Vision-Conditioned Gating for Product-Store-Level Demand Forecasting of New Fast-Fashion Products

초록

Fast fashion demand forecasting at the product-store level is a critical task for improving market responsiveness, profitability, and operational stability. Nevertheless, product-store level data are often sparse and highly volatile, and forecasting becomes particularly challenging in cold-start settings where new products lack historical sales records. In response, this study proposes Vision-Conditioned Gated RNN (VIGRNN), based on the observation that the predictive usefulness of auxiliary information varies across product-store instances. VIGRNN is an adaptive forecasting model that dynamically controls attribute, release-date, and trend information on a per-sample basis, conditioned on visual information. Experimental results show that, under cold-start settings, the proposed model outperforms an existing multimodal RNN-based baseline by reducing WAPE by 3.6% and MAE by 2.47%. In addition, it reduces GFLOPs by 5.33% and the number of parameters by approximately 32%. These findings suggest that the proposed model effectively balances forecasting accuracy and computational efficiency, highlighting its potential for practical deployment in real-world industrial environments.

키워드