基于时序矩阵分解的缺失销售数据估计

企业的历史销售记录是供应链优化研究的基础数据来源,然而,在日常的研究中,几乎所有可以通过公开途径获得的销售记录都是高度不完整的,这为研究者开展工作带来了极大的不便。为解决此问题,本文提出,以销售数据集中已有的数据为基础,使用面向时序数据的矩阵分解模型MAFTIS对其缺失的部分进行估算,从而把残缺的数据集补全完整。进一步地,为提高MAFTIS的计算效率,本文还为该模型设计了一种基于交替最小二乘法的求解策略MAFTIS^ALS。在评估实验中,MAFTIS^ALS被用于三个真实销售数据集的缺失记录估计,结果显示,与其它估计模型相比,MAFTIS^ALS能获得更准确的估计结果,并且具有更高的收敛速度。

Estimating the Missing Sales Data with the Time Series Matrix Factorization Model

The historical sales data are one of the most important data sources for supply chain optimization (SCO) research. However, for most SCO researchers, the publicly accessible sales data is usually highly incomplete, which makes their work difficult. In order to solve this problem, we propose to use MAFTIS, atime series matrix factorization model, to estimate the missing sales data with respective to the collected ones.In addition, to speed up the computation procedure of MAFTIS, we also present an alternating least square (ALS) based solution algorithm for the model.For evaluations, we apply the proposed method to estimate the missing entries of three real sales data sets, and all results show that, compared with competitive algorithms, our proposed model achieve the best performance in terms of prediction accuracy and convergence speed.