基于多模态信息融合的卷烟销量预测方法

The Multimodal-Based Method for Tobacco Sales Forecast

融合多模态信息的数据科学对智能营销至关重要.该文提出了一种融合了视觉、自然语言和结构化数据的基于多模态信息的烟草销售预测方法.首先,引入扩散模型来生成高质量的香烟图像样本以进行数据扩增,在柜台香烟识别阶段,采用深度耦合网络和多元组排序损失来提高香烟识别率;其次,在销售预测中,使用多模态信息作为输入,包括柜台位置、采用双向文本编码的品牌表示以及相应价格;最后,通过预测算法得到香烟的预测销量.通过全面综合的分析为营销提供了有价值的建议,促进了多模态信息在卷烟营销科学上的应用.

Data science with multimodal information is of great importance for intelligent marketing.In the paper,a tobacco sales prediction method based on multimodal information is proposed,including visual,natural language,and structured data.Firstly,a diffusion model is introduced to generate high-quality cigarette image samples for cigarette recognition.In the cigarette recognition stage,a deep coupled network and multiple sets of ranking losses are employed to improve cigarette recognition at the counter.Secondly,in sales prediction,cigarettes are represented with multimodal information,including counter location,brand representation encoded with Bidirectional Encoders Representations from Transformers(BERT),and corresponding prices.Finally,the monthly sales of cigarettes are provided.Through comprehensive analysis,valuable strategic recommendations are provided for marketing,promoting the multimodal-based application in tobacco science marketing.