IFF: An Intelligent Fashion Forecasting System

Abstract

Fashion evolves continuously, and there is a real need for technological solutions to aid the fashion industry in predicting its dynamic nature to reduce waste, enhance quality output, and augment sales. Exploration of this task so far does not account for fine-grained elements and spans shorter time periods, which is not sufficient to analyze a trend’s temporal behavior. In this work, we introduce a framework that can provide a probabilistic popularity score for the apparel in an image. It consists of a neural network that is trained to identify and extract all the fashion attributes from an image and a forecasting model trained over a dataset of attributes and their trendiness over the last century. When provided with an image, the framework extracts a list of all the fashion attributes. The forecasting model then predicts the future trend for each of the attributes in the list using the dataset. Finally, an average score for the dresses’ trendiness is generated. A number of models were tested for the forecasting task, and the best model was selected based on Mean Absolute Error. Models tried ranged from statistical time series forecasting models such as ARIMA and SARIMAX to neural network-based models such LSTMs and Seq2Seq models. The best model was found to be the Seq2Seq model, with an F1 score of 0.813, and hence, it was chosen for building the framework. The framework can be applied to analyze how different attributes of a dress affect the popularity scores of a dress, given a picture.