Alessandro Flammini. 2013b. Traveling trends: social
NSF (grant CCF-1101743). EF acknowledges the support by butterflies or frequent fliers?. In Proceedings of the first
DARPA grant W911NF-12-1-0034. ACM conference on Online social networks. ACM,
213–222.
REFERENCES
14. T. S. Fox. 2014. How Instagram is Changing Fashion
1. Nazanin Andalibi, Pinar Ozturk, and Andrea Forte.
Week. (February 2014). http://hypebeast.com/2014/
2015. Depression-related Imagery on Instagram. In
2/how-instagram-is-changing-fashion-week
Proceedings of the 18th ACM Conference Companion
[Online; posted 14-February-2014].
on Computer Supported Cooperative Work & Social
Computing. ACM, 231–234. 15. Georg Franck. 1999. Scientific Communication–A
Vanity Fair? Science 286, 5437 (1999), 53–55.
2. Patrik Aspers and Frédéric Godart. 2013. Sociology of
Fashion: Order and Change. Annual Review of 16. Yoav Freund and Robert E Schapire. 1995. A
Sociology 39, 1 (2013), 171–192. desicion-theoretic generalization of on-line learning and
an application to boosting. In Computational learning
3. Sitaram Asur, Bernardo A Huberman, Gabor Szabo, and
theory. Springer, 23–37.
Chunyan Wang. 2011. Trends in Social Media:
Persistence and Decay. In Fifth International AAAI 17. Cheryl D Fryar, Qiuping Gu, and Cynthia L Ogden.
Conference on Weblogs and Social Media. 2012. Anthropometric reference data for children and
adults: United States, 2007-2010. Vital and health
4. Sitaram Asur and Bernardo A Huberman. 2010.
statistics. Series 11, Data from the national health
Predicting the future with social media. In Web
survey 252 (2012), 1–48.
Intelligence and Intelligent Agent Technology (WI-IAT),
2010 IEEE/WIC/ACM International Conference on, 18. Eric Gilbert, Saeideh Bakhshi, Shuo Chang, and Loren
Vol. 1. IEEE, 492–499. Terveen. 2013. ”I Need to Try This”?: A Statistical
Overview of Pinterest. In Proceedings of the SIGCHI
5. Sushil Bikhchandani, David Hirshleifer, and Ivo Welch.
Conference on Human Factors in Computing Systems
1992. A theory of fads, fashion, custom, and cultural
(CHI ’13). ACM, New York, NY, USA, 2427–2436.
change as informational cascades. Journal of political
Economy (1992), 992–1026. 19. Frédéric C Godart and Ashley Mears. 2009. How do
cultural producers make creative decisions? Lessons
6. Leo Breiman. 2001. Random forests. Machine learning
from the catwalk. Social Forces 88, 2 (2009), 671–692.
45, 1 (2001), 5–32.
20. Lisa Green, Olivier Zimmer, and Yarden Horwitz. 2015.
7. Leah Cassidy and Kate Fitch. 2013. Beyond the catwalk:
Google Fashion Trends Report (U.S.). (2015).
Fashion public relations and social media in Australia.
https://www.thinkwithgoogle.com/features/
Asia Pacific Public Relations Journal 15, 1 (2013).
spring-2015-fashion-trends-google-data.html
8. Giovanni Luca Ciampaglia, Alessandro Flammini, and Accessed 2015-05-18.
Filippo Menczer. 2015. The production of information
21. Thomas Hillman and Alexandra Weilenmann. 2015.
in the attention economy. Scientific Reports 5 (2015),
Situated Social Media Use: A Methodological Approach
9452.
to Locating Social Media Practices and Trajectories. In
9. Pasquale De Meo, Emilio Ferrara, Giacomo Fiumara, Proceedings of the 33rd Annual ACM Conference on
and Alessandro Provetti. 2014. On Facebook, most ties Human Factors in Computing Systems (CHI ’15). ACM,
are weak. Commun. ACM 57, 11 (2014), 78–84. New York, NY, USA, 4057–4060.
10. Marie-Laure Djelic and Antti Ainamo. 1999. The 22. CJ Hutto and Eric Gilbert. 2014. Vader: A parsimonious
coevolution of new organizational forms in the fashion rule-based model for sentiment analysis of social media
industry: a historical and comparative study of France, text. In Eighth International AAAI Conference on
Italy, and the United States. Organization Science 10, 5 Weblogs and Social Media.
(1999), 622–637.
23. Mohsen JafariAsbagh, Emilio Ferrara, Onur Varol,
11. Emilio Ferrara, Roberto Interdonato, and Andrea Filippo Menczer, and Alessandro Flammini. 2014.
Tagarelli. 2014. Online popularity and topical interests Clustering memes in social media streams. Social
through the lens of instagram. In Proceedings of the 25th Network Analysis and Mining 4, 1 (2014), 1–13.
ACM conference on Hypertext and social media. ACM,
24. Yuniya Kawamura. 2004. Fashion-ology: an
24–34.
introduction to fashion studies. Berg.
12. Emilio Ferrara, Mohsen JafariAsbagh, Onur Varol,
25. Jan H. Kietzmann, Kristopher Hermkens, Ian P.
Vahed Qazvinian, Filippo Menczer, and Alessandro
McCarthy, and Bruno S. Silvestre. 2011. Social media?
Flammini. 2013a. Clustering memes in social media. In
Get serious! Understanding the functional building
2013 IEEE/ACM International Conference on Advances
blocks of social media. Business Horizons 54, 3 (2011),
in Social Networks Analysis and Mining. IEEE,
241–251.
548–555.

