MMDT: تصمیم گیری طبقه بندی درختی چندگانه و چند نشاندار برای داده کاوی
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|22062||2005||14 صفحه PDF||سفارش دهید||6937 کلمه|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : Expert Systems with Applications, Volume 28, Issue 4, May 2005, Pages 799–812
We have proposed a decision tree classifier named MMC (multi-valued and multi-labeled classifier) before. MMC is known as its capability of classifying a large multi-valued and multi-labeled data. Aiming to improve the accuracy of MMC, this paper has developed another classifier named MMDT (multi-valued and multi-labeled decision tree). MMDT differs from MMC mainly in attribute selection. MMC attempts to split a node into child nodes whose records approach the same multiple labels. It basically measures the average similarity of labels of each child node to determine the goodness of each splitting attribute. MMDT, in contrast, uses another measuring strategy which considers not only the average similarity of labels of each child node but also the average appropriateness of labels of each child node. The new measuring strategy takes scoring approach to have a look-ahead measure of accuracy contribution of each attribute's splitting. The experimental results show that MMDT has improved the accuracy of MMC.
The purpose of the decision tree classifier is to classify instances based on values of ordinary attributes and class label attribute. Traditionally, the data set is single-valued and single-labeled. In this data set, each record has many single-valued attributes and a given single-labeled attribute (i.e. class label attribute), and the class labels that can have two or more than two types are exclusive to each other or one another. Prior art decision tree classifiers, such as ID3 ( Quinlan, 1979 and Quinlan, 1986), Distance-based method ( Mantaras, 1991), IC ( Agrawal, Ghosh, Imielinski, Iyer, & Swami, 1992), C4.5 ( Quinlan, 1993), Fuzzy ID3 ( Umano et al., 1994), CART ( Steinberg & Colla, 1995), SLIQ ( Mehta, Agrawal, & Rissanen, 1996), SPRINT ( Shafer, Agrawal, & Mehta, 1996), Rainforest ( Gehrke, Ramakrishnan, & Ganti, 1998) and PUBLIC ( Rastogi & Shim, 1998), all focus on this single-valued and single-labeled data set. However, there is multi-valued and multi-labeled data in the real world as shown in Table 1. Multi-valued data means that a record can have multiple values for an ordinary attribute. Multi-labeled data means that a record can belong to multiple class labels, and the class labels are not exclusive to each other or one another. Readers might have difficulties to distinguish multi-labeled data from two-classed or multi-classed data mentioned in some related works. To clarify this confusion, we discuss the exclusiveness among classes, number of class and representation of the class label attribute in the related works as follows: 1. Exclusiveness: Each data can only belong to a single class. Classes are exclusive to one another. ID3, Distance-based Method, IC, C4.5, Fuzzy ID3, CART, SLIQ, SPRINT, Rainforest and PUBLIC are such examples. 2. Number of class: Data with classes classified into two types in the class label attribute is called two-classed data. ID3 and C4.5 are such examples. Data with classes classified into more than two types in the class label attribute is called multi-classed data. IC, CART and Fuzzy ID3 are such examples. 3. Label representation: Data with a single value for the class label attribute is called single-labeled data. ID3, Distance-based Method, IC, C4.5, Fuzzy ID3, CART, SLIQ, SPRINT, Rainforest and PUBLIC are such examples.According to the discussion above, a multi-valued and multi-labeled data as we defined here can be regarded as a non-exclusive, multi-classed and multi-labeled data. In our previous work (Chen, Hsu, & Chou, 2003), we have explained why the traditional classifiers are not capable of handling this multi-valued and multi-labeled data. To solve this multi-valued and multi-labeled classification problem, we have designed a decision tree classifier named MMC ( Chen et al., 2003) before. MMC differs from the traditional ones in some major functions including growing a decision tree, assigning labels to represent a leaf and making a prediction for a new data. In the process of growing a tree, MMC proposes a new measure named weighted similarity for selecting multi-valued attribute to partition a node into child nodes to approach perfect grouping. To assign labels, MMC picks the ones with numbers large enough to represent a leaf. To make a prediction for a new data, MMC traverses the tree as usual, and as the traversing reaches several leaf nodes for the record with multi-valued attribute, MMC would union all the labels of the leaf nodes as the prediction result. Experimental results show that MMC can get an average predicting accuracy of 62.56%. Having a decision classifier developed for the multi-valued and multi-labeled data, this research steps further to improve the classifier's accuracy. Considering the following over-fitting problems (Han and Kamber, 2001 and Russell and Norving, 1995) of MMC, improvement on its predicting accuracy seems possible. First, MMC neglects to avoid the situation when the data set is too small. Therefore, it may choose some attributes irrelevant to the class labels. Second, MMC appears to prefer the attribute which splits into child nodes with larger similarity among multiple labels. Therefore, MMC exists inductive bias (Gordon & Desjardins, 1995). Trying to minimize the over-fitting problems above, this paper proposes solutions as: (1) Set a constraint of size for the data set in each node to avoid the data set being too small. (2) Consider not only the average similarity of labels of each child node but also the average appropriateness of labels of each child node to decrease the bias problem of MMC. Based on the propositions above, we have designed a new decision tree classifier to improve the accuracy of MMC. The decision tree classifier, named MMDT (multi-valued and multi-labeled decision tree), can construct a multi-valued and multi-labeled decision tree as Fig. 1 shows.The rest of the paper is organized as follows. In Section 2, the symbols will be introduced first. In Section 3, the tree construction and data prediction algorithms are described. In Section 4, the experiments are presented. And, finally, Section 5 makes summaries and conclusions.
نتیجه گیری انگلیسی
This research has designed a decision tree classifier, MMDT, to improve the accuracy of MMC by minimizing the over-fitting problems. In MMDT, we set a constraint of size for the data set in each node to avoid the data set being too small; and we consider not only the average similarity of labels of each child node but also the average appropriateness of labels of each child node to decrease the bias problem. The experimental results show that MMDT has improved the accuracy of MMC. The main works of MMDT can be summarized as follows: 1. To select the best multi-valued attribute: MMDT proposes a new measure named similarity ratio, which combines the measure of weighted label ratio with the measure of weighted similarity, as the measure for selecting the best multi-valued attribute. 2 To grow a multi-valued and multi-labeled decision tree: MMDT constructs a decision tree by traversing and splitting each internal node recursively. When the traversing has reached a leaf node, a label-set would be assigned to the leaf node. To avoid the data set in each node being too small, MMDT sets a threshold. To acquire a smaller decision tree, MMDT sets a degree constraint for each internal node containing numeric attribute to reduce branches of the internal node. The constraint can reduce the number of rules. 3. To assign multiple labels to a leaf: Several thresholds and parameters of MMDT are used to decide whether a node can be a leaf and what label-set can be assigned to the leaf. 4. To make a prediction based on the decision tree: MMDT predicts the class label of a record by traversing the tree from its root until the traversing finally reaches a leaf node. If a record has a multi-valued attribute and the traversing reaches several leaf nodes, MMDT will union all the leaves' label-sets as the prediction result. The capability of classifying multi-valued and multi-labeled data could be applied to many real-world commercial data. It also could be applied to the handling of more complicated data. For example, it makes a multi-valued and multi-labeled meta-data for semi-structured data (Wang et al., 1999 and Zaiane and Han, 1995) or object-oriented data (Han, Nishio, Kawano, & Wang, 1998) meaningful and manageable. This certainly extends our ability in data management. Therefore, continued improvement on the multi-valued and multi-labeled classification work is important.