动态置信度的序列选择增量学习方法

贝叶斯在训练样本不完备的情况下,对未知类别新增训练集进行增量学习时,会将分类错误的训练样本过早地加入到分类器中而降低其性能,另外增量学习采用固定的置信度评估参数会使其效率低下,泛化性能不稳定.为解决上述问题,提出一种动态置信度的序列选择增量学习方法.首先,在现有的分类器基础上选出分类正确的文本组成新增训练子集.其次,利用置信度动态监控分类器性能来对新增训练子集进行批量实例选择.最后,通过选择合理的学习序列来强化完备数据的积极影响,弱化噪声数据的消极影响,并实现对测试文本的分类.实验结果表明,本文提出的方法在有效提高分类精度的同时也能明显改善增量学习效率.

Incremental Learning Method of Dynamic Confidence Level and Sequence Selectable

Under the condition of insufficiency of the training sets, Bayesian will easily make the classification of the new incremental and unlabeled training texts incorrectly. If these incorrectly labeled texts are added to the Bayesian classifier early, it will reduce the performance of Bayesian classifier. In addition, incremental learning with fixed confidence level parameter will cause low learning efficiency and instable generalization ability. In order to solve the above problems, this paper proposes an incremental learning method of dynamic confidence level and sequence selectable. Firstly, the new incremental training subsets are made up of these texts which are classified by current Bayesian classifier correctly. Secondly, it uses confidence level to dynamically monitor the performance of classifier, and then chooses texts from the new incremental training subsets. Finally, strengthen the positive impact of the more mature data, weaken the negative impact of the noise data, and complete the text classification of the test sets by choosing reasonable learning sequence. The experimental results show that the classification efficiency and precision are both advanced by using the method this paper proposes.