72
CSCW '16, FEBRUARY 27–MARCH2, 2016, SAN FRANCISCO, CA, USA

26. Angella J Kim and Eunju Ko. 2012. Do social media 38. Fabian Pedregosa, Gaël Varoquaux, Alexandre
marketing activities enhance customer equity? An Gramfort, Vincent Michel, Bertrand Thirion, Olivier
empirical study of luxury fashion brand. Journal of Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss,
Business Research 65, 10 (2012), 1480–1486. Vincent Dubourg, and others. 2011. Scikit-learn:
Machine learning in Python. The Journal of Machine
27. Peter Klemperer, Yuan Liang, Michelle Mazurek,
Learning Research 12 (2011), 2825–2830.
Manya Sleeper, Blase Ur, Lujo Bauer, Lorrie Faith
Cranor, Nitin Gupta, and Michael Reiter. 2012. Tag, you 39. Alexander Michael Petersen, Santo Fortunato, Raj K.
can see it!: using tags for access control in photo Pan, Kimmo Kaski, Orion Penner, Armando Rungi,
sharing. In Proceedings of the SIGCHI Conference on Massimo Riccaboni, H. Eugene Stanley, and Fabio
Human Factors in Computing Systems. ACM, 377–386. Pammolli. 2014. Reputation and impact in academic
careers. Proceedings of the National Academy of
28. Haewoon Kwak, Changhyun Lee, Hosung Park, and Sue Sciences 111, 43 (2014), 15316–15321.
Moon. 2010. What is Twitter, a social network or a news
media?. In Proceedings of the 19th international 40. Matthew J. Salganik, Peter Sheridan Dodds, and
conference on World wide web. ACM, 591–600. Duncan J. Watts. 2006. Experimental Study of
Inequality and Unpredictability in an Artificial Cultural
29. Yusan Lin, Heng Xu, Yilu Zhou, and Wang-Chien Lee. Market. Science 311, 5762 (2006), 854–856.
2015. Styles in the Fashion Social Network: An
Analysis on Lookbook.nu. In Social Computing, 41. Matthew Schneier. 2014. Fashion in the Age of
Behavioral-Cultural Modeling, and Prediction. Instagram. (April 2014).
Springer, 356–361. http://www.nytimes.com/2014/04/10/fashion/
fashion-in-the-age-of-instagram.html?_r=0
30. Yusan Lin, Yilu Zhou, and Heng Xu. 2014. The Hidden [Online; posted 9-April-2014].
Influence Network in the Fashion Industry. In The 24th
Annual Workshop on Information Technologies and 42. Arnout van de Rijt, Soong Moon Kang, Michael
Systems (WITS). Restivo, and Akshay Patil. 2014. Field experiments of
success-breeds-success dynamics. Proceedings of the
31. Yusan Lin, Yilu Zhou, and Heng Xu. 2015. National Academy of Sciences 111, 19 (2014),
Text-Generated Fashion Influence Model: An Empirical 6934–6939.
Study on Style.com. In 48th Hawaii International
Conference on System Sciences, HICSS 2015, Kauai, 43. Yiran Wang, Melissa Niiya, Gloria Mark, Stephanie M.
Hawaii, USA, January 5-8, 2015. 3642–3650. Reich, and Mark Warschauer. 2015. Coming of Age
(Digitally): An Ecological View of Social Media Use
32. Ashley Mears. 2011. Pricing beauty: The making of a Among College Students. In Proceedings of the 18th
fashion model. Univ of California Press. ACM Conference on Computer Supported Cooperative
33. Ashley Mears and William Finlay. 2005. Not Just a Work & Social Computing (CSCW ’15). ACM, New
Paper Doll How Models Manage Bodily Capital and York, NY, USA, 571–582.
Why They Perform Emotional Labor. Journal of 44. Alexandra Weilenmann, Thomas Hillman, and Beata
Contemporary Ethnography 34, 3 (2005), 317–343. Jungselius. 2013. Instagram at the museum:
34. Robert K Merton. 1988. The Matthew effect in science, communicating the museum experience through social
II: Cumulative advantage and the symbolism of photo sharing. In Proceedings of the SIGCHI
intellectual property. Isis (1988), 606–623. Conference on Human Factors in Computing Systems.
ACM, 1843–1852.
35. Márton Mestyán, Taha Yasseri, and János Kertész. 2013.
Early prediction of movie box office success based on 45. Wikipedia. 2014. Fashion Model Directory. (2014).
Wikipedia activity big data. PloS one 8, 8 (2013), http://en.wikipedia.org/w/index.php?title=
e71226. Fashion_Model_Directory&oldid=602195571 [Online;
accessed 20-April-2015].
36. Shuyo Nakatani. 2010. Language Detection Library for
Java. (2010). 46. Elizabeth Wissinger. 2009. Modeling Consumption:
http://code.google.com/p/language-detection/ Fashion modeling work in contemporary society.
Journal of consumer culture 9, 2 (2009), 273–296.
37. Oded Nov, Mor Naaman, and Chen Ye. 2008. What
drives content tagging: the case of photos on Flickr. In 47. Kate Zernike. 2004. Sizing Up America: Signs of
Proceedings of the SIGCHI conference on Human Expansion From Head to Toe. (March 2004).
factors in computing systems. ACM, 1097–1100. http://goo.gl/E8I8Co [Online; posted
1-March–2004].

73