《计算机应用》唯一官方网站 ›› 2023, Vol. 43 ›› Issue (S1): 107-111.DOI: 10.11772/j.issn.1001-9081.2022121877

• 数据科学与技术 • 上一篇    

基于时空信息转换方程的药品销量预测模型

靳东辉1,2,3,4,5, 杨小博1,4,5(), 郭炳晖1,2,3,4,5,6,7   

  1. 1.北京航空航天大学 数学科学学院, 北京 100191
    2.北京航空航天大学 人工智能研究院, 北京 100191
    3.中关村实验室, 北京 100094
    4.软件开发环境国家重点实验室(北京航空航天大学), 北京 100191
    5.数学信息与行为教育部重点实验室(北京航空航天大学), 北京 100191
    6.未来区块链与隐私计算高精尖中心(北京航空航天大学), 北京 100191
    7.鹏城实验室, 广东 深圳 518055
  • 收稿日期:2022-12-21 修回日期:2023-03-16 接受日期:2023-03-17 发布日期:2023-07-04 出版日期:2023-06-30
  • 通讯作者: 杨小博
  • 作者简介:靳东辉(2000—),男,河南洛阳人,硕士研究生,CCF会员,主要研究方向:数据科学、机器学习与时序预测、复杂网络
    杨小博(1994—),男,山东临沂人,博士研究生,CCF会员,主要研究方向:数据科学、复杂网络.yangxb@buaa.edu.cn
    郭炳晖(1982—),男,河北张家口人,副教授,博士,CCF会员,主要研究方向:动力系统、计算复杂性、人工智能基础、复杂系统及网络的信息传播行为、在线社会网络的舆情动力学。
  • 基金资助:
    科技创新2030—“脑科学与类脑研究”重大项目(2021ZD0201302);广东省重点领域研发计划项目(2021B0101420003)

Pharmaceutical sales prediction model based on spatiotemporal information transformation equations

Donghui JIN1,2,3,4,5, Xiaobo YANG1,4,5(), Binghui GUO1,2,3,4,5,6,7   

  1. 1.School of Mathematical Sciences,Beihang University,Beijing 100191,China
    2.Institute of Artificial Intelligence,Beihang University,Beijing 100191,China
    3.Zhongguancun Laboratory,Beijing 100094,China
    4.State Key Laboratory of Software Development Environment(Beihang University),Beijing 100191,China
    5.Key Laboratory of Mathematics,Informatics and Behavioral Semantics(Beihang University),Beijing 100191,China
    6.Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing(Beihang University),Beijing 100191,China
    7.Peng Cheng Laboratory,Shenzhen Guangdong 518055,China
  • Received:2022-12-21 Revised:2023-03-16 Accepted:2023-03-17 Online:2023-07-04 Published:2023-06-30
  • Contact: Xiaobo YANG

摘要:

针对药品销售中高维短时间序列预测问题,利用时空信息转换方程及储备池计算方法构建了一种基于时空信息(STI)转换方程的药品销售量预测模型。首先针对药品销售时间序列数据样本量较小的特点,引入储备池计算方法拓展数据样本信息维度,将多个不同药品销售量时序数据中的动力学信息引入储备池。使用时空信息转换方程对时间信息与空间信息进行转化,最后在储备池运算的基础上对时空信息转化方程求解,对目标药品的销售量进行有效的时间序列预测。通过将提出的基于时空信息转换方程的时序预测模型与神经网络预测模型在特定药品销售数据集进行时序预测验证并进行横向对比,相较于GRU(Gated Recurrent Unit),所提模型在测试时间节点上的均方根误差(MSE)及运算时间分别减小了13.27%和95.60%、皮尔逊相关系数提高了34个百分点;相较于长短期记忆模型(LSTM),所提模型在测试时间节点上的均方根误差及运算时间分别减小了69.85%和98.00%,而皮尔逊相关系数提高了44个百分点;相较于卷积神经网络模型(CNN),在测试节点的均方根误差及运算时间分别减少了48.96%和88.53%,皮尔逊相关系数提高了33个百分点。证明了基于时空信息转换方程的药品销售预测模型在测试集时间节点上的预测效果要优于GRU、LSTM、CNN时序预测模型,同时也说明模型具有更高的运算效率。

关键词: 医药销售, 小样本学习, 时序预测, 时空信息转换方程, 储备池计算, 相关性分析。

Abstract:

In order to address the problem of high-dimensional short time series forecasting in pharmaceutical sales, a pharmaceutical sales prediction model based on Spatiotemporal Information(STI) transformation equations and reservoir computing method was proposed. To address the small sample size of pharmaceutical sales time series data, the reservoir computing method was introduced to expand the data sample dimension by incorporating the dynamics information from multiple different pharmaceutical sales time series data into the reservoir pool. Furthermore, the spatial information was transformed into temporal information using STI transformation equations. And the prediction of the target pharmaceutical sales was achieved by solving the equations based on the reservoir computing method. To validate the proposed STI-transformation-based time series prediction model, it was compared with neural network-based prediction models on specific pharmaceutical sales dataset. Compared with the Gated Recurrent Unit (GRU), the proposed model reduces the Mean Squared Error (MSE) and computation time by 13.27% and 95.60%, respectively, and increases the Pearson correlation coefficient by 34 percentage points at the test time node. Compared with the Long Short Term Memory (LSTM) model, the proposed model reduces the MSE and computation time by 69.85% and 98.00%, respectively, and increases the Pearson correlation coefficient by 44 percentage points. Compared with the Convolutional Neural Network (CNN) model, the proposed model reduces the MSE and computation time by 48.96% and 88.53%, respectively, and increases the Pearson correlation coefficient by 33 percentage points. These results demonstrate that the spatiotemporal information transform based pharmaceutical sales prediction model outperforms GRU, LSTM, and CNN-based prediction models in terms of prediction accuracy and computational efficiency on the test dataset.

Key words: pharmaceutical sales, small sample learning, time series prediction, spatiotemporal information transformation equation, reservoir computing, correlation analysis

中图分类号: