Upsampling, a comparative study with new ideas

We give a brief tour of sampling techniques not uncommon in dealing with imbalanced data having binary tags. We mention the rather ‘naive’ upsampling techniques, classical bootstrap (B. Efron, 1979) and a more recent technique called SMOTE (Chawla, et al, 2002). We demonstrate those techniques using features derived from the Android OS source code with tags derived from the common vulnerabilities & exposures public database (CVE, Mitre) which contains known software bugs in widely-used applications. Evidently, the data which contains approximately 0.5% positive tags with remaining approximately 99.5% negative tags is highly, in fact extremely imbalanced. A new idea is to synthesize samples from a minority class using an idea from the Bayesian domain. Forthcoming is an empirical analysis of the procedure. Keywords: Unstructured Data, Android Source Code Imbalanced Data, Binary Tags, Binary Classification, ROC, AUC, Lift, Specificity, Recall, Upsampling, Boot-strapping, SMOTE, Bayesian Upsampling, EM, Autoregressive