面向电商营销决策的多模态服装销量预测方法研究

Research on Multimodal Clothing Sales Forecasting Methods for E-Commerce Marketing Decision-Making

摘要: 服装行业商品上新频繁、需求波动大，传统仅依赖历史销量序列的预测方法难以准确刻画多因素共同作用下的销量变化规律。为提升预测准确性，本文提出一种基于多模态信息融合的服装销量预测方法，将商品图像、文本描述、用户评价与历史销量等多源数据统一建模，以全面刻画影响销量的关键因素。模型由多模态编码、门控融合、跨模态对齐和GRU解码器四个模块组成。各模态特征分别编码并标准化至统一表示空间；通过门控机制与跨模态注意力进行动态融合；对比学习增强图文时序特征的一致性；最终通过GRU完成未来多步销量预测。实验表明，该方法在MAE、RMSE、MAPE与R²等指标上均显著优于基线模型；多步预测结果显示，模型在短期补货及中期库存规划任务中具备更高稳定性。消融实验进一步验证了门控融合与对比学习模块对提升性能的重要性。研究结果得出，多模态融合能够有效弥补结构化数据表达不足，为服装行业的补货策略与库存管理提供更可靠的预测支持。

Abstract: The clothing industry is characterized by frequent product updates and significant demand fluctuations, and traditional forecasting methods relying solely on historical sales sequences struggle to accurately capture sales variation patterns under the combined influence of multiple factors. To improve prediction accuracy, this paper proposes a multimodal clothing sales forecasting method based on multimodal information fusion, which jointly models multiple data sources, including product images, textual descriptions, user reviews, and historical sales data, to comprehensively characterize the key factors affecting sales. The proposed model consists of four modules: multimodal encoding, gated fusion, cross-modal alignment, and a GRU decoder. Features from different modalities are separately encoded and normalized into a unified representation space; dynamic fusion is achieved through a gating mechanism and cross-modal attention; contrastive learning is employed to enhance the consistency of image, text, and temporal features; and finally, a GRU decoder is utilized to perform future multi-step sales forecasting. Experimental results show that the proposed method significantly outperforms baseline models in terms of MAE, RMSE, MAPE, and R². Multi-step forecasting results further demonstrate that the model exhibits higher stability in short-term replenishment and medium-term inventory planning tasks. Ablation studies further verify the importance of the gated fusion and contrastive learning modules in improving predictive performance. The findings indicate that multimodal fusion can effectively compensate for the limitations of structured data representation and provide more reliable forecasting support for replenishment strategies and inventory management in the clothing industry.