Study On Classification Algorithms Based On Application Field Character

With the rapid development of computer technology, we face manynew challenges in the classification field. For example, in the field ofbioinformatics, we could get the gene expression profile using DNAmicroarrays, and through analysis of the data we have got, we could minemeaningful biological knowledge and information out of the data.Because Biological Datasets are typically Long, recent research hasconcentrated on sample classification and feature set selection; In thefield of color transfer, recent research concentrated on applyingMulti-valued and Multi-labeled to multi-images color transfer; In thefield of wireless sensor networks, according to the features of eventdetection and sensor networks, how to construct distributed and weightedalgorithm for fault-tolerant event region detection also proposed newdemands to classifier. Aiming at these problems, this paper concentrateson the algorithms of classification. The work has been done and the mainachievements obtained are presented below:1. Aiming at the classification of gene expression data of the tumoursubtype and the feature gene selection, an all-round index called GBindex is proposed to eliminate the irrelevant genes and noise, bycombining the Gini index and the Bhattacharyya distance; Theinformative genes, according to the Euclidian distance, are saved byfurther deleting some redundant genes with strong relevancy; the featuregenes are selected from the informative genes with searching approaches.We obtained 2 feature sets that contained 4 feature genes from the acuteleukaemia data set and 1 feature set that contained 7 feature genes fromthe SRBCT data set. Using SVM and ANN as the classifier, the accuracyof classification is 100ï¼…. The feature sets in this work are more compactand superior than previous work.2. Combining classification and mining of association rules, weconstructed the classifier based on closed patterns. Gene expression datarepresent three status of genes: Rising, descending and unchanging. Theclassic algorithms based on association rules, e.g. CBA, are mining all thefrequent itemsets. In the frequent itemsets mining of gene expression profile, we may obtain many redundant and valueless rules. Theclassification algorithm based on closed pattern is one of the approachesto solve this problem. This work proposed a classification algorithm,DMAC, based on closed pattern. According to the feature of geneexpression profile, we build row FP-tree based on row enumeration, thenuse PEA to mine classed patterns. We also proposed weight algorithm QZto improve the performance of classifier, reclassifying the, unrecognizedgenes by constructing weight function. This algorithm is also proved tobe correct and efficient by experiments in four-class data sets. Thisalgorithm is proved to be efficient in multi-class classification.3. This paper analyzed the algorithm of MMC and MMDT, andproposed a new formula Sim3 which considered both similarity andbehavior coherence. The decision tree algorithm SCC_SP based on thisformula therefore has better performances than MMC and MMDT.Aiming at the color transfer, this work analyzed the color transferalgorithm of single image, and proposed Multi-labeled Decision Treealgorithm of multi-images color transfer. Selecting multi-images as thetraining samples and constructing Multi-labeled Decision Tree accordingto the color and texture information of source images, constructingMulti-labeled Decision Tree and making use of the Multi-labeledDecision Tree for classification, this work realized the color transfer ofobject image. This algorithm tried and realized color transfer ofmulti-images which was superior to color transfer of single-image byproviding more reference information.4. Aiming at the problem of fault-tolerant event region detection inwireless sensor networks, this paper proposes a distributed and weightedalgorithm for fault-tolerant event region detection. Considering thefault-tolerant event region of neighborhood's neighborhood, we first useinformation exchange between neighbor nodes and their nearby nodes toestimate the status of the neighbor nodes. Then we use the weightedfault-tolerant algorithm to fuse the status of neighbor nodes for faultdetection of the central node. The simulation results show that 90ï¼…offaults can be detected and corrected using this algorithm, even when 20ï¼…nodes are faulty. Compared with other detection algorithm, the proposed algorithm improves the accuracy of detecting fault-tolerant event regionas well as the fault detection of boundary nodes of the event region, and itachieves a better balance between detection accuracy and energy usage.