Research And Application Of Active Learning In Batch Mode

In many data mining and machine learning applications,obtaining the labels of samples is very difficult and time-consuming,and sometimes costly,because they may require a lot of labor.However,unlabeled samples are ubiquitous because they are available in large quantities and cheaply.Active learning attempts to select the most uncertain and important unlabeled samples for manual labeling and use as few labeled samples as possible to train an effective classifier.Therefore,active learning is an important learning paradigm in machine learning and data mining.In the field of active learning,the key is how to select the appropriate unlabeled samples for manual annotation,and the selection method is the so-called query strategy.This paper mainly studies the batch model active learning algorithm under the pool based active learning scenario,including the following two parts:[1].We propose a batch mode active learning algorithm combining active learning and semi supervised learning.Specifically,we fully consider the informativeness and representativeness of unlabeled samples and propose a stricter risk constraint of active learning.Based on this risk constraint,we derive the objective function of active learning in batch mode.Then,we propose an alternating direction multiplier algorithm to solve the objective function.The essence of the algorithm is to train a semi supervised classifier,and alternately select informative and representative unlabeled samples for manual labeling.In particular,in order to avoid retraining the semi supervised classifier after each query,we design two unique processes based on path tracking technology,which can delete multiple query samples from the unlabeled sample set,effectively add the query samples to the labeled sample set,and directly update the new semi supervised classifier.A large number of experimental results on various benchmark datasets and real-world datasets show that our algorithm not only has better generalization performance than the existing active learning methods,but also shows its remarkable efficiency.[2].We propose a batch mode active learning algorithm combining active learning and reinforcement learning.Specifically,we use reinforcement learning to learn the process of sample selection in active learning,and use the selection strategy of deep reinforcement learning to replace the heuristic sample selection strategy of traditional active learning.We use semi-supervised support vector machine as the model classifier,and select more valuable sample points to train the model through reinforcement learning,which improves the accuracy of the algorithm.In particular,in reinforcement learning,in order to accelerate the interaction between agents and the environment and get new state and feedback rewards,we propose a security sample screening rule to accelerate the training of semi-supervised support vector machine model.A large number of experimental results on various benchmark data show that our algorithm can effectively learn the selection strategy of active learning.