Predict or Explain?

Predict or explain? Or otherwise stated, conduct predictive or explanatory analysis? There is a trade-off between the two, driven by model complexity.

Predict or explain? An important decision for the data scientist to take when trying to assess/compare several competitive analytic model(s) and ultimately select one or two for productive use.

We know that the higher the model complexity the better the predictive performance of a learning model. Yet, simpler models prove to be more explanatory (or interpretable) in the sense that they can describe/estimate the effects of the predictors on the response and possibly allow users make inferential claims (how much confident they are) about the estimates of individual effects.

If we think of a model as a mapping between the predictors and the response then this mapping can have dual use: 1) make predictions of the response given specific predictor values, 2) provide "explanations", such as mapping rules or effect weights, that describe the connection between the predictors and the response. A complex model, say a sate-of-the-art ensemble method, can undoubtedly give more accurate predictions, yet a self-explanatory, simpler model, like a linear model fed with features derived from some additional feature engineering, can provide a better description of the predictors-to-response association.

Before starting the assessment/comparison of the many, diverse models that were built to resolve the domain problem at hand, we first need to consider the potential use of modeling. |If our clients' primary purpose is to make predictions, then, we must place emphasis on assessing the predictive performance of the models. If, on the other hand, our clients primarily wish to develop a better understanding of the effects on the response of potential variables, then special emphasis must be placed on the explanatory ability of the models. Once the potential model use has been made clear, we can, then, accordingly employ the appropriate model families and model assessment tools in our analytic solution.