The present paper describes a methodological framework developed to select a multi-label dataset transformation method in the context of supervised machine learning techniques. We explore the rectangular 2D strip-packing problem (2D-SPP), widely applied in industrial processes to cut sheet metals and paper rolls, where high-quality solutions can be found for more than one improvement heuristic, generating instances with multi-label behavior. To obtain single-label datasets, a total of five multi-label transformation methods are explored. 1000 instances were generated to represent different 2D-SPP variations found in real-world applications, labels for each instance represented by improvement heuristics were calculated, along with 19 predictors provided by problem characteristics. Finally, classification models were fitted to verify the accuracy of each multi-label transformation method. For the 2D-SPP, the single-label obtained using the exclusion method fit more accurate classification models compared to the other four multi-label transformation methods adopted.