Supervised machine-learning approaches are frequently applied on biological data to learn a regression or classification model, whether used for prediction, classification or for gaining an understanding on the biological process that has generated the data (Larranaga et al., 2006). Arguably however, it is sometimes the case that sophisticated and complicated methods are employed, published and advocated as advances without a comparison even against the simplest baseline methods. We consider a baseline method as the simplest method that an expert analyst can conceive within a few minutes of consideration of the problem and does not require any engineering or scientific ingenuity or novelty. A baseline can take the form of comparing against predicting by the mean of the outcome on a known dataset, without use of any special predicting variables, or comparing against random guessing. Because of this lack of comparison, the added-value of the sophisticated methods—if any—is not quantified; it remains unknown whether the extra effort for implementing or applying it is worth. A false perception about the difficulty of the problem may be created.
展开▼
