"Woman-Metal-White vs Man-Dress-Shorts": Combining Social, Temporal and Image Signals to Understand Popularity of Pinterest Fashion Boards
““Woman-Metal-White vs Man-Dress-Shorts”:Combining Social, Temporal and Image Signals toUnderstand Popularity of Pinterest Fashion Boards
Suman Kalyan Maity ∗§ , Anshit Chaudhary ∗† and Animesh Mukherjee ‡ § Northwestern University; ‡ Dept. of CSE, IIT Kharagpur, India
Abstract
Pinterest is a popular photo sharing website. Fashion is onethe most popular and content generating category on thisplatform. Most of the popular fashion brands and designersuse boards on Pinterest for showcasing their products. How-ever, the characteristics of popular fashion boards are notwell-known. These characteristics can be used for predictingpopularity of a nascent board. Further, newly formed boardscan organize their content in a way similar to the popularfashion boards to garner enhanced popularity. What proper-ties on these fashion boards determine their popularity? Canthese properties be systematically quantified? In this paper,we show how social , temporal and image signals can togetherhelp in characterizing the popular fashion boards. In particu-lar, we study the sharing/borrowing behavior of pins and theimage content characteristics of the fashion boards. We ana-lyze the sharing behavior using social and temporal signals,and propose six novel yet simple metrics: originality score , retention coefficients , production coefficients , inter-copyingtime , duration of sharing and speed coefficients . We furtherstudy the image based content properties by extracting fash-ion , color and gender terms embedded in the pin images. Weobserve significant differences across the popular (highly fol-lowed or highly ranked by the experts) and the unpopular(less followed) boards. We then use these characteristic fea-tures to early predict the popularity of a board and achievea high correlation of . with low RMSE value. Our keyobservation is that likes and repin retention coefficients arethe most discriminatory factors of a board’s popularity apartfrom the usage of various color, gender and fashion terms. Introduction
Pinterest is an image-based online social network which hasgrown with unprecedented pace attaining a mark of 110 mil-lion monthly active users. It was also the fastest site to breakthe 10 million unique visitors mark . Although Pinterest isfairly new in the social media gamut, it is being heavily usedby many big business houses like Etsy, The Gap, Allrecipes,Jettsetter, Nike, Adidas etc. to advertise their products. Fur-ther, Pinterest drives more revenue per click than Twitter or ∗ Most of the work was done when the author was at IIT Kharag-pur, India.Copyright c (cid:13) https://techcrunch.com/2012/02/07/pinterest-monthly-uniques/ Facebook . This stupendous growth makes it interesting tostudy Pinterest. Fashion industry and social media
Social media is an amazing marketing tool for the fashion in-dustry. Fashion brands can leverage public perception avail-able on social media over various fashion items. The con-tinuous feedback received by the fashion brands in the formof likes and comments on their social media posts lets themgauge and further viralize their product chain in the mar-ket. Image-based social media platforms like Pinterest, In-stagram have become popular venues for fashion brand mar-keting and advertising.
Role of Pinterest in fashion industry
Fashion is an integral part of Pinterest. All the major brandslike Nike, Adidas Originals, Dolce Gabbana, Louis Vuittonetc. have their presence on the Pinterest platform. What arethe characteristic features of these popular brands? Do theybear certain signatures – social, temporal or image based –that make them distinct from the not-so-popular ones?In Pinterest, users save images ( pins ) and categorize themon different boards . Thus, a board is an important entityin Pinterest, and it has various influences on interest-drivenpin propagation or pin sharing. Sharing is an important as-pect in social media. If one likes a content, one might tendto use it in the same/modified form. On Pinterest, shar-ing and borrowing of pins (images) from various boards isa routine phenomena. This motivates us to consider shar-ing/borrowing behavior in understanding the popularity ofthe fashion boards. Similarly, we hypothesize that the (im-age) content of the post should also be a key factor deter-mining the popularity of a fashion board.There are multiple boards where Pinterest advertises var-ious fashion contents. In this paper, we study the popular-ity of fashion boards by analyzing their originality, shar-ing/borrowing behavior, and the characteristics of the imagecontent.
Research objectives and contributions
In this paper, we analyze a massive Pinterest dataset consist-ing of pins and boards and make the following contributions. https://en.wikipedia.org/wiki/Pinterest a r X i v : . [ c s . S I] D ec We investigate three major factors that can potentiallycharacterize popularity of a fashion board: social , tempo-ral (i.e., how pins are shared or borrowed across boards)and image content characteristics (usage of fashion, colorand gender terms etc. in the image) of the pins belongingto a board. • We observe that generally popular fashion boards areable to make an existing non-popular pin popular ,whereas less popular fashion boards do not exhibit thischaracteristic. Note that this is a very non-intuitive find-ing indicating a non-assortative behavior (the ‘popular’pins making the non-popular pins popular) as opposed towhat is usually observed in most social networks. • Another key observation is that same content in differentboards achieve different levels of popularity . If a pin hasoriginated from a popular board, it achieves higher popu-larity on the originating board than the subsequent boardsto which it gets shared possibly pointing to dampeningof the popularity due to re-sharing. In addition, pins keepgetting shared for longer durations in popular boards . • We perform extensive image analysis of the pins on theboards and extract fashion, color, and gender terms fromthem. Popular fashion boards have more female faces thanthe unpopular ones. Further, the popular boards have arich collection of pins in which both the gender co-appear .We also observe significant characteristic differences be-tween popular and unpopular boards in the usage of colorand fashion words . • Our characterization further helps us to predict whethera given fashion board would become ‘popular’ or not. Inprecise, we attempt to predict the popularity in terms ofthe future number of followers of the boards. We achievea very high correlation coefficient of 0.874 with very lowRMSE. A post-hoc analysis of the importance of the fea-tures indicates that the likes and the repin retention coef-ficients are the most discriminative ones followed by thecolor and gender terms embedded in the image.
Lessons for newbie fashion houses
The insights gained from this work can highly impact thenew and upcoming fashion brands. For instance, allowingfor more female faces or both male and female faces to-gether, certain color terms (white, black, blue, brown, pinketc.) and color combinations (blue-pink, black-pink etc.) canincrease the chances of the boards getting popular. Also theycould ‘engineer’ campaigns to promote their boards in sucha way that the originating boards are able to retain the ‘likes’and ‘repins’ of their pins in the face of constant sharing ofthese pins.
