Computational Methodology for Knowledge Discovery

从各个领域收集的数据量呈指数级增长，分析数据并从中提取有用信息的任务也变得越来越困难。为了从数据中提取有用的信息，提取过程和数据之间必须有某种适当的交互。通过相互作用，各种过程，如记忆某些信息，学习，进化，并可能发现知识将被执行。从海量数据中自动提取知识的主要障碍是计算资源的限制。A03小组旨在提出和发展知识发现的计算模型和方法。为了实现这一目的，我们探讨了各种主题，包括处理异构数据的算法，这些数据可能是强结构化的，也可能是低结构化的。在本项目的研究成果中，涉及到从超大型数据库中寻找高效有效规则的计算机制的成果有：通过查询学习从大型数据库中高效挖掘；AdaBoost自适应采样方法的改进基于树的线性分类器提升；高斯密度估计的极大极小策略。此外，还开发了解决某些具体问题的算法；一种寻找最佳子序列模式的实用算法生物序列压缩算法。通过压缩方案学习；样本大小对文本分类的影响运用实验方法和理论方法进行知识发现；基于mdl压缩的定义句共性发现。

The amount of data collected from various fields is growing exponentially and the task of analyzing data to extract useful information behind it is becoming more and more difficult accordingly. To extract useful information from data, there must be certain appropriate interaction between the extraction process and data. Through the interaction various processes, such as memorizing certain information, Iearning, evolution, and possibly discovering knowledge will be performed. The major hurdles to automatically extracting knowledge from huge amount of data is the limitations on computational resources. Group A03 aims to propose and develop computational models and methodologies for knowledge discovery. To achieve the purpose we explore various topics including algorithms dealing with heterogeneous data which may be strongly structured or poorly structured.Among the results of this project, the ones concerning computational mechanisms to find efficiently effective rules from very large databases are as follows : Efficient mining from large databases by query learning ; A modification of AdaBoost for adaptive sampling methods ; Tree-based boosting using linear classifier ; The minimax strategy for Gaussian density estimation. Furthermore, algorithms to solve certain concrete problems are developed ; A practical algorithm to find the best subsequence patterns ; Biological sequence compression algorithms - Learning via compression schemes ; Effect of sample size in text categorization ; Knowledge discovery by using both experimental and theoretical methods ; Discovery of commonality among definition sentences by MDL-based compression.