Identification of cancer biomarker using gene expression data is a challenging task. Many strategies have been proposed to identify signature genes for distinguishing cancer from normal cells. Recently, ANOVA-based Feature Set (AFS) has been used to successfully identify the gene sets as markers from multiclass gene expression data. Nevertheless, AFS does not take network data into consideration, resulting in gene-set markers that may not be functionally related to the cancer. Thus, in this work, a gene-set-based biomarker identification method termed Gene-Network-based Feature Set (GNFS) is proposed by integrating gene-set topology derived from expression data with network data. For each gene-set, GNFS identifies a subnetwork as a marker by superimposing those genes onto the network obtained from pathway data and gene-gene relationship, and applying greedy search to identify gene subnetworks. Then, the representative level of each gene-set or gene-set activity is calculated based on the best subnetwork and utilized in cancer classification to evaluate the potentiality of the identified markers. In a comparative study, the classification performance of GNFS is benchmarked against two existing methods, i.e., AFS and Paired Fuzzy SNet (PFSNet). Besides, the identified markers are validated using the online text-mining tool HugeNavigator. The results show that the use of GNFS provides more biologically significant markers while maintaining comparable classification performance.