Abstract:

Supervised categorization algorithms can yield better categorization performance in datasets with enough and balanced labels. However, various real-world categorization tasks suffer from the class imbalance problem which has been known to hinder the learning performance of categorization algorithms. This paper, demonstrates that SLDA model is capable of solving the class imbalance problem by sampling unlabeled instances. In order to yield a better prediction performance with minority classes, the semantic relationship between topics and minority classes is derived by the SLDA topic model. An efficient way of calculating confidence and sampling valuable unlabeled instances is proposed. The proposed method reduces the skewness of the imbalanced datasets efficiently and improves the categorization performance of minority classes. Our experimental results show that the the proposed method, ITC-SLDA algorithm, can significantly improve Macro-F1 and G-mean values in imbalanced text categorization.

Key words: supervised topic model, semi-supervised learning, imbalanced text, categorization

摘要：

在标签均衡分布且标注样本足够多的数据集上，监督式分类算法通常可以取得比较好的分类效果。然而，在实际应用中样本的标签分布通常是不均衡的，分类算法的分类性能就变得比较差。为此，结合SLDA（Supervised LDA）有监督主题模型，提出一种不均衡文本分类新算法ITC-SLDA（Imbalanced Text Categorization based on Supervised LDA）。基于SLDA主题模型，建立主题与稀少类别之间的精确映射，以提高少数类的分类精度。利用SLDA模型对未标注样本进行标注，提出一种新的未标注样本的置信度计算方法，以及类别约束的采样策略，旨在有效采样未标注样本，最终降低不均衡文本的倾斜度，提升不均衡文本的分类性能。实验结果表明，所提方法能明显提高不均衡文本分类任务中的Macro-F1和G-mean值。

关键词: 有监督主题模型, 半监督学习, 不均衡文本, 分类