基于费希尔信息度量的随机近邻嵌入算法

Fisher Information Metric Based on Stochastic Neighbor Embedding

摘要: 为提高文本分类的准确率,提出了费希尔信息度量随机近邻嵌入算法(Fisher information metric based on stochastic neighbor embedding, FIMSNE). 首先,把文本的词频向量看作统计流形上的概率密度样本点, 利用费希尔信息度量计算样本点之间的距离;然后,从信息几何的观点出发, 对 t分布随机近邻嵌入( t-stochastic neighbor embedding, t-SNE)进行改进,实现了新算法. 真实文本数据集上的二维嵌入和分类实验的结果表明:FIMSNE的性能在总体上优于 t-SNE、费希尔信息非参数嵌入(Fisher information nonparametric embedding,FINE)和主成分分析(principal components analysis,PCA).

Abstract: To improve the classification accuracy of text classification, Fisher information metric based on stochastic neighbor embedding (FIMSNE) was proposed. In this paper, text word frequency vectors were taken as probabilistic density functions that were points on a statistical manifold, and their distances were defined by Fisher information metric. From the view of information geometry, t-stochastic neighbor embedding ( t-SNE) was improved to FIMSNE. That FIMSNE outperforms t-SNE, Fisher information nonparametric embedding (FINE) and principal components analysis (PCA) in the whole was verified with 2D-embedding and classification task to real text dataset.