基于逻辑回归的多任务域快速分类学习算法

发布时间：2018-06-28 23:28

本文选题：多任务分类 + 罗杰斯特回归　；参考：《计算机工程与应用》2017年15期

【摘要】：多任务学习通过寻找并共享不同任务域之间的共性特征来完成学习,利用知识迁移加速不同任务域的学习为每个任务域构建一个分类器。提出了一种基于罗杰斯特回归模型的多任务学习方法 MTC-LR(Multi-task Coupled Logistic Regression)。"罗杰斯特回归模型"已经被成功应用于单任务分类器上,该模型被众多实验证明是有效的,正是这种方法给人们带来了启示。从理论上证明了通过构造多任务分类器的"开销函数"和"差异性度量函数",MTC-LR算法可以提高多任务分类器的各自分类精度。相比传统的基于SVM的多任务学习方法,MTC-LR并不依赖于核方法而是通过共轭梯度下降法寻找各个分类器的最优参数。同时MTC-LR与采用"罗杰斯特回归模型"的快速算法CDdual更容易结合,可扩展至大样本的多任务分类学习。正是基于上述发现,为了充分高效利用大样本的多任务域数据,满足大样本的快速运算,在MTC-LR算法的基础上,结合最新的CDdual(The Dual Coordinate Descent Method)算法,提出了MTC-LR的快速算法MTC-LR-CDdual,并对该算法进行了相关的理论分析。将该算法在人工数据集和真实数据集上进行了验证,实验结果表明该算法有着较高的识别率、快速的识别速度和较好的鲁棒性。
[Abstract]:Multi-task learning accomplishes learning by finding and sharing the common features between different task domains, and constructs a classifier for each task domain by using knowledge migration to accelerate the learning of different task domains. A multitask learning method, MTC-LR (Multi-task coupled Logistic Regression), is proposed based on Rogers regression model. " The Rojester regression model has been successfully applied to single task classifier. The model has been proved to be effective by many experiments. It is this method that has brought enlightenment to people. It is proved theoretically that by constructing the "overhead function" and "difference measure function" of the multitask classifier, the MTC-LR algorithm can improve the classification accuracy of the multitask classifier. Compared with the traditional SVM based multitasking learning method, MTC-LR does not depend on kernel method but uses conjugate gradient descent method to find the optimal parameters of each classifier. At the same time, MTC-LR is easier to combine with CDdual, a fast algorithm based on "Rogers regression Model", and can be extended to multi-task classification learning with large samples. Based on the above findings, in order to make full and efficient use of multi-task domain data of large samples and satisfy the fast operation of large samples, the MTC-LR algorithm is combined with the latest CDdual (the dual coordinate Descent method) algorithm. A fast algorithm of MTC-LR, MTC-LR-CDdual. is proposed, and the related theoretical analysis of the algorithm is given. The algorithm is verified on artificial data set and real data set. Experimental results show that the algorithm has higher recognition rate, faster recognition speed and better robustness.
【作者单位】：华中科技大学光学与电子信息学院;江苏北方湖光光电有限公司;南京大学软件学院;
【分类号】：TP181;TP391.41

【参考文献】