Autonomous Data Mining System based on Constructive Learning

提出了一种动态偏差选择预测规则的自主发现方法。研制了原型系统，并通过实验验证了其有效性。在当前的数据挖掘系统中，用户既要参与数据集的预处理，又要参与知识发现。为了减轻他选择和调整多种挖掘算法的负担，我们提出了一种基于构造归纳法的知识发现系统，该系统可以自主选择学习方法。我们的任务是预测规则发现。预测规则的目的是预测未知样本的类别，由于其在探索性数据分析和知识库自动构建等各个领域的有用性，值得特别关注。我们的方法包括两个阶段：1)通过自主离散化对数据集进行预处理；2)通过知识表示的自主决策和评价标准的自主调整来发现知识。我们的方法基于新颖的数据驱动标准和约束，选择适当的偏差，每个偏差都是学习算法的一个组成部分。可用的偏差有等频法和最小熵法进行离散化；知识表示的连接规则和M (N)规则；j评价标准的测度性和预见性。我们的方法已经通过47个真实世界数据集（如零售数据）的发现任务进行了验证。讨论了预测规则发现的定量评价标准，提出了交叉验证的j测度。与最佳偏差组合相比，我们的方法在30个任务中实现了90%以上的交叉验证j测度。仔细分析表明，除非提供的数据集非常小，否则我们的方法是有效的。我们还假设了一个大规模的数据集，并在多台个人计算机上开发了一个并行系统。

This research presents an autonomous method for discovering prediction rules with dynamic bias selection. A prototype system has been developed, and its effectiveness was demonstrated by experiments.In current data mining systems, an user is involved in both pre-processing of a data set and knowledge discovery. In order to reduce his burden of choosing and adjusting multiple mining algorithms, we propose a knowledge discovery system, which autonomously selects learning methods based on constructive induction. Our task is prediction rule discovery. A prediction rule, which is aimed at predicting the class of an unseen example, deserves special attention due to its usefulness in various domains such as exploratory data analysis and automatic construction of a knowledge base.Our method consists of two phases : 1) pre-processing of a data set by autonomous discretization ; 2) knowledge discovery by autonomous decision of knowledge representation and autonomous adjustment of evaluation criteria. Our method, based on novel data-driven criteria and constraints, selects appropriate biases, each of which is a component of a learning algorithm. Available biases are an equal-frequency method and a minimum entropy method for discretization ; a conjunction rule and an M of N rule for knowledge representation ; J-measure and predictiveness for evaluation criterion.Our approach has been validated using 47 discovery tasks with real-world data sets such as retail sale data. We have discussed quantitative evaluation criteria for prediction rule discovery, and proposed J-measure with cross-validation. Our method) compared with the best combinations of biases, achieved more than 90% J-measure with cross-validation in 30 tasks. Careful analysis revealed that our approach is effective unless provided data set is extremely small. We have also assumed a large-scale data set, and developed a parallel system on multiple personal computers.