Related work
Content characteristics, sharing, and engagement
Content sharing ensures user engagement and commitmentin future (Burke, Marlow, and Lento 2009). There are di-verse motivations to share content on social media (Lee andMa 2012). Apart from network structure, the content matter also play important role in sharing. Users may share use-ful content to appear knowledgeable or simply to help outothers (Wojnicki and Godes 2008). The emotional valencebehind content also drive its extent of being shared (Jamaliand Rangwala 2009; Berger and Milkman 2012).Though there have been various studies on diffusion, shar-ing, engagement in social media, very less work has beendone in the domain of visual analysis of image content.Hochman et al. (2012) show differences in local color us-age, cultural production rate, varied hue intensity (blue-grayin New York vs red-yellow in Tokyo) by analyzing imagesfrom New York and Tokyo posted on Instagram. Bakhshiet al. (2014) study the engagement characteristic of imagescontaining human faces. They observe that images with hu-man faces in them, have higher chances of receiving likesand comments. Bakhshi and Gilbert (2015) study the role ofcolor in online diffusion of pins in Pinterest. They observethat color significantly impacts the diffusion of images andadoption of content. Red, purple and pink seem to promotediffusion, while green, blue, black and yellow suppress it.
Popularity:
There have been few studies in the domain ofFashion trend and popularity. Sanchis-ojeda et al. (2016) ex-plore various statistical models using clients’ temporal reac-tion to style units change for identification and quantifica-tion of linear and cyclical fashion trends. Lee et al. (2017)propose a classifier for identification of fashion-related Twit-ter accounts whereas we focus on understanding the popularfashion boards. Hessel et al. (2017) propose a relative popu-larity prediction framework based on content characteristicswith minimal influence of other external factors like timingeffects, community preferences, and social networks. Wu etal. (2017) study the sequential prediction of popularity forimage posts using a deep learning framework by incorporat-ing temporal context and temporal attention into the frame-work.
Fashion brand marketing
Yamaguchi et al. (2014) study the effects of visual, textual,and social factors on the popularity in a large real-world net-work focused on fashion. There are several studies that fo-cus on understanding the growing interest in social mediamarketing (Dubois and Duquesne 1993; Kim and Ko 2012;Kim and Ko 2010). Manikonda et al. (2015) study the in-fluence of social media in various behavior of fashion brandmarketing. They also analyze fashion brands’ audience re-tention and social engagement.
Color in affective marketing:
There are several researchworks that have studied the role of color in affective mar-keting. Most of these works focus on various kinds of adver-tisements, for example, the research on role of specific col-ors used in magazine ads (Lee and Barnes 1989; Schindler1986), the efficiency of color ads compared to black andwhite ads (Meyers-Levy and Peracchio 1995; Sparkman Jrand Austin 1980). There are also studies on understand-ing the effects of colors on consumer responses (Bellizzi,Crowley, and Hasty 1983; Crowley 1993). This line of re-search suggests that red backgrounds elicit greater feel-ings of arousal than blue ones, whereas products presentedagainst blue backgrounds are liked more than products pre-ented against red ones (Bellizzi, Crowley, and Hasty 1983;Middlestadt 1990). Gray et al. (2014) study the relationshipbetween color coordination and ‘fashionableness’. They ob-serve that maximum fashionableness is attained by selectinga color combination that is neither completely uniform, norcompletely different, i.e., fashionable outfits are those thatare moderately matched, not those that are ultra-matched(“matchy-matchy”) or zero-matched (“clashing”). This bal-ance of extremes supports Goldilocks principle regardingaesthetic preferences that seeks to balance of simplicity andcomplexity.
Studies on Pinterest
There have been several works on Pinterest.
Gender Roles:
Gilbert et al. (2013) perform analysis focus-ing on the influence of gender, geography, language usageon Pinterest. They identify features of pins that could predictthe activity of the board. They observe that being female onthe site leads to more repins while having fewer followers.Our work draws motivation from this work where we studyvarious social, temporal and image content factors as drivingfactors of popularity. Ottoni et al. (2013) study differencesin gender role in platform usage and social interaction onPinterest. They observe that females invest more effort inreciprocating social links, are more active and generalist incontent generation whereas male are more likely to be spe-cialists and tend to describe themselves in an assertive way.Also men and women possess different interests. Chang etal. (2014) study users’ topical specialization and homophily.
User interaction and experience : Zarro et al. (2013) in-vestigate professional and personal uses of Pinterest withinterview data and observations of online activity. Zhonget al. (2014) perform analysis on how borrowing behav-ior facilitates social interactivity and experience. Yang etal. (2015) study recommendation of Pinterest boards for theTwitter users. Linder et al. (2014) investigate social and cog-nitive aspects of creativity that affect the digital curationpractices of everyday ideation with Pinterest users. Milleret al. (2015) study perception on Pinterest of the users andnon-users and show that there exist differences among thesetwo groups and how exploring Pinterest changes the non-users’ experience.
Diffusion, Popularity:
Zhong et al. (2016) study the im-pact of social ties on Pinterest. Lo et al. (2016) study useractivity and purchasing behavior on Pinterest for character-ization of temporal user purchase intent. Lo et al. (2017)in another paper characterize the growth of Pinterest boards(size) and analyze how initial growth can be used to predictfuture growth behavior. Han et al. (2017) study popular andviral image diffusion in Pinterest. Deeb-Swihart et al. (2017)study selfie presentation in everyday life on Instagram. Youet al. (2017) study various spatio-temporal patterns of Face-book photographs as well as its diffusion pattern via socialties.
The present work:
Our study is different from the aboveones in that we analyze the popularity aspects of (fashion)boards and attempt to understand its relationship with (i) so-cial and temporal sharing/borrowing behavior and (ii) thegender, fashion and color terms embedded in the images posted on these boards. Our analysis sheds light on strategiesand mechanisms an upcoming fashion brand could adopt tomake itself popular on Pinterest which can eventually en-hance their overall business.
Pinterest terminologies and the dataset
Entities on Pinterest
There are several entities on Pinterest. We shall provide abrief discussion on the essential functional entities of Pin-terest platform. • Pin : A pin (analogous to a post in Facebook or a tweetin Twitter) is an image which is forms the basic buildingblock of Pinterest. Pin is a visual bookmark. Each Pin onecan see on Pinterest site links back to the website it camefrom. The activity related to posting a pin is known as‘pinning’, and the user who posts a pin is known as the‘pinner’. Pins can be liked and shared. Each of these pinshas the following meta-data associated with it - uniquepin-id, description, number of likes, number of comments,number of repins, board name, source, and content of thecomments. Sharing an already existing pin is referred toas ‘repinning’ (similar to retweet in Twitter). • Board : A board is a user-generated collection where onesaves pins. Boards can be made in secrecy or publicly.One can add collaborators to boards. Each board has aurl, a name, a description (optional) and a category (op-tional, e.g. Art, Architecture, Celebrities, Food and Drink,Entertainment, Education, Fashion etc.). This analogy ofpins and boards replicates the real-world concept of clas-sifying images into photo albums.
