Array-based comparative genomic hybridization (array-CGH) has been used to detect DNA copy number variations at genome scale for molecular diagnosis and prognosis of cancer. A special property of arrayCGH data is that, among the spot-intensity variables in the arrayCGH data, there are spatial relations introduced by the layout of the probes along the chromosomes. Standard classification algorithms are not capable of capturing the spatial relations for accurate cancer classification or biomarker identification from the arrayCGH data. We introduce a hypergraph based learning algorithm to classify arrayCGH data with spatial priors modeled as correlations among variables for cancer classification and biomarker identification. In the experiments, we show that, by incorporating the spatial relations among the spots as prior, our algorithm is more accurate than other baseline algorithms on a bladder cancer array-CGH data. Furthermore, some discriminative regions identified by our algorithm contain genomic elements that are cancer-relavent.