用Boosting方法组合增强Stumps进行文本分类

为提高文本分类的精度,Schapire和Singer尝试了一个用Boosting来组合仅有一个划分的简单决策树(Stumps)的方法.其基学习器的划分是由某个特定词项是否在待分类文档中出现决定的.这样的基学习器明显太弱,造成最后组合成的Boosting分类器精度不够理想,而且需要的迭代次数很大,因而效率很低.针对这个问题,提出由文档中所有词项来决定基学习器划分以增强基学习器分类能力的方法.它把以VSM表示的文档与类代表向量之间的相似度和某特定阈值的大小关系作为基学习器划分的标准.同时,为提高算法的收敛速度,在类代表向量的计算过程中动态引入Boosting分配给各学习样本的权重.实验结果表明,这种方法提高了用Boosting组合Stump分类器进行文本分类的性能(精度和效率),而且问题规模越大,效果越明显.

Improved Stumps Combined by Boosting for Text Categorization

Stumps, classification trees with only one split at the root node, have been shown by Schapire and Singer to be an effective method for text categorization when embedded in a boosting algorithm as its base classifiers. In their experiments, the splitting point (the partition) of each stump is decided by whether a certain term appears or not in a text document, which is too weak to obtain satisfied accuracy even after they are combined by boosting, and therefore the iteration times needed by boosting is sharply increased as an indicator of low efficiency. To improve these base classifiers, an idea is proposed in this paper to decide the splitting point of each stump by all the terms of a text document. Specifically, it employs the numerical relationship between the similarities of the VSM-vector of text document and the representational VSM-vector of each class as the partition criteria of the base classifiers. Meanwhile, to further facilitate its convergence, the boosting weights assigned to sample documents are introduced to the computation of representational VSM-vectors for possible classes dynamically. Experimental results show that the algorithm is both more efficient for training and more effective than its predecessor for fulfilling text categorization tasks. This trend seems more conspicuous along with the incensement of problem scale.