Helping Novices Avoid the Hazards of Data: Leveraging Ontologies to Improve Model Generalization Automatically with Online Data Sources

ods produce a model from a set of examples. Despite the maturity of these algorithms, decisions that result from models are unlikely to be correct if data have been used indiscriminately. This is part of the so-called data-dredging problem (Smith and Shah 2002). Figure 1 shows a bemusing example: U.S. spending per annum on science, space, and technology is highly correlated with suicides by hanging, strangulation, and suffocation. Logically, we know that this correlation does not imply causation (that is, higher spending on technology cannot cause more suicides by hanging or vice versa). Unfortunately, without the capacity to distinguish real and spurious correlations, learning methods are prone to picking up such correlations in producing models (Tukey 1977). The onus to be judicious ultimately falls on the person building the model. Otherwise, data dredging may lead to specious models, which may overfit, generalize poorly, and suggest conclusions that are fallacious. Articles