Abstract

A method to produce personalized classification models to automaticallyreview online dating profiles on Tinder is proposed, based on the user'shistorical preference. The method takes advantage of a FaceNet facialclassification model to extract features which may be related to facialattractiveness. The embeddings from a FaceNet model were used as the featuresto describe an individual's face. A user reviewed 8,545 online dating profiles.For each reviewed online dating profile, a feature set was constructed from theprofile images which contained just one face. Two approaches are presented togo from the set of features for each face, to a set of profile features. Asimple logistic regression trained on the embeddings from just 20 profilescould obtain a 65% validation accuracy. A point of diminishing marginal returnswas identified to occur around 80 profiles, at which the model accuracy of 73%would only improve marginally after reviewing a significant number ofadditional profiles.

 

Introduction (beta)

None

 

Conclusion (beta)

None