| Most of the traditional machine learning algorithms are based on batch learning model,and the model training must be started after a large number of samples have been obtained.However,nowadays the amount of data grows rapidly in our society.When the new data arrives,the retraining of the model will consume lots of time and computational costs.The dimension of novel social networks data and financial data may higher than one hundred thousand.Feature selection can help the algorithm to eliminate low correlation and redundant features,to reduce the number of features and computational cost,at the same time,avoids some disadvantages,such as much too complex model and weak normalization capability.In this paper,we study the feature selection algorithm based on online learning.For the single task of online learning feature selection algorithm,this thesis presents a new learning algorithm named PABS based on online feature selection by linking online feature selection algorithm and efficient PA online learning algorithm.The PABS algorithm can achieve the same or even higher classification accuracy than the PA algorithm,while still enjoying the good characteristics of feature selection algorithm.The algorithm can keep the core features of the classification with less computing resources,and reduce the burden of model training.When the single task training data is not enough,it can significantly improve the generalization ability of the algorithm by using a number of interrelated tasks.For multitask learning,this thesis proposed the CMOFS algorithm.The algorithm build a global model and many Independent model at the same time.And according to the model prediction effect,dynamically adjusting the weights between the global model and the Independent model.Finally improve the learning effect of all tasks.In this paper,we use the classical machine learning data set to test the performance of each algorithm.The Experiments show that algorithms in this thesis can get better training effect. |
PDF Full Text Request