We are now reaching the final phase, and we are going to implement our models into the production system. But, what if one of the previous phases doesn't go well? This is when we understand that the CRISP-DM model is an iterative one. If the previous evaluation phase terminates showing that an unsatisfactory level of performance was reached, it would be pointless to develop a deployment plan, since the deployed solution would not meet the business expectations, and this would later produce undesired costs required to fix the problem.
In these circumstances, it would definitely be more appropriate to invest some more time to understand what went wrong, to define which phase of the CRISP-DM process needs to be resorted to. A model performance analysis could, for instance, reveals a poor level of accuracy of the model due to bad data quality of the training dataset employed to estimate model parameters. This would then involve a step back to the data preparation phase, or even to...