Dataset
The dataset used in this study contains information about 0.3million boards and their 63 million pins. We use PinterestAPI v1 to crawl information about boards and pins. Boardinformation constitutes of the following: board description,number of followers, and the creator. Pin information has thefollowing attributes: pin description, number of likes, num-ber of comments, number of repins, board name, and thecreator. The data collection process is divided into two partsas follows
Crawling of the massive dataset
We crawl a large datasetwhich should be useful for doing various analysis of thefashion boards. • Initial pin collection : We initiate the data collection pro-cess by obtaining the pin-id of 1000 pins from by generating au-tomatic scrolls. Now, each pin-id of these pins are pickedand the trailing 6 digits were permuted to generate newpin-ids. About 10 million new pin-ids are generated bythis process. Information of all these pins are crawled sep-arately. • Massive information collection : We extract the board-urlof each of these pins from the information crawled above.About 0.3 million unique board urls are obtained. Now,or all the board-urls, board information and their indi-vidual pin’s information are crawled. This results in 59million unique pins out of a total of 63 million pins. Thismassive dataset is used to find out the origin of the pins.
Fashion boards dataset
We extract names of fashionboards from the following sources: i) Fashion categories onPinterest and ii) Expert rankings from Ranker , Mashable and Stylecaster . Finally, we could obtain information about ∼ fashion boards. We discard those boards that havevery less number of pins. A total of 4 million pins are foundon these 3600 boards. For each of these board urls, we crawlthe detailed information about the board and its pins fromFeb, 2016 till March, 2016.We then further categorize the 3600 boards above into fol-lowing two categories of popularity: popular boards and un-popular boards. In popular boards we define two popularityclasses - Highly Followed (socially ranked) Boards (HFB)and Expert Ranked Boards (ERB). We denote the unpopu-lar ones as the Less Followed Boards (LFB). From the col-lection of the fashion boards, we assume the top 20% mostfollowed boards as HFB and the bottom 20% as LFB. Wehave tried to use other percentage values also but choosing20% allowed us to have a sizeable data for conducting mean-ing experiments. The 1200 boards which we obtain from theexpert rankings are noted as expert ranked boards (ERB). Characterization of fashion boards
In this section, we shall discuss the various factors whichcharacterize the popularity of fashion boards. There are sev-eral factors that governs the popularity of a board - the orig-inality/novelty, sharing/borrowing behavior as well as thecontent on the board (the image-characteristics of the pins).
Originality
Originality/novelty of boards is an important aspect. If oneobserve the creation of pins over the years (see figure 1),one can conclude that the total no. of pins are continuouslygrowing whereas the no. of unique pins have increased onlyin the early few years but then started decreasing. This indi-cates that originality in this social media is on a decline overtime due to heavy content sharing. Motivated by figure 1, westudy board originality as an indicator of popularity.
Pin originality:
Pins on a board can be classified into twotypes: original pins and duplicate pins. If a pin has origi-nated from the board b , then it is called an original pin withrespect to the board b whereas, if a pin has not originatedfrom the board b , but is a result of a copy from another boardto b , then it is called a duplicate pin with respect to b . Board originality score:
Using the concept of pin original-ity, we define a measure to compute the originality score of http://mashable.com/2012/08/06/top-fashion-pinterest-accounts http://stylecaster.com/fashion-pinterest-accounts Figure 1:
Evolution of unique (i.e., absolute) and total number ofpins created per year. a board. Originality score ( orig score ) of a board ( b ) can bedefined as the ratio of the original pins ( o b ) on it to the totalnumber of pins ( t b ) on it. orig score ( b ) = o b t b Originality score of a board lies in interval [0 , . Boardshaving originality score close to 1 constitute of mainlyoriginal pins, which means that they are content genera-tors. Boards having originality score close to 0 constituteof mainly duplicate pins, which means that they are contentcopiers/consumers.In figure 2(a), we observe that the originality scores arehighly correlated with follower count. We then group theboards in less followed, highly followed and expert rankedboards and measure the originality scores in these popular-ity buckets. We observe that originality scores of highly fol-lowed and expert ranked boards are high, whereas that ofless followed boards are on the lower side. Thus, originalityof a board is an important indicator of its popularity. LFB, HFB: ****LFB,ERB: ****
Originality Score F r a c t i o n (a) (b) Originality Score F o ll o w e r c o u n t ( L o g ) Figure 2: (a) Relationship between originality scores and followercounts of boards. (b) Distribution of originality scores across theless followed, highly followed and expert ranked boards. The K-Stest for significance among the relevant distributions are measured.****,***,**,*, ns denote p -values of significance to be < < < < Originality of the top fashion brands
We further studythe originality scores of the boards corresponding to the topfashion brands. Toward this objective, we consider the topfashion clothing brands and attempt to compute their orig-inality scores. We separately collect the board information nd the pins of these top fashion clothing brands. We com-pute the originality score of these boards and observe thatthey have highly original content (see table 1).We then attempt to find the extent of correlation betweenthe originality scores of these boards with their popularity(in terms of the number of followers). The Spearman’s rankcorrelation comes out to be . . This establishes that thereis a strong positive correlation between originality and pop-ularity of the top fashion brands. Table 1:
Originality scores of top fashion brands
Rank Brand Name Originality Score1 Nike 0.8933334940642 Target 0.995608131343 Adidas 1.04 Macy’s 0.985390778175 JCPenney 0.9928222030766 Converse 0.8772293317427 Van’s 0.9904414632518 Ralph Lauren 0.9968354430389 Forever 21 0.95384726423210 Victoria’s Secret 0.98873243246211 Levi’s 0.94462846532412 Chanel 0.87640755639413 Under Armour 0.89765342723214 Aeropostale 0.916352963232
Sharing/borrowing behavior
Sharing/borrowing of pins are very common on the Pin-terest platform. We introduce board retention coefficients and board production coefficients based on the shar-ing/borrowing behavior dynamics of the pins on a board.On Pinterest, the ‘social behavior’ of a pin can be measuredbased on three factors: the number of likes , the number ofrepins and the number of comments generated by the pin.We however observe that commenting is not practiced exten-sively in this platform. Hence, we only take number of likes and repins generated by a pin. The two coefficients we de-fine next roughly correspond to the direction and magnitudeof flow of information from one board to another. Each ofthem is independently able to portray meaningful informa-tion about sharing.
