基于选择性抽样的SVM增量学习算法的泛化性能研究

Research on Generalization Performance of SVM Incremental Learning Algorithm Based on Selective Sampling

摘要:

针对大数据环境中存在很多的冗余和噪声数据,造成存储耗费和学习精度差等问题,为有效的选取代表性样本,同时提高学习精度和降低训练时间,提出了一种基于选择性抽样的SVM增量学习算法,算法采用马氏抽样作为抽样方式,抽样过程中利用决策模型来计算样本间的转移概率,然后通过转移概率来决定是否接受样本作为训练数据,以达到选取代表性样本的目的。并与其他SVM增量学习算法做出比较,实验选取9个基准数据集,采用十倍交叉验证方式选取正则化参数,数值实验结果表明,该算法能在提高学习精度的同时,大幅度的减少抽样与训练总时间和支持向量总个数。

Abstract:

For the large data environment, there are many redundancy and noisy data, resulting in storage costs and poor learning accuracy problems, in order to effectively select representative samples, while improving learning accuracy and reduce training time, the thesis presents a selective sampling of SVM incremental learning algorithm, based on markov sampling as a sampling method. In the sampling process, the decision-making model is used to calculate the transition probability between samples, and then the transition probability is adopted to decide whether to accept the sample as the training data, in order to select the representative sample. Compared with other SVM incremental learning algorithms, the experiment selects 9 benchmark data sets and uses 10-fold cross-validation to select regularization parameters, and the numerical experiments show that the algorithm can greatly reduce the total time of sampling and training and the number of support vectors while improving learning accuracy.