Research on the Methodology of Information Extraction and Knowledge Discovery Based on Statistical Time Seeries Modeling

对这笔赠款的研究涉及以下四个领域：(1)统计建模的新方法，(2)基于统计建模的知识发现，(3)实际数据分析的应用和(4)软件开发。对于(1)，改进了蒙特卡罗滤波中的重采样方案，通过自组织状态空间模型稳定了模型的自动估计。此外，还尝试了几种新的方法，以缓解固定参数估计中经常遇到的数值困难。(2)针对基因芯片数据分析，提出了一种新的贝叶斯网络估计算法。利用时间序列模型、主成分分析和聚类分析对分分钟股指数据进行了分析，探讨了日历效应。(3)基于海底地震仪数据的多元时间序列模型，提出了一种估计地下结构的近似时空平滑算法。对大规模场向电流进行了详尽的分析，提出了一种自动识别方法。(4)实现了单变量和多变量AR模型估计程序的并行化。这是通过对Household变换进行并行化来实现的。一般状态空间模型的滤波和平滑算法也仅在实验基础上实现。

Research upon this grant has been concerned in the following four fields ; (1) new methodologies in statistical modeling, (2) knowledge discovery based on statistical modeling, (3) application to actual data analysis and (4) software development. As for (1), resampling scheme in Monte Carlo filter is improved to stabilize the automatic model estimation via self-organizing state-space model. Also, several new method are tried to alleviate numerical diffculties often encountered in estimation of fixed parameters. (2) New algorithm for estimating Bayesian network is proposed which is designed for cDNA micro array data analysis. Minute by minute stock index data is analyzed by time series models, principal component analysis and cluster analysis to explore the calendar effect. (3) Based on multivariate time series model for ocean bottom seismograph data, an approximate space-time smoothing algorithm is proposed to estimate underground structure. Large-scale field-aligned currents are exhaustively analyzed, and automatic identification procedure is proposed. (4) Programs to estimate univariate and multivariate AR models are parallelized. This was done by the parallelization of Householder transformation. Filtering and smoothing algorithm for general state-space models is also implemented only on an experimental basis.