Multimodal Quasi-AutoRegression: forecasting the visual popularity of new fashion products

• DOI: 10.1007/s13735-022-00262-5
• Publication Date: 2022-10-08T14:03:29Z
• Authors: Stefanos-Iordanis...; Christos Koutlis; Symeon Papadopoulos; Ioannis Kompatsiaris
• Keywords: Popularity; Closed captioning

Abstract Estimating the preferences of consumers is utmost importance for fashion industry as appropriately leveraging this information can be beneficial in terms profit. Trend detection a challenging task due to fast pace change industry. Moreover, forecasting visual popularity new garment designs even more demanding lack historical data. To end, we propose MuQAR, Multimodal Quasi-AutoRegressive deep learning architecture that combines two modules: (1) multimodal multilayer perceptron processing categorical, and textual features product (2) neural network modelling “target” time series product’s attributes along with “exogenous” all other attributes. We utilize computer vision, image classification captioning, automatically extracting descriptions from images products. Product design initially expressed visually these represent products’ unique characteristics without interfering creative process its designers by requiring additional inputs (e.g. manually written texts). employ target proxy temporal patterns, mitigating data, while exogenous help capture trends among interrelated perform an extensive ablation analysis on large-scale datasets, Mallzee-P SHIFT15m assess adequacy MuQAR also use Amazon Reviews: Home Kitchen dataset generalization domains. A comparative study VISUELLE shows capable competing surpassing domain’s current state art 4.65% 4.8% WAPE MAE, respectively.