Abstract: Gaussian process has the ability to deal with complicated problems such as high dimension, small samples, and nonlinearity. A new algorithm called confidence-weighted learning is proposed by introducing Gaussian process into the online classification. It maintains a Gaussian distribution over weight vectors and makes adjustment over the weight vector with every sample instance. By this way, it correctly classifies examples with a specific probability. Experimental results show that the confidence-weighted classifiers are always more accurate than the perceptron classifier in the simulated datasets, as well as in the real datasets.

Key words: confidence-weighted learning, perceptron, online learning, large margin, Gaussian distribution

摘要： 鉴于高斯过程对处理高维数、小样本、非线性等复杂问题具有较好的适应性，将其引入到在线分类器学习算法中，形成一种新型的在线分类算法，即信任权算法。该算法的信任权超参数为模型向量的高斯分布，每训练一次样本就修正一次模型向量的信任权，并使样本正确分类的概率在某个特定信任域内。采用人工和实际数据进行实验，结果表明信任权算法优于传统的感知器算法。

关键词: 信任权学习, 感知器, 在线学习, 大间隔, 高斯分布

CLC Number: 

• TP301.6