【应用】预测新产品上市后的基于时间的销售数据
【领域】Neural networks; RNNs; Encoder-Decoder;
【文章要点】
1. 使用历史数据进行训练,预测一个新产品上市后的销售情况
2. 数据:产品图像数据I+产品属性数据x(如 design attributes such as color, pattern, sleeve style etc. or merchandising attributes such as list price, promotion etc.)
3. 问题定义:销售数据:![](https://img2020.cnblogs.com/blog/985935/202009/985935-20200903105701805-1590595806.png)
![](https://img2020.cnblogs.com/blog/985935/202009/985935-20200903105701805-1590595806.png)
输入数据:![](https://img2020.cnblogs.com/blog/985935/202009/985935-20200903105703347-1813156022.png)
![](https://img2020.cnblogs.com/blog/985935/202009/985935-20200903105703347-1813156022.png)
需要求取的是:![](https://img2020.cnblogs.com/blog/985935/202009/985935-20200903105704711-541319099.png)
![](https://img2020.cnblogs.com/blog/985935/202009/985935-20200903105704711-541319099.png)
同时,还需要考虑一些外部因素,如周末,节日,重大节日,促销等, 外部的因素表示为![](https://img2020.cnblogs.com/blog/985935/202009/985935-20200903105706612-1572712124.png)
![](https://img2020.cnblogs.com/blog/985935/202009/985935-20200903105706612-1572712124.png)
综上,求取的是![](https://img2020.cnblogs.com/blog/985935/202009/985935-20200903105708157-255586468.png)
![](https://img2020.cnblogs.com/blog/985935/202009/985935-20200903105708157-255586468.png)
4. 系统整体示意:
![](https://img2020.cnblogs.com/blog/985935/202009/985935-20200903105709867-209638806.png)
5. 传统方法:使用KNN。即将新产品与历史产品做比对,得到最为相似的K个旧产品,使用旧产品的历史数据集成并做响应的预测
1) Attribute KNN. 使用商品属性的距离,求取紧邻,使用距离作为权重的参考,将k个商品的历史销售数据集成。θ为距离相关的数据
![](https://img2020.cnblogs.com/blog/985935/202009/985935-20200903105711300-2029856273.png)
2) Embedding KNN. 等不能直接量化的特征embedding,embedding后再求K个紧邻,随后同上。Φ为求embedding
![](https://img2020.cnblogs.com/blog/985935/202009/985935-20200903105712425-134236782.png)
6. 基于Encoder-Decoder的时间序列模型。探索了多个模型
1)Sequence learning with encoded image input (Image RNN)
Encoder模块为给定的输入图像计算一个紧凑的嵌入,并将其与时间特征合并,然后再把encoder生成的数据输入到RNN decoder
2) Sequence learning with encoded multi-modal inputs.(Multi-modal RNN)
与1)不同的是,将属性标签做了embedding,一起加进去网络中
![](https://img2020.cnblogs.com/blog/985935/202009/985935-20200903105720254-1832910013.png)
3) Explainable sequence learning with attended multi-modal inputs.(Cross-Attention RNN)
加入可解释性的部分?使用Cross-Attention
![](https://img2020.cnblogs.com/blog/985935/202009/985935-20200903105721914-52234625.png)
![](https://img2020.cnblogs.com/blog/985935/202009/985935-20200903105723151-305955975.png)
6. 实验结果
![](https://img2020.cnblogs.com/blog/985935/202009/985935-20200903105729379-79501169.png)
7. 可解释性的结果
![](https://img2020.cnblogs.com/blog/985935/202009/985935-20200903105734929-186197103.png)