PCA and projection pursuits on high dimensional data reduction

Abstract

The flood of multimedia data has affected various fields of research, including multimedia retrieval, database management, data mining, machine learning, social media analysis, image processing and so on. Many digital multimedia data collections are made in various fields and application contexts not only in advertising or journalism, but also often used in the context of cultural heritage applications, web search, or geographical information. Digital multimedia data is created in various formats and media modalities and available worldwide every day. The higher the quantity of data, the higher the complexity, diversity, dimensions and multy modality, so that multimedia data grows exponentially. To reduce the impact of high-dimensional data, data reduction is carried out. Data reduction plays an important role in classification. Therefore the selection of data reduction techniques must be done carefully. Principal Component Transform is widely used in data reduction. This analysis attempts to eliminate correlations between bands and further determines the optimal linear combination of original data dimensions for variations in the value of data pixels. The mathematical principle of Principal Component depends on the decomposition of eigenvalues from the covariance matrix of the data matrix. When multimedia data is smaller or covered by data noise, Principal Component is no longer efficient at reducing. To overcome this, Projection Pursuit was introduced using sphering matrices and invariant translations in the transformations. In Projection Pursuits, only data that is considered interesting will be transformed. Optimization of reduced data selection is based on the maximum value of the projection index generated. The projection index used is Skewness and Kurtosis because it is considered in accordance with the data used, the CASI image of Bogor city.