In many business applications the traditional predictive models that were the norm for many years are being replaced with machine learning models. We often hear the statement “machine learning models predict the outcome without being explicitly programmed to do so” and hence require minimum manual work by the modelers.
While many practitioners quote this verbatim, they often miss the real meaning behind. What it really means is that in traditional predictive models the underlying structure of the relationship between the features impacting the outcome needs to be fixed. Hence, the modeler would manually choose the features that have a specific statistical or behavioural relationship with the outcome.
In machine leaning models, the main objective it to maximise the predictive power of the model by using the large volume of data. While the modeler needs to specify the parameters to be used in the model along with the objective function and the higher level constraints, the machine will discover complex interdependencies between the features and outcomes without the need to specify a fixed mathematical formula. In other words without fixing the “functional form” of the relationship between the features and outcome.
This means that the machine can evaluate a much higher number of potential versions and combinations of input features thereby enhancing the predictive power of the final model. On the flip side this complexity in the final model leads to lack of explainability of the model which then contributes to challenges in model adoption by users of the model
Photo by Mykola Makhlai on Unsplash
Comments