A Multi-strategy Approach For Mining Multimedia Data Repositories
Free (open access)
H. L. Viktor & E. Paquet
This paper introduces a multi-strategy learning environment, consisting of a number of heterogeneous classifiers and cluster analysis techniques, used for the mining of so-called integral data records containing 3-D objects, together with a large number of relational attributes. In this multi-strategy learning approach, the two data types are mined together, thus focusing on the interrelationships between them and utilizing their unique characteristics in a synergic way. Experimental results, when applying this approach to an anthropometric database, which contains 3-D human body scans together with (relational) demographic and anthropometric attributes, show promising results. Through the use of cluster analysis and information fusion, a streamlined grouping of human subjects, based on clothing size, was obtained. In addition, the demographic data of each cluster was explored by means of a classifier in order to obtain new insights into the population. Keywords: multi-strategy learning, hybrid systems, cluster analysis, classification, 3-D objects, integral data records, information fusion. 1 Introduction Multimedia databases, which are used to store huge amounts of image, audio and video data, amongst others, are increasingly becoming commonplace [1, 2]. These databases are abundant in many domains, including medicine, anthropometry, bioinformatics, heritage applications and the arts. Consider a multimedia database consisting of so-called integral data records, containing three-dimensional (3-D) objects, together with traditional relational data. This type of database is inherently huge and unstructured, containing diverse data
multi-strategy learning, hybrid systems, cluster analysis, classification, 3-D objects, integral data records, information fusion.