Retention coefficients
Board retention coefficients are anovel set of measures concerning the like / repin ‘retention’capabilities of a board. It addresses the question that - howmany likes/repins shall a board be able to retain if otherboards copy content from it . We calculate the likes retentioncoefficient using the algorithm 1. Similarly, we also computethe repins retention coefficient . Algorithm 1
Calculation of likes retention coefficient .temp ← [ ] for each original pin p on board b do temp.append ( 1 + likes of p on b avg. likes of p on other boards ) end for likes retention coefficient of board b = average(temp) Both the retention coefficient values lie in the interval (1 , ∞ ) . If a board b has a higher likes (repins) retention co-efficient, then the subsequent boards that copy pins from thisboard shall be able to garner less likes (repins) than the board b . LFB,HFB: ****LFB,ERB: * LFB,HFB: ****LFB,ERB: ***LFB,HFB: ****LFB,ERB: ****LFB,HFB: ****LFB,ERB: ****(a) (b)(c) (d)
Figure 3:
Distribution of a) likes retention coefficient b) repinsretention coefficient (c) likes production coefficient b) repins pro-duction coefficient for less followed, highly followed and expertranked boards.
A significant fraction of highly followed and expertranked boards have higher retention coefficients comparedto the less followed boards (see figure 3(a) and (b)). Thus,the original pins on highly followed and expert rankedboards are significantly more liked/repinned among their re-spective duplicate (shared) pins, whereas the original pinson less followed boards are not popular among their dupli-cate pins. Hence, the likes/repins of the content on highlyfollowed and expert ranked boards do not decline even afterthey are duplicated through copying.
Production coefficients
Board production coefficients area novel set of measures that capture the like/repin produc-tion capacities of a board. It addresses the question that - how many likes/repins shall other boards gain if they copycontent from a board . We compute the likes production co-efficient using the algorithm 2. Similarly, we also compute repins production coefficient . Algorithm 2
Calculation of likes production coefficient .temp ← [] for each duplicate pin p on board b do temp.append ( 1 + likes of p on its original board likes of p on b ) end for likes production coefficient of board b = average(temp)Both the production coefficients lie in the interval (1 , ∞ ) .If a board b has a lower likes (repins) production coefficient,then the duplicate pins on b shall garner more likes (repins)compared to that on their board of origin.A vast majority of highly followed and expert rankedboards have lower production coefficients than the less fol-lowed boards (see figure 3 (c) and (d)). Thus, the duplicateins on highly followed and expert ranked boards gener-ate more number of likes/repins compared to that gener-ated on their corresponding boards of origin. On the otherhand, duplicate pins on less followed boards generate lesslikes/repins compared to that in their corresponding boardsof origin. Hence, highly followed boards and expert rankedboards are able to make an existing pin more liked/repinnedby copying it. Temporal dynamics of sharing/borrowing
In this section, we introduce two measures based on the tem-poral aspects of sharing: inter-copying time and duration ofsharing . In addition, we also define speed coefficients thatindicate the speed at which likes/repins are gained.
Inter-copying time is a measure defined for the original pins. Thisis expressed as the average time-gaps between instances ofsharing of an original pin on the subsequent boards. For anoriginal pin p , we compute ICT as explained in algorithm 3.We now average the value of
ICT s for all the original pinson board b , and call this as inter-copying time of board b .A significant number of pins belonging to highly fol-lowed and expert ranked boards have a higher value of inter-copying time than pins on less followed boards (see fig-ure 4(a) and (b)). This shows that pins on less followedboards have smaller time-gaps between consecutive sharesas compared to pins on highly followed and expert rankedboards. Algorithm 3
Calculation of inter-copying time for an orig-inal pin p . temp ← [] for each duplicate pin p (cid:48) generated from pin p do temp .append(time-stamp of p (cid:48) ) end for sort temp in non-decreasing order for each i in range(0, len( temp )) doif i == 0 then temp [ i ] ← else temp [ i ] ← temp [ i ] - temp [ i − end ifend for ICT for pin p ← average ( temp ) Duration of sharing (DoS)
Similar to inter-copying time , duration of sharing is also defined for original pins. It canbe interpreted as the life-cycle of sharing of a pin. For anoriginal pin p , we compute DoS as explained in algorithm 4.We now average the value of
DoS s for all original pins onboard b , and call this as duration of sharing of board b .A large fraction of pins on highly followed boards and ex-pert ranked boards have high duration of sharing comparedto the less followed boards (see figure 4(c) and (d)). Hence,a pin on highly followed and expert ranked boards is likelygoing to have a longer life span than a pin on the less fol-lowed boards. Algorithm 4
Calculation of duration of sharing for an orig-inal pin p . temp ← [] for each duplicate pin p (cid:48) generated from pin p do temp .append(time-stamp of p (cid:48) ) end for sort temp in non-decreasing order DoS for pin p ← temp [ len ( temp ) − − temp [0] Speed coefficients
In this section, we attempt to combinethe likes/repins on a board with its temporal characteristics.Toward this objective, we define likes and repins speed co-efficients as follows.we compute likes speed coefficient as explained in algo-rithm 5. We then average the value of likes speed coefficient for all the original pins on a board b , and call this as the likesspeed coefficient of board b . We similarly calculate repinsspeed coefficient . Algorithm 5
Calculation of likes speed coefficient for anoriginal pin p . likes ← [] for each duplicate pin p (cid:48) generated from pin p do likes .append(number of likes on p (cid:48) ) end for likes speed coefficient of p ← sum ( likes ) DoS ( p ) Speed coefficients are greater for highly followed and ex-pert ranked boards compared to the less followed boards (seefigure 4(e) and (f))). Thus, original pins on highly followedand expert ranked boards gain popularity much more quicklythan the original pins on less followed boards.
Image-based content analysis
Pinterest being an image sharing social media, the charac-teristics of image (pins) also should have an impact on theirpopularity. In this section, we analyze the content character-istics of images (pins) on the various boards. Toward thisobjective, we use densecap (Johnson, Karpathy, and Fei-Fei 2016), an image captioning tool that extracts salient re-gions from an image and describes them in natural language(English). We perform tokenization, stemming, lemmatiza-tion and stop-words removal of these generated ‘dense’ cap-tions to obtain key tokens/phrases demonstrating the image.We further group these tokens into three key types: genderterms , fashion terms and color terms . Gender terms whichwe analyze are male and female . We obtain an exhaustivelisting of fashion terms from Myvocabulary . A universal setof all color terms are available in Wikipedia . The fashion https://myvocabulary.com/word-list/fashion-and-clothing-vocabulary/ https://en.wikipedia.org/wiki/Lists_of_colors FB,HFB: ****LFB,ERB: **** LFB,HFB: ****LFB,ERB: ****LFB,HFB: ****LFB,ERB: ****LFB,HFB: ****LFB,ERB: **** LFB,HFB: nsLFB,ERB: **** LFB,HFB: nsLFB,ERB: ****(a) (b) (c)(d) (e) (f)
Figure 4:
Distribution of a) inter-copying time for pins b) inter-copying time for boards c) duration of sharing for pins d) duration of sharingfor boards e) likes speed coefficient f) repins speed coefficient for pins in less followed, highly followed and expert ranked boards. and color terms which we find both in the above respectivelistings and our data-set are shown in table 2 and 3.
