Predicting consumers' garment fit satisfactions by using machine learning

Abstract

The objectives of this study were to apply alternative machine learning (ML) algorithms to predict consumers' garment fit satisfactions (real fit satisfaction [RFS]) and compare the efficiencies of these algorithms to predict RFS. Skirts made from different fabrics were used as test garments. Mechanical properties of the skirts' fabrics were assigned as predictor variables to estimate RFS. Study participants' virtual body models were created by using 3D body scanner and used for virtual fitting. Each participant physically tried on the skirts and evaluated the fit. Participants also viewed the skirt simulations on their avatars and evaluated the virtual fit, which represented participants' virtual fit satisfactions (VFS). Random Forest (RF), support vector machine (SVM), and conditional tree (CT) algorithms were used to learn from the data to predict participants' RFSs. The mean correlations between the predicted and observed RFS values in the validation sets were 0.74 (RF), 0.70 (SVM-linear kernel), 0.72 (SVM-radial kernel), and 0.55 (CT). According to the variable importance analysis, VFS had the highest importance among 35 predictor variables. ML is used mostly for sales forecasting and manufacturing purposes in the fashion industry. However, garment fit, which restrains consumers from shopping online, did not get enough attention in ML studies. Besides, the ML algorithms used in fashion and apparel studies are often genetic algorithms and neural networks; therefore, there is a need to test other algorithm types. In this study, we offered alternative ML algorithms (i.e., RF, SVM, and CT) to predict consumers' garment fit satisfactions.