Extraction of nonlinear structural change of time series from incomplete large-scale data

作为理论观点，我们提出了以下模型，并用统计语言S实现了这些模型的程序。为了验证所提出的方法，我们进行了案例研究的护理服务评估数据，POS（销售点）数据与客户ID和Web访问日志数据。此外，我们还参加了由日本运筹学会等主办的数据分析竞赛。1、研究了提取具有多元历史的个体类型的方法。我们使用SOM（自组织映射）来提取在特定时间点的基本类型的个人。接下来，我们使用分布函数定义SOM图上基本类型分布之间的新距离。然后，我们绘制行为类型的分布图，以便通过SOM获得长期的客户类型（Seki等人，2006）。此外，我们还考虑了一种基于空间拓扑的索引方法。2、提出了一种提取非均匀时间序列非线性结构变化的模型。(1)我们提出了一种模型合并方法，以获得一个分段的变量之间具有齐次函数关系的分割。(2)我们推广SOM时，有两个变量组。我们提出了两阶段SOM，它保留了两个变量组的每个结构，并提取类型的个人。

As theoretical viewpoints, we proposed the following models, and implemented the programs of these models with the statistical language S. In order to verify proposed methodology, we made case studies about care service evaluation data, POS (Point Of Sales) data with customer ID and Web access log data. In addition, we participated in the data analysis competition sponsored by the Operations Research Society of Japan etc..1, We consider methods to extract types of individuals which have multivariate history. We use SOM (Self-Organizing Maps) to extract basic types of individuals at a specified point in time. Next, we define new distances between distributions of basic types on the SOM map using distribution functions. We then map the distributions of behavior types in order to obtain customer types over the long-term by SOM (Seki, et al. 2006). In addition, we consider a method to make index using spatial statistics.2, We propose models which extract nonlinear structural changes in heterogeneous time series.(1) We propose a model merge method, in order to obtain a segmentation whose segment has a homogeneous functional relation between variates.(2) We generalize the SOM when there are two variable groups. We propose two stage SOM, which retains each structure of two variable groups, and extracts types of individuals.