TeamTraditionalWorkflow

CRISP-DM for Data Science Teams: 5 Actions to Consider

While there is no standard process for a team to use when working on a data science project, CRISP-DM (CRoss-Industry Standard Process for Data Mining) is one framework that is often considered for data science projects. Perhaps because of this, there are lots of web sites describing the 6 phases of a CRISP-DM project, and noting that CRISP-DM can be used for a data science project (even though CRISP-DM was developed 20+ years ago, prior to the rise of data science).

However, there are fewer discussions / blog posts on the challenges of using CRISP-DM for data science projects (or when leading a data science team), or how to augment the key concepts of CRISP-DM within a data science context.

So, after briefly reviewing CRISP-DM (phases and deliverables), I’ll highlight 5 key actions teams should think about if they are using CRISP-DM.

 

CRISP-DM Phases

As noted in our overall description of CRISP-DM, CRISP-DM defined six phases of project:

  1. Business Understanding: determine business objectives; assess situation; determine data mining goals; produce project plan
  2. Data Understanding: collect initial data; describe data; explore data; verify data quality
  3. