Data representation has a significant impact on how well the standard multi-viewmachine learning algorithms perform. Existing data representation methods commonly construct a latent subspace such that transformed projections of sample datafrom multiple views are maximized in the common subspace. Different views aregenerally heterogeneous and may contain either similar or complementary information. To be able to remove common information from multiple views such that thedata contains unique information, an extension of the Canonical Correlation Analysis is presented in this paper as the Generalized Similarity Distance based CanonicalCorrelation Analysis (GSDCCA). The proposed approach exploits complementaryand coherent information between the views and intrinsic structural informationwithin the view to build a comprehensive latent space for multi-view machine learning algorithms. The Bhattacharya similarity distance approach is used to exploit thecomplementary and coherent information between the views to project the sampledata onto a common subspace for correlation analysis. Moreover, local neighbouring information is exploit as to obtain the view that preserves the local structureinformation of sample data while performing global dimensionality reduction. Experiments on real-world datasets and synthetic datasets demonstrate that the GSDCCAapproach can build a vast latent space that consistently captures the complementaryand coherent information between views and intrinsic structure information withinthe view. The performance of the proposed approach was compared to other relatedapproaches, and the results show that it is an effective and promising approachfor real-world applications and superior to other state-of-the-art approaches. It isalso observed that the method is highly suitable where views contain lots of similarinformation.