Table 2:
Fashion terms in the dataset. clothes tshirt skin shirts jacketfeathers pillows sunglasses buttons shoesuit curtains skirt leather pantstrouser striped shorts strap dressjeans pillow necklace umbrella bag
Table 3:
Color terms in the dataset. black blue white brown green purple redyellow grey metal wooden pink silver
Gender term analysis
We analyze the occurrences ofboth the genders in the pins across all the three categories ofboards. In table 4, we report the occurrences of each gender.The cell corresponding to
Male and
Less Followed Boards has a value of 0.45. This means that 0.45 fraction of pins be-longing to less followed boards have male faces on them. Weobserve that the number of female faces on highly followedboards and expert ranked boards is high, whereas they aresignificantly lower ( ∼ lower) in less followed boards.This shows that having more female faces on a board couldincrease the popularity of a board. Further, if a board hasmore female faces, it has more chances of being listed in ex-pert ranked boards. Another very interesting observation isthat the more popular boards have a higher fraction of pinscontaining both male and female faces together on a singlepin compared to the less followed boards.In summary, highly followed boards and expert rankedboards have more female faces than the less followed boards.Moreover, the boards in the former category have a richercollection of pins that together feature faces of the genders. Table 4:
Fraction of pins on boards with various gender combina-tions.
Gender LFB HFB ERBMale only 0.45 0.48 0.49Female only 0.49
Male-Female 0.26
Fashion term analysis
We analyze the occurrences of var-ious fashion terms in table 2 in the imagery of the pins acrossall the three categories of boards. We compute number of oc-currences of the fashion terms from table 2 appearing in thepins across the boards. We use the torso of this frequencydistribution to extract the most discerning fashion terms. Wechoose the top 10 fashion terms from the torso to computeresults in table 5. Each cell ( x, y ) in the table represents thefraction of pins belonging to a particular popularity class‘y’ having the fashion term ‘x’. Therefore, the cell valueof 0.318 corresponding to shirt and Less Followed Boards means that 0.318 fraction of pins belonging to less followedboards have the fashion term shirt in them. We observe that20% of fashion terms have equal distributions in highly fol-lowed boards and expert ranked boards whereas their distri-butions in the less followed boards are quite different.We further study the co-occurrences of fashion terms inmore detail. In table 6, we report the number of occurrencesof two co-occurring fashion terms in a pin. Each cell ( x - z, y ) in the table represents the fraction of pins belonging to aparticular popularity class ‘y’ having the co-occurring fash-ion term ‘x-z’. Thus, the value of 0.0964 corresponding to jacket-trouser and Less Followed Boards means that 0.0964fraction of pins belonging to
Less Followed Boards haveboth jacket and trouser together in them. We observe that co-occurring bi-terms have almost similar distributions able 5:
Fraction of pins having top 10 fashion terms which are ob-tained from the torso of the frequency distribution of all the fashionterms in table 2. The boldface values indicate similar distributionamong
ERB and
HF B but a different distribution in LFB. Weadopted this convention in the subsequent tables also.
Fashion Terms LFB HFB ERBshirt 0.318 0.389 0.330 bag 0.221 0.291 0.312 dress 0.221 0.179 0.225pants 0.210 0.154 0.216shoe 0.187 0.231 0.201jacket 0.159 0.148 0.139umbrella 0.158 0.139 0.146necklace 0.147 0.134 0.156pillow 0.143 0.131 0.153 jeans 0.092 0.127 0.152 in highly followed boards and expert ranked boards, whereastheir distributions in the less followed boards are very differ-ent from the other two.
Table 6:
Fractions of pins having top 10 co-occurring bi-terms,which are obtained from the torso of the distribution of the fashionterms in table 2.
Bi-terms LFB HFB ERB jacket-trouser 0.0964 0.1276 0.1198bag-umbrella 0.0753 0.1243 0.1134bag-striped 0.0683 0.1223 0.1143 necklace-strap 0.0879 0.1124 0.0953 jeans-shoe 0.0762 0.1057 0.1049 bag-shorts 0.1032 0.1242 0.0643 bag-trouser 0.0923 0.1243 0.1142leather-strap 0.0923 0.1214 0.1203dress-umbrella 0.0812 0.1143 0.1043 dress-skirt 0.1023 0.0854 0.1053
We also consider the co-occurring tri-terms (three fash-ion terms together). We observe similar discriminating re-sults for tri-terms (see table 7). The discrimination be-comes more prominent when we use the co-occurring bi-terms and tri-terms. We thus conclude that the collection offashion terms together used in pins affect the popularity oftheir boards. Hence, a popularity seeking board can host im-ages having a particular collection of fashion terms from theabove analysis.
Color term analysis
Colors are a very important factorin fashion (Bakhshi and Gilbert 2015). In this section, weanalyze the occurrence of color terms from table 3 appear-ing in the pins across all three categories of boards. Wecompute the number of occurrences of each color term, co-occurring bi-terms generated from the colors in table 3. Wechoose the top 10 color terms (once again, from the torso ofthe frequency distribution) in table 8. We observe that 30%color terms have similar distributions in the highly followedboards and expert ranked boards, whereas their distributionsin the less followed boards are different from the other two.White, black and blue are found to be the three mostly usedcolor terms whereas purple is the least favored one.
Table 7:
Fraction of pins having top 10 co-occurring tri-terms,which are obtained from the torso of the distribution of the fash-ion terms in table 2.
Tri-terms LFB HFB ERB leather-pillow-shirt 0.1032 0.1343 0.1763pants-strap-trouser 0.0913 0.1132 0.1298pants-shoe-skirt 0.0613 0.1232 0.1265 jeans-leather-pants 0.1265 0.0942 0.1135jeans-shirt-shorts 0.1175 0.0823 0.0732bag-necklace-skirt 0.0786 0.0974 0.1296dress-shirt-sunglasses 0.0874 0.0925 0.1145 bag-pants-trouser 0.0874 0.1134 0.1341bag-shirt-umbrella 0.0112 0.1324 0.1142dress-pants-shorts 0.0931 0.1121 0.1321
In table 9, we show the occurrence distribution of thetop 10 most co-occurring color bi-terms from the torso ofthe frequency distribution. Once again, we obtain a betterdiscrimination while using bi-terms over single term oc-currences ( over ). Black-yellow and blue-yelloware the most co-occurring color terms for the less followedboards whereas black-pink and pink-red are the most dom-inating color terms occurring together for the highly fol-lowed boards in the torso region. For the expert rankedboards, blue-pink and pink-red are the most used color termswhereas blue-white, black-white, blue-black are the mostdominant color combinations in all the three categories ofboards when we consider the whole distribution. Therefore,we observe that the color composition of images (pins) af-fect the popularity of their boards. Hence, a popularity seek-ing board can host images having a particular color compo-sition from the above analysis.
