Development of Knowledge Discovery Systems by using the Hierarchical Bayesian Time Series Models

在这个项目中，我们处理了去除人工噪声的问题，这是在从大规模时间序列数据中执行知识发现的自动过程中的绊脚石。更具体地说，我们关注的问题是排除由于观测仪器灵敏度变化而导致的趋势（背景平均值）的快速变化，并识别异常值。本文采用层次贝叶斯模型中的自组织状态空间模型来解决这些问题。即使时间序列中的噪声分量表现出随时间变化的结构，它也能估计出趋势分量；例如，它的方差取决于时间。我们开发的程序使我们能够自动检测大规模时间序列数据集的调谐相关平均结构。我们希望这个程序能够为我们打开一扇门，重新分析由于各种噪声明显污染信号而没有详细检查的大量累积数据。我们已经在Web.http://tswww.ism.ac.jp/higuchi/index e/Soft/index.htm上发布了这个程序和使用说明

In this project, we dealt with a removal of the artificial noises that is a stumbling block in the effort to perform an automatic procedure for knowledge discovery from a large-scale time series data. More specifically, we focused on the problems to exclude a rapid change in a trend (background mean) due to changes in sensitivity of the observation instruments, and to identify an outlier. The self-organizing state space model which belongs to the hierarchical Bayesian model has been employed to solve these problems. It is capable of estimating the trend component even if the noise component in a time series shows a time-dependent structure ; ex., its variance depends on time. The program that we developed allows us to detect the tune-dependent mean structure automatically for a large-scale time series datasets. We hope this program will open a door for us to re-analyze huge accumulated dataset that has not been examined in detail owing to an apparent signal contamination by various noises. We have already post this program with an explanation for usage on Web.http://tswww.ism.ac.jp/higuchi/index e/Soft/index.htm