With the advancement of packages such as tensorflow, sklearn and paid softwares or services such as sage maker, building, training and validating a machine learning algorithm has been extremely simple. The decisions of what goes as a final model is not solely dependent on the results metrics but largely on the usability of the model. A model that is efficient in the production system with f1 score 0.8 will definitely be preferred over an overly sized model of f1 score 0.82
Furthermore, it’s not just models that are useful, but the entire pipeline. This is because for giant organizations the data distribution is dynamic, implying changing clusters and trends. Hence that requires model training and selection in the real time.
Overall, the challenge for the day is to bring the machine learning product into the light of the day. This will require a researcher to have a strong understanding of the MLOps as well.