在新服装产品销售预测任务中，由于缺乏历史销售数据，通常需要充分利用其他模态的数据作为补充。然而，多模态服装数据通常具有冗余性和异构性。为解决这些问题，提出一种包括三个主要元素的层次化多模态注意力循环神经网络(hierarchical multi-modal attention based recurrent neural network，HMA-RNN)。层次化结构将高层语义信息与低层语义信息分离，以避免信息冗余。在模态融合阶段引入多模态注意力机制(multi-modal attention，MMA)以减轻固有的数据不对齐问题。采用共享注意力机制构建跨多模态数据的依赖关系。在 Visuelle 2。0 数据集上的试验结果表明，所提出的方法加权平均百分比误差(weighted average percentage error，WAPE)为 72。07，平均绝对误差(mean absolute error，MAE)为0。80，明显优于现有的方法，表明了该研究所提出的方法的有效性。
Sales Forecasting of New Clothing Products Based on Hierarchical Multi-Modal Attention Recurrent Neural Network
In the task of sales forecasting of new clothing products,the lack of historical sales data often necessitates the full utilization of data from other modalities as a supplement.However,multi-modal clothing data are usually redundant and heterogeneous.To solve the problems,a hierarchical multi-modal attention based recurrent neural network(HMA-RNN)including three main elements is proposed.The hierarchical structure separates high-level semantic information from low-level semantic information to avoid information redundancy.The multi-modal attention(MMA)is introduced in the fusion stage to mitigate inherent data non-alignment.The shared attention mechanism is utilized to build the dependencies across the multi-modal data.Experimental results on the Visuelle 2.0 dataset show that the proposed approach achieves promising results with 72.07 on the weighted average percentage error(WAPE)and 0.80 on the mean absolute error(MAE),outperforming existing works significantly,which indicates the effectiveness of the proposed approach.
国家科技期刊平台
NSTL
万方数据
