Making room for the modeler in data mining

I would like to use this perspective to advocate further research in data mining toward methods that allow and formalize greater input from the modeler. By ‘modeler’, I mean the domain expert, scientist, business analyst, or researcher who has questions to answer using data. Soliciting input from such users is atypical of most data mining algorithms (e.g., see the introductory discussion in ref. [1]). Although excluding pesky users enables one to enforce protections against, say, overfitting, those same users may come to resent the exclusion. Many fields remain deeply suspicious of black-box models that generate predictions without asking for help or supplying explanations. Handpicked regression models remain the model of choice in the social sciences, for instance, because these allow the user to pick the features and specify the relationships. Regression allows the model to express a theory about how the world works, and the success of the model depends on this specification. To this community, the fact that we can automate the process of selecting features or supply black-box predictions that are more accurate than those of regression is secondary to the absence of interpretative hooks. Allowing the user to customize a statistical model introduces any number of problems. With practice and a bit of collinearity in the data, one can mold a regression model to support a variety of theories by iteratively adding and removing variables. We know how to keep an automated search for features from overfitting; the challenge is to involve users in this process. Modelers who invest years gathering data are not eager to lay down their favorite tools unless we provide them with alternatives that respect their knowledge of the context and their desire to influence the analysis. A modeler who has a better theory and understanding of the data should be able to get a better model than one who lacks these insights. Black-box models remove the competitive advantage: both modelers pour the data into the model and pop out the same predictions. The methodologies that I want to encourage reward the modeler who has useful insights. I have a couple of illustrative examples in mind. First, consider the grouped LASSO [2] in the classical regression setting: for a sample of n observations, we observe a response y and k explanatory features in the vector x. The goal is a precise estimate of E y = x ′β. The original LASSO provides a method for identifying useful features. Rather than relying on stepwise or heuristic searches, LASSO selects the coefficient vector β̂ that solves