The Research Of Semi-supervised Learning Based On Boosting

Boosting algorithm is a method which use the tactics of the assemble classifier toimprove the performance of the algorithm. In recent years, boosting method has become a hotspot in the area of machine learning. And now there are many boosting method such asadaboost etc. Boosting algorithm has a good performance in machine learning, and it’s a wellboost method. Boosting series of algorithms are widely used in many fields in recent years.Semi-supervised learning is a more popular direction of the current field of machinelearning.Semi-supervised learning can take full advantage of the unlabeled examples in thetraining data. It can train the classifier with a small amount of labeled samples and a largenumber of unlabeled samples. And because semi-supervised learning requires only a smallamount of labeled examples so the algorithm can save the cost used in getting the labels of theexamples. Semi-supervised learning is significant to reduce costs to improve the classifier’sperformance.Boosting method and the semi-supervised learning algorithms are all the algorithmswhich can improve classification performance. In recent years, many experts and scholars inthe study of these two aspects have made good progress. But there are few people whocombine the semi-supervised learning and the Boosting algorithm.In this paper, we will combine the Boosting and the Semi-supervised learning. And weresearch the Data Mining based on the WEKA platform. We implement our algorithm on theWEKA. Our algorithm IMSB has improved the algorithm MCSSB. MCSSB algorithmcombines the Semi-supervised Learning and the Boosting method and it is used to solve themulti-classification problems. It has achieved a good performance in many UCI data sets. Inthis paper our algorithm IMSB gets an improvement based on MCSSB. We add aclassification algorithm at the start of our algorithm to deal with the training data and get apredicted label for the unlabeled data. Experiments results show that our algorithm can get agood performance when the labeled data sets are little. Our algorithm improved the MCSSBin some extent.To predict the unlabeled examples’s labels we compose a object function to guideprediction. We get an object function similar to the object function in MCSSB. We get thepseudo labels of the unlabeled examples with a high confidence. In the meanwhile we get thebase classifier’s weights by optimizing the object function. We combine the base classifierswith different weights and get a composite classifier. We add a supervised algorithm in thefirst stage of our algorithm to predict the unlabeled examples’s labels. We choose the KNNand NaiveBayes in our algorithm. For our algorithm propose in this paper, we will implement the experimental verificationexperiment on the WEKA platform. And we will use the java language to program ouralgorithm and integrate it into the WEKA platform to test our algorithm.