CRISP-ML(Q)

Have you ever wondered why CRISP-DM methodology which was one of the best process methodologies for Data Mining project development is not being practiced anymore?

Why has CRISP-ML(Q) emerged??

The major focus was on ‘model development’ in the CRISP-DM methodology. The data pre-processing and the training of the model happen on the given sample data. Model tuning and evaluation process then help in identifying the best model from various experiments and finally, the best model will be saved for serving purposes. This is a typical flow of steps for the ML development life cycle.?

The typical architecture diagram of CRISP-DM:?

CRISP-DM process does not consider the applicability, reusability, and generalization of the model for business needs. Every business eventually generates data, and the application is expected to work on this new data that flows in. CRISP-DM methodology does not consider these aspects and only focuses on developing a solution for the current data.?

To overcome this drawback and to ensure the application is generalized and reusable with reproducibility aspects CRISP-DM had to get revamped with best practices the industry follows for application deployment. Therefore, the emergency of CRISP ML(Q).?

The CRISP-ML(Q) methodology focuses on ensuring the maintenance of the application after the deployment. Continuous training strategies have been inducted into the ML lifecycle steps.?

Right from phase/step-1 of the MDLC (machine learning development life cycle), the process has quality check metrics included to ensure that quality assurance is also maintained.?

The major difference in CRISP-ML(Q) compared with CRISP-DM is the last step which takes care of post-deployment activities for maintenance and support to customers for the consistent and expected experience.

Pipelines are used for model development to ensure the steps are generalized for any relevant new data which flows into the system as well. Many tools are available to capture the Data drift and generate reports to highlight these differences in the data versions.

Any application over a period degrades and this is called Model drift - this can be caused due to various factors like concept drift, and data drift.?

Tools like MLflow are used to track the model performances and easily compare them to ensure the retraining can be triggered to maintain consistency in the ML model.

Check below a high-level architecture design for CRISP-ML(Q) methodology which is designed with MLOps strategies to enable CI/CD pipelines.

What other differences do you see between the CRISP-DM and CRISP-ML(Q) methodologies?

Credit: Thank you to all the original creators. Some content is picked up from online sources.