N-gram字符序列能有效捕捉文本中作者的个体风格信息,但其特征空间稀疏度高,且存在较多噪音特征.针对该问题,提出一种基于半随机特征采样的中文书写纹识别算法.该算法首先采用一种离散度准则为每个作者选取一定粒度的个体特征集,然后将个体特征集以一种半随机选择机制划分成多个等维度的特征子空间,并基于每个子空间训练相应的基分类器,最后采取多数投票法的融合策略构造集成分类模型.在中文真实数据集上与基于随机子空间和Bagging算法的集成分类器进行了对比试验,结果表明,该算法在正确率和差异度方面优于随机子空间和Bag-ging算法,并且取得了比单分类模型更好的识别性能.%Character N-gram can be used to effectively capture individual-author stylistic information in texts. To deal with the problems of high-sparsity and high-redundancy in the feature spaceman ensemble classification algorithm based on semi-random feature sampling was proposed in this study. Firstly, the whole feature space is divided into several individual-author feature sets by a divergence rule. Then each of them is divided into equally sized subspaces by a semi-random selection method,and a base classifier is trained on each random subspace. Finally, these base classifiers are combined to construct an ensemble via the majority voting method. To examine the algorithm, the experiment was conducted on a real-life dataset. It is observes that the algorithm achieved a considerable improvement in accuracy and robustness compared with the benchmark technique in Chinese writeprint identification (random subspace method, bagging and support vector machine).
展开▼
