Research And Application Of The Semi-supervised Algorithm Based On Classifier Combination

Traditional supervised learning methods generally require a large amount oflabeled data to accomplish their training tasks. But in the practical applications, due toresources, human and insurmountable conditions, obtaining vast labeled examples isdifficult. The lack of labeled samples is one of the important bottlenecks forsupervised learning methods. On the other hand, unsupervised learning methods donot require labeled samples, but the lack of the effective guidance, model accuracy isnot guaranteed. Therefore, Semi-supervised learning is the newly proposed theory thatfocuses on learning both labeled and unlabeled samples．Studies on it are valuable. Onthe one hand, semi-supervised learning uses vast unlabeled data to assist in training amore effective classification. On the other hand, the results of unsupervised learningare more effective and reasonable with the guidance of prior knowledge.Since classifier combination algorithm can to some extent improve theperformance of a single classifier. The technique has been widely used in variousfields of Data Mining. In this paper, we discuss the semi-supervised classifierextensional algorithm base on classifier combination algorithm. At the same timeintroduces a density-sensitive metric which can effectively describe the clusteringinformation. Two semi-supervised algorithm is proposed correspond to the twogenerate form of the ensemble. Finally It was applied to texture image classification.The main works are listed follows:1.To better utilize clustering information in the co-training algorithm, adensity-sensitive metric is adopted to construct clustering information graph. Basedon this structure distribution information, a novel co-training algorithm is proposedand improves the performance of semi-supervised learning algorithm. Experimentsshow that the algorithm proposed by this paper is more effective.2.There are two forms to generate Combining Classifiers: parallel or serial form.The algorithm mentioned above is the former. To further strengthen and improveClassifiers combination for semi-supervised extension. the density-sensitive metric isalso introduced into the serial algorithm and an new semi-supervised algorithm baseon AdaBoost is proposed. It improves the performance of the AdaBoost. 3.To solve the practical problem of classifying the texture image, dual treecomplex wavelet transform is used to extract the features of texture image since it hasadvantage on direction selectivity, translational invariance and limited redundancy.And then the improved semi-supervised learning algorithm is applied on it. Improvethe accuracy of texture image classificationThe proposed algorithms were tested on some public datasets and real texturedatasets. The experimental results show that our studies on semi-supervised learningare feasible, which will extend the existing studies and have certain application value.