Table 8:
Fractions of pins having top 10 color terms which areobtained from the torso of the frequency distribution of all colorsterms mentioned in table 3.
Color LFB HFB ERBwhite 0.638 0.664 0.696black 0.555 0.581 0.538blue 0.543 0.566 0.492 brown 0.497 0.379 0.389red 0.346 0.244 0.262wooden 0.338 0.218 0.236 green 0.224 0.221 0.233metal 0.256 0.205 0.198pink 0.122 0.098 0.086purple 0.012 0.008 0.012
Gender infused fashion analysis
In this section, we shallstudy gender based usage of fashion and color terms acrossthe three board categories. In table 10, we show the gen-der based usage of the fashion terms (bi-terms). Each cell ( x, y ) in the table represents the fraction of pins belongingto a particular popularity class ‘y’ having the gender (‘g’)based fashion bi-term ‘a-b’. Here ‘x’ corresponds to ‘g-a-b’. For example, the cell value corresponding to man-bag-jeans and Less Followed Boards means that 0.1375 fraction able 9:
Fractions of pins having top 10 co-occurring bi-terms,which are obtained from the torso of the distribution of all possiblecolor bi-terms from table 3.
Color bi-terms LFB HFB ERBblack-yellow 0.1043 0.0745 0.0943 blue-pink 0.0744 0.0935 0.1064black-pink 0.0824 0.1053 0.1034 metal-red 0.0743 0.0723 0.1053blue-yellow 0.1024 0.0923 0.0814blue-silver 0.0908 0.0956 0.0932 pink-red 0.0824 0.1025 0.1057 metal-silver 0.0823 0.0675 0.0723 red-yellow 0.0923 0.0814 0.0774 grey-white 0.0452 0.0423 0.0424 of pins belonging to less followed boards have male, bag andjeans in them. We observe that of combinations havealmost equal distributions in the highly followed boards andexpert ranked boards, whereas their distributions in the lessfollowed boards are very different from the other two. Suchdifferences in the combination of gender and fashion termsaffect the popularity of boards. It is seen that some combi-nations increase the popularity of boards, whereas rest de-crease it. Another interesting observation we obtain fromthis analysis is that all the 40% combinations which have woman in them correspond to more popular boards.
Table 10:
Fractions of pins having top 10 gender-based fashionbi-terms.
Gender and Fashion Bigrams LFB HFB ERBman-bag-jeans 0.1375 0.1323 0.1250 man-dress-shorts 0.1175 0.1442 0.1525woman-bag-jeans 0.1275 0.1567 0.1489woman-bag-strap 0.1325 0.1578 0.1652 woman-shirt-striped 0.1450 0.1682 0.1575 man-bag-shoe 0.1675 0.1324 0.1434 woman-bag-shoe 0.1424 0.1550 0.1523 woman-shirt-skirt 0.1375 0.1576 0.1503 woman-necklace-pants 0.1324 0.1425 0.1232man-shirts-shorts 0.1453 0.1486 0.1502
Gender infused color analysis
In table 11, we show thedistribution of top five gender-based color bi-terms amongpins across the three board categories. We observe that ∼
60% combinations have equal distributions in highly fol-lowed boards and expert ranked boards, whereas their distri-butions in less followed boards are different from the othertwo. Though black-metal is the dominant color combinationfor female in both less and highly followed boards, white-metal is the prominent color combination in expert rankedboards. In general, metal colors seem to go very well withwomen.
Prediction model
The previous section demonstrates how several factors serveas indicators of popularity of the fashion boards on Pinter-est. In this section, we shall leverage these factors to predictthe future popularity of fashion boards. The popularity ofa board is governed by the number of followers it has. To
Table 11:
Fractions of pins having top five gender-based co-occurring color bi-terms.
Gender-Color Trigrams LFB HFB ERBwoman-metal-white 0.2853 0.2753 0.2895 woman-pink-white 0.2514 0.2657 0.2644 man-pink-white 0.2425 0.2400 0.2350 woman-black-metal 0.2850 0.2675 0.2643woman-brown-green 0.2325 0.2184 0.2135 prevent any from of data leakage we separately re-crawl thenew follower counts of all the fashion boards in our datasetin the month of April, 2017. This follower count statisticstherefore is at a distance of 12 months from the training data. For the prediction task, we shall use the following featureseach of which is motivated by the analysis in the previoussection. • Originality score; • Likes retention coefficient; • Repins retention coefficient; • Likes production coefficient; • Repins production coefficient; • Total number of pins; • Avg. no. of likes on pins; • Avg. no. of repins of pins; • Avg. no. of comments on pins; • Inter-copying time; • Duration of sharing; • Likes speed coefficient; • Repins speed coefficient; • Gender counts (2 bins); Gender bi-term count; • Fashion term count (10 bins); Fashion bi-term count (10bins); Fashion tri-term count (10 bins); • Color term count (10 bins); Color bi-term count (10 bins); • Gender infused fashion bi-term count (10 bins); Genderinfused fashion tri-term count (10 bins); Gender infusedcolor count (5 bins).
Predicting the popularity class of the boards
We have seen that the factors we have discussed earlierhighly discriminate the unpopular class (LFB) from thetwo popular classes (HFB and ERB). The factors, however,can only moderately discriminate one of the popular class(HFB) from the other (ERB). We consider equal number ofdata points for each of the classes and then perform a 10-fold cross validation for generating results. In table 12, wepresent the classification results for i) HFB vs LFB ii) ERBvs LFB and iii) HFB vs ERB. As evident from the table,we can discriminate both popular (HFB or ERB) from theunpopular class (LFB) very well with a very high accuracy(95.96% for HFB vs LFB and 93.95% for ERB vs LFB) andvery high precision, recall and area under ROC curve. Notethat we are only able to obtain a moderate accuracy (65.1%)in classifying the two popular classes since the boards be- Note that we do not use temporal statistics between March2016 and April 2017 for enhanced robustness of the model; theidea is to make efficient predictions using minimal set of featuresthat can be easily obtainable at any static time point. onging to these two classes have very similar characteris-tics.For the classification task, we have used Support Vec-tor Machines (SVM), Logistic Regression (LR) and Ran-dom Forest (RF) classifiers implemented in the WekaToolkit (Hall et al. 2009). We choose the three classifiersfor their diversity since they are known to be able to solve avast range of different types of classification problems. Eachof these classifiers represent different schools of thoughtsand have their own set of strengths and advantages . Allthe classifiers yield similar performance results with Ran-dom Forest classifier performing the best. Table 12:
Performance of various classifiers for classification of i)HFB vs LFB ii) ERB vs LFB ii) ERB vs HFB.Categor-ies Classif-iers Accu-racy Preci-sion Recall F-Score ROCAreaHFB vsLFB SVM 92.51% 0.935 0.925 0.925 0.928LR 94.91% 0.95 0.949 0.949 0.981
RF 95.96% 0.96 0.96 0.96 0.995
ERB vsLFB SVM 92.06% 0.921 0.921 0.921 0.92LR 93.27% 0.934 0.933 0.933 0.977
RF 93.95% 0.94 0.94 0.94 0.99
ERB vsHFB SVM 61.08% 0.62 0.611 0.603 0.611
LR 65.1%
RF 64.81%
Predicting the followership counts of the boards
To study the robustness of our prediction model, we furthertry to predict the actual popularity, i.e., the logarithmic val-ues of the followership counts of the boards. Toward thisobjective, we use Support Vector Regression (SVR) due tonon-linearity of the problem. We use sequential minimal op-timization (SMO) algorithm for training the SVR. We per-form both separate training and testing as well as 10-foldcross validation method. We consider Pearson VII function-based universal kernel (PUK) due to its flexibility and adapt-ability through adjusting kernel parameter. We set the costparameter (C) as 1. For evaluating how good the predictionis, we use Pearson correlation coefficient ( ρ ), normalizedroot mean square error (RMSE). We achieve high correlationcoefficient ( . ) and low normalized root mean squareerror ( . ) which establishes the fact that the featuresobtained are robust and discriminating in nature (see ta-ble 13 ). Both cross-validation and separate training/testingproduces very similar results. Table 13:
Regression results.
Method ρ Normalized RMSE10-fold cross-validation 0.8659 0.146Separate training/testing(4:1 ratio)
Discriminative features:
In order to determine the discrim-inative power of each feature, we use the
RELIEF F fea- https://bit.ly/2LkuSf0 We have also tried linear regression model which gives corre-lation coefficient of 0.7363 and 0.2 as normalized the RMSE valuefor 10-fold cross validation setting.
Table 14:
Top predictive features and their ranks.Rank Features1 LRC2 RRC3 bag-striped (fashion)4 white (color)5 black (color)6 blue (color)7 female8 male-female9 brown (color)10 blue-pink (color)11 pink (color)12 black-pink (color)13 woman-pink-white (gender-color)14 red (color)15 man-shirts-shorts (gender-fashion) ture selection algorithm (Kononenko, Simec, and Robnik-Sikonja 1997) available in the Weka Toolkit. Table 14 showsthe rank of the features in terms of their discriminatingpower for prediction. The rank order clearly indicates thatfor popularity prediction the sharing/borrowing features, thecolor terms and some of the fashion terms are important.Likes retention coefficient, repins retention coefficient arethe top discriminative features followed by various colorterm based features. Therefore, color (sometimes in con-junction with fashion and gender term) seems to be one ofthe most important discriminator for popularity of fashionboards.
Discussions and conclusions
In this section we outline various insights and implicationsof the current work. We also discuss the generalizability ofthe current work and finally draw the conclusions.
Insights and implications
Insights : The current study puts forward a lot of insightsespecially for new and upcoming fashion brands. • Certain social sharing behavior of users can make boardspopular. The most crucial among these are the retentioncoefficients. Popular boards are able to retain their pro-portion of ‘likes’ and ‘repins’ despite a lot of sharing andre-sharing of pins. Specially, engineered campaigns by thefashion houses can be made to ensure/promote such reten-tions. • More female faces or both male and female faces togethermay be promoted by the fashion houses since that, as wehave seen, could lead to enhanced popularity. • Certain choices of colors (e.g., white, black, blue, pinketc.) and color combinations (e.g., blue-pink, black-pinketc.) may be more advertised to enhance the chances ofbeing more popular. • Certain fashion items like ‘striped bags’ seem to be verycommon in popular brands and could be more promotedby the newbies. • For male fashion, ‘shirts’ and ‘shorts’ are the items thatseem to propel popularity and can therefore be more vig-orously advertised by the new outlets. Many articles , in act, have noted that shorts like boxers and bathing suitsthat end above the knee enhance the sex appeal of men. • For female fashion, colors like pink and white seem tobe good indicators of popularity. In fact, pink has beenthe most favorite color of garments for women for a verylong time . These therefore can be items of more focusedpublicity by the upcoming fashion agencies. Implications : Our findings make several contributions to ex-isting research. We believe, this research opens new pathwayto understand new factors like colors, faces, fashion termswhich are influential for understanding popularity. Our workalso echoes some of the previous findings on impact of coloron diffusion. We also suggest color combinations that makesa board popular. For newbie fashion houses and fashiontrend-setters, our findings shed light on how images can beconstructed so that they become popular. Pins of certain col-ors, more female faces or male-female joint faces could besome of the prime suggestions. One could also launch cam-paigns to promote their boards in such a way that the orig-inating boards are able to retain the ‘likes’ and ‘repins’ oftheir pins in the face of constant sharing of these pins. In fact,Pinterest can make such ‘tips-n-tricks’ application availablein exchange of a small amount of subscription from everynewbie. This could potentially be a premium/paid supportand could be a business model for the company for possibil-ity of enhanced revenues.There are several mobile apps which provide users withphoto-editing tools. One of the widely used techniques inphoto-editing is applying filters to them. These filters canchange saturation, brightness, and color distribution of theimage. Our findings can be used to design new filters forphoto editing. Filters that increase saturation or enhance thewarmness of the image will likely increase engagement withthe photo.
Generalizability
Though the entire study has been performed on Pinterest, thefindings can be generalized in other similar websites focusedon images, for example, professional photography site likeFlickr, or people-focused website like Instagram. Instagramis also a quite popular website for fashion trend. We believethese findings in the form of importance of color combina-tions and fashion terms influencing popularity can be gener-alized to Instagram, Flickr and Tumblr as well, though thepopularity figures might vary which is mostly dependent onthe website’s underlying usage among communities, rankingalgorithms etc.
Conclusions
In summary, we study various aspects of fashion boards onPinterest. Our proposed measures – retention coefficients,production coefficients, inter-copying time and duration ofsharing portray the sharing dynamics evident in less fol-lowed, highly followed and expert ranked fashion boards.We observe that generally highly followed and expertranked fashion boards are able to make an existing non- guys-wear/ popular pin popular, whereas less popular fashion boardsdo not exhibit this characteristic. Further, if a pin has origi-nated from highly followed or expert ranked fashion boards,it would achieve high popularity on this board than the sub-sequent boards on which it would be shared in future. Wealso observe that the pins on the highly followed and expertranked fashion boards keep getting shared for a long time,whereas this happens for a short time for the pins on lessfollowed fashion boards.Gender, fashion and color terms embedded in images alsoyield interesting and conclusive results. We observe thatboth highly followed and expert ranked fashion boards ex-hibit similar trend in the usage of fashion bi- and tri-terms.We also observe that a large number of pins having femalefaces are present in highly followed and expert ranked fash-ion boards; the number of female faces is 20% lower forthe less followed boards. Similar trend is observed for pinshaving both male and female faces. We also study occur-rences of gender-based fashion and color terms. We identifycombinations which give good discriminatory results acrossthe three board categories. We try to leverage various shar-ing/borrowing characteristics, image-based content charac-teristics of fashion boards to predict their future popularity(logarithm of follower count). We achieve a high correlationcoefficient of . and low RMSE. Limitations : We acknowledge that there is some limitationof the current study. We specifically note the fact that someof the features and outcomes might be influenced by the par-ticulars of the Pinterest ranking algorithms (e.g., what getsfeatured on the homepage, how personalization affects theprobability a pin will be surfaced, etc.). There is no waywe can control the internal algorithm promoting pins andboards. However, we believe, the factors we come up withare indeed influential as they strongly correlate with popu-larity studied on a large-scale data.
References [Bakhshi and Gilbert 2015] Bakhshi, S., and Gilbert, E.2015. Red, purple and pink: The colors of diffusion on pin-terest.
PloS one
CHI ’14 , 965–974. ACM.[Bellizzi, Crowley, and Hasty 1983] Bellizzi, J. A.; Crowley,A. E.; and Hasty, R. W. 1983. The effects of color in storedesign.
Journal of retailing .[Berger and Milkman 2012] Berger, J., and Milkman, K. L.2012. What makes online content viral?
Journal of market-ing research
CHI ’09 , 945–954.[Chang et al. 2014] Chang, S.; Kumar, V.; Gilbert, E.; andTerveen, L. G. 2014. Specialization, homophily, and gen-der in a social curation site: Findings from pinterest. CSCW’14.Crowley 1993] Crowley, A. E. 1993. The two-dimensionalimpact of color on shopping.
Marketing letters
ICWSM , 42–51.[Dubois and Duquesne 1993] Dubois, B., and Duquesne, P.1993. The market for luxury goods: Income versus culture.
European Journal of Marketing
PloS one
SIGKDD Explor.Newsl.
ICWSM , 82–91.[Hessel, Lee, and Mimno 2017] Hessel, J.; Lee, L.; andMimno, D. 2017. Cats and captions vs. creators and theclock: Comparing multimodal content to context in predict-ing relative popularity. In
WWW ’17 , 927–936.[Hochman and Schwartz 2012] Hochman, N., and Schwartz,R. 2012. Visualizing instagram: Tracing cultural visualrhythms. In
SocMedVis in conjunction with ICWSM ’12 ,6–9.[Jamali and Rangwala 2009] Jamali, S., and Rangwala, H.2009. Digging digg: Comment mining, popularity predic-tion, and social network analysis. In
WISM ’09 , 32–38.IEEE.[Johnson, Karpathy, and Fei-Fei 2016] Johnson, J.; Karpa-thy, A.; and Fei-Fei, L. 2016. Densecap: Fully convolutionallocalization networks for dense captioning. CVPR ’16.[Kim and Ko 2010] Kim, A. J., and Ko, E. 2010. Impacts ofluxury fashion brands social media marketing on customerrelationship and purchase intention.
Journal of Global Fash-ion Marketing
Journal of BusinessResearch
Applied Intelligence
Journal ofAdvertising Research . [Lee and Ma 2012] Lee, C. S., and Ma, L. 2012. News shar-ing in social media: The effect of gratifications and prior ex-perience.
Computers in Human Behavior arXivpreprint arXiv:1512.01174 .[Meyers-Levy and Peracchio 1995] Meyers-Levy, J., andPeracchio, L. A. 1995. How the use of color in advertisingaffects attitudes: The influence of processing motivationand cognitive demands.
Journal of Consumer Research
ACR North American Advances .[Miller, Chang, and Terveen 2015] Miller, H. J.; Chang, S.;and Terveen, L. G. 2015. “i love this site!” vs. “it’s a littlegirly”: Perceptions of and initial user experience with pin-terest. CSCW ’15, 1728–1740.[Ottoni et al. 2013] Ottoni, R.; Pesce, J. P.; Las Casas, D. B.;Franciscani Jr, G.; Meira Jr, W.; Kumaraguru, P.; andAlmeida, V. 2013. Ladies first: Analyzing gender roles andbehaviors in pinterest. ICWSM ’13.[Sanchis-Ojeda, Sibley, and Massimi 2016] Sanchis-Ojeda,R.; Sibley, D.; and Massimi, P. 2016. Detection of fashiontrends and seasonal cycles through the analysis of implicitand explicit client feedback. In
KDD Fashion Workshop .[Schindler 1986] Schindler, P. S. 1986. Color and contrast inmagazine advertising.
Psychology & Marketing
Journal of Advertising
HBS Market-ing Research Paper .[Wu et al. 2017] Wu, B.; Cheng, W.-H.; Zhang, Y.; Huang,Q.; Li, J.; and Mei, T. 2017. Sequential prediction of socialmedia popularity with deep temporal context networks. In
IJCAI ’17 .[Yamaguchi, Berg, and Ortiz 2014] Yamaguchi, K.; Berg,T. L.; and Ortiz, L. E. 2014. Chic or social: Visual popularityanalysis in online fashion networks. MM ’14, 773–776.Yang, Li, and Luo 2015] Yang, X.; Li, Y.; and Luo, J. 2015.Pinterest board recommendation for twitter users. MM ’15,963–966.[You et al. 2017] You, Q.; Garca-Garca, D.; Paluri, M.; Luo,J.; and Joo, J. 2017. Cultural diffusion and trends in face-book photographs. ICWSM ’17.[Zhong et al. 2014] Zhong, C.; Salehi, M.; Shah, S.;Cobzarenco, M.; Sastry, N.; and Cha, M. 2014. Socialbootstrapping: How pinterest and last.fm social communi-ties benefit by borrowing links from facebook. WWW ’14,305–314.[Zhong, Kourtellis, and Sastry 2016] Zhong, C.; Kourtellis,N.; and Sastry, N. R. 2016. Pinning alone? a study of therole of social ties on pinterest. In