基于信息几何构建朴素贝叶斯分类器

Constructing Naive Bayesian Classifier Based on Information Geometry

朴素贝叶斯分类器是机器学习中一种简单而又有效的分期方法。但是由于它的属性条件独立性假设在实际应用中经常不成立，这影响了它的分类性能。本文基于信息几何和Fisher分，提出了一种新的创建属性集的方法。把原有属性经过Fisher分映射成新的属性集，并在新属性集上构建贝叶斯分类器。我们在理论上探讨了新属性间的条件依赖关系，证明了在一定条件下新属性间是条件独立的。试验结果表明，该方法较好地提高了朴素贝叶斯分类器的性能。

The Naive Bayesian Classifier （NBC） is a simple yet effective technique for machine learning. But the unpractical condition independence assumption of the Naive Bayesian Classifier （NBC） greatly degrades the performance of classifying. This paper improved this method based on information geometry theory and Fisher score. We map the original attributes to new attribute set according to the Fisher score, and construct the NBC on the new attribute set. We further prove that these new attributes are condition independent of each other on certain conditions. This method shows excellent performance in experiments.