An Experimental Platform Integrating Variety Clustering Techniques And Its Application On 3D Model Retrieval

The researches on KDD rise to extract valuable information from thehuge databases. As a branch of KDD, the aim of cluster analysis technique isto obtain "a collection of data objects that are similar to one another withinthe same cluster and are dissimilar to the objects in other clusters". Thecluster analysis techniques are widely applied in data mining, patternrecognition, content-based image retrieval etc. And clustering analysistechnique can be studied to apply it on 3D model retrieval field.Recently, clustering analysis technique often is used to handle the largehigh-dimensional databases, but the increase of the dimension leads to theworse clustering result. So cluster analysis techniques should work togetherwith other techniques. For instance, the dimensionality reduction techniquecan reduce the side-effect of the dimensionality curse;the samplingtechniques have been applied to increase the efficiency of the clusteringalgorithm. As clustering algorithm is complex and difficult to understand, thevisualization techniques can remedy the complexity of the clusteringalgorithm by its intuition and help user better understand the clusteringprocess and result.It is a critical problem for the clustering analysis techniques to selectappropriate value for parameter(s), which is also an important researchdirection in KDD. The most clustering algorithms require the user-specifiednumber of final clusters k. However, it is very difficult to choose anappropriate k because of lacking the valuable prior knowledge. The clusteringalgorithms are also short at detecting outliers and waste the valuableinformation.First part of work of this article is on study of the new hierarchicalclustering algorithm. The main part of work proposes a new outlier identifyalgorithm and the several new auto-stopped hierarchical clustering algorithmsby the study on the several hierarchical clustering algorithms and outlieridentify algorithm. The new outlier identify algorithm refers to a moregeneral case, the even distribution pattern, considers the local character of thedistribution and uses only one parameter. The article proposes three newalgorithms such as CURED, ASHCA and AS-ROCK by amending thetraditional hierarchical clustering algorithms' lack. Compared to thetraditional method, they have the following merits: the proposed algorithmsmake clustering work together with outlier detection quite well, handle theseoutliers very well, auto-stop hierarchical clustering by regarding outliers asvaluable information and cancel the parameter k. CURED amends CUREalgorithm only, but ASHCA develops the whole hierarchical algorithms bythe auto-stopped idea of combining outlier identify;AS-ROCK need lessparameters, and it is also suitable for both real attribute and categoricalattribute.With the rapid improvements on clustering analysis techniques, itbecomes important to evaluate variety clustering algorithms in a friendenvironment in order to explore their usages in the new application fields.However, traditional software cannot meet this requirement when they stilladopt the old fashion clustering algorithms. Therefore, the article introducesan experimental platform, which realizes proposal new hierarchical clusteringalgorithms, many newly emerged clustering algorithms and integrates theoutlier detection, dimensionality reduction and visualization techniques withcluster analysis technique. We can integrate all kinds of clustering algorithmseasily in the platform. The platform provides a friendly method for users toadjust parameters' value and select suitable dataset for clustering algorithms.The article analyses and compares the several kinds of clustering algorithmson the different datasets by the visual tools provided in the platform.In the field of shape-based 3D model retrieval, clustering analysistechnique will have wild application foreground, but a large number of 3DModels have not been classed perfectly, so we can't know the final clusternumber of the clustering algorithm in advance, and it is lack of the availableprior knowledge, which need induct unsupervised and auto-stop clusteringalgorithm. The article choose two two typical auto-stopped clusteringalgorithm in the platform, and applies them to 3D Models' feature vectordataset to study their effect comparatively.In order to study the clustering techniques' application on 3D modelretirval, the article implements a 3D model retrieval system prototype, andintroduces the cluster results obtained in clustering platform to the systemand test their effect. The system adopts java program language and MVCstruct, and it has a better cross-platform and transplantable property. Thegreat character of 3D model retrieval system is to introduce clusteringteniques into it and use clustering algorithm to organise model database.The study of 3D model retrieval system based on the clustering analysistechniques validates that the system can reduce retrieval time, improveretrieval speed and have a high precision by clustering algorithms.The study of clustering algorithms, clustering analysis platform and theirapplication on 3D model retrieval still needs improvement.As a new researchtopic about 3D model retrieval, it is not clear that some clustering algorithmsare suitable to the field. The new type clustering algorithms proposed in thearticle still need improvement, or we need study new algorithms for bettereffect to apply on 3D model retrieval field. And In the real fields such aselectronic government, clustering techniques should have a wide applicationforeground, so the future work will improve the clustering analysis platformto apply it in the real field such as electronic government. The clusteringtechniques' application in 3D model retrieval field is not enough, the futurework will continue to study to use the clustering result to impove the effect ofthe retrieving 3D model database largely.