結合分類分群技術建立推測法則之研究

Conjecturable Rules Discovery by Clustering-Classification Hybrid Approach

許武先(Wu-hsien Hsu)

指導教授 ： 陳彥良

國立中央大學/管理學院/資訊管理學系/博士(2011年)

摘要

資料探勘的主要目的是發掘隱藏或未知的知識。分類技術可以透過分析具有分類標簽的訓練資料，建立各項法則以便未來對新資料進行分類。然而若資料集並未存在已知的分類標籤，分類技術則無法發揮。而分群技術可將無標籤的資料依據各資料點的相似程度，分成若干群，各群因具有高度相似的屬性值，可將各群歸類為某種概念。雖然分群技術可將無標籤的資料分為特定的數個概念，分群技術的特性卻無法如同分類技術一樣，將分群的規則留下來，以便於未來推測之用。
所謂「推測」係針對不熟悉或無法提供分類標簽之資料集進行兩組不同屬性之分析，期能發掘出兩組資料屬性之關係，進而建立推測的法則。
本研究延伸了先前的研究，提出新的方法，藉以發掘隱性法則與改善推測正確率。除了運用分類技術建立決策樹，作為推測法則，同時以分群方式來解決無標籤資料的困境。也透過模糊理論的實踐與離群值處理，對於隱性法則的發掘，以及正確率的提升都有顯著的結果。實驗結果顯示本研究所提出的方法，能有效建立推測法則，所發掘的規則也可彌補過去方法的缺憾。

關鍵字

資料探勘 ； 分類 ； 分群 ； 推測規則 ； 決策樹 ； 數值分析 ； 模糊理論

並列摘要

Discovering hidden or unknown knowledge is the major theme of most data mining studies. In this dissertation, we propose a new approach to discover conjecturable rules, which categorize observations of a data set into classes of similar attribute values instead of classes of crisp labels. The proposed approach is developed based on the two most developed data mining techniques: Classification and Clustering.
Classification is the problem of identifying the sub-population to which new observations belong. The result is decided according to a set of rules which discovered from a training set of data of observations whose sub-population is known. The technique is known as supervised learning, i.e. pre-defined labels are necessary for the process. The result is a set of rules which are able to predict which label a new observation is belonged to. However, when there is no label existed in the dataset, this technique fails to apply. On the other hand, Clustering is the process of grouping a set of objects into classes of similar objects. No pre-defined label is necessary for the process. It is known as unsupervised learning. Yet no any rule is preserved after the process for future prediction.
The object of this dissertation is to discover conjecturable rules from those datasets which do not have any predefined class label. Furthermore, the technique extends our two previous studies with fuzzy concept and outliers handling. Thus recessive conjecturable rules can be discovered as well as the accuracy is improved. The proposed technique covers the convenience of unsupervised learning as well as the ability of prediction of decision trees. The experiment results show that our proposed approach is capable to discover conjecturable rules as well as recessive rules. Sensitivity analysis is also given for practitioners’ reference.

並列關鍵字

Data Mining ； Cluster Analysis ； Conceptual Cluste

參考文獻

Agrawal, R., Gehrke, J., Gunopulos, D., Raghavan, P. (1998). Automatic subspace clustering of high dimensional data for data mining applications. In: Proceedings of the ACM SIGMOD Int''l Conference on Management of Data, 94-105.

Basak, J. and Krishnapuram, R. (2005). Interpretable Hierarchical Clustering by Constructing an Unsupervised Decision Tree. IEEE Transactions on Knowledge and Data Engineering, 17(1), 121- 132.

Berkhin, P., (2002). Survey of clustering data mining techniques. Technical Report, CA: Accrue Software.

Bezdek, J., (1981). Pattern recognition with fuzzy objective function algorithms. Plenum Press, New York.

Bezdek, J.C., Ehrlich, R., and Full, W. (1984). FCM: The fuzzy c-means clustering algorithm. Computers & Geosciences Vol. 10, Issue 2-3, 191-203.