Computational models for fashion models

Who’s going to be the next top model? Researchers from the Indiana University used Instagram as a dataset for their algorithm that can predict the popularity of new faces with an accuracy greater than 80%. They have used 400 models from the Fashion Model Directory, analyzed their Instagram accounts and their relevant statistics, for example the number of posts, likes, comments and whether these comments were positive or negative. A subset of this data was chosen and marked as the new faces, and of these eight were expected to achieve the greatest popularity, while the real result was that six of these were accurately predicted. The same goes for the most unpopular ones – of the seven predicted to be non that popular, six of seven were accurately identified.



The parameter of popularity was established as the number of catwalks in which a new model participated during the fall and winter 2015 season (to account for a homogeneous set of runway experiences). The six most popular models of this season were Sofia Tesmenitskaya, Arina Levchenko, Renata Scheffer, Sasha Antonowskaia, Melanie Culley and Phillipa Hemphrey, which correlated greatly with the expectations. The non-model part, of course, consisted of Indiana University researchers Giovanni Luca Ciampaglia, Emilio Ferrara and Jaehzuk Park. The researchers have explained that they chose the fashion industry for their research because it represented a strong mentality of the winner takes it all, where a certain kind of social darvinism comes into picture most prominently.

The computational method was based on the fact that a high number of likes, comments and frequent posts were correlated with the runway successes. A huge amount of comments did not affect popularity, but a higher than average number of posts yielded a higher chance of walking a runway. The surprising fact was that more likes than usual could lower these changes! Perhaps as an overkill? All of these factors show that models’ online activity does play a great and an important role in their popularity, and consequently, success. Of course, that’s not the only factor, physical ones are important too, as well as adequate modeling agency representation. The winning combination seems to be having physical prerequisites and adequate online and offline representation.

The Fashion Model Directory data consisted of the following attributes: name, hair color, eye color, height, hip size, dress size, waist size, shoe size, list of agencies, nationality, details about runways, and these researchers have discarded the data about hair and eye color, since color coding was not reliable enough for meaningful characterizations, along with the unnecessary nationality information. They annotated each agency to reflect its reputation in the fashion industry, and used a simple binary classification system for the classification. Social media activity was analyzed with sentiment analysis on English comments by a simple Naive Bayes classifier (using probabilistic methods and assuming independence between features). For the prediction, they used three widely-used machine learning algorithms, based on ensemble methods and boosting: decision trees (where leaves are class labels and branches are conjunctions of features), Random forest (multitudes of decision trees), and AdaBoost (reducing bias and variance, and the weak learners are tweaked in favor of those instances misclassified by previous classifiers).

These methods could potentially be applied outside of the modelling world as well, for example to predict the popularity of actors and actresses, or musicians and other entertainment industry people, which could be a potential asset in further cultural and sociological research and studies. The study will be presented at the 2016 conference in San Francisco called Computer-Supported Cooperative Work and Social Computing.

_______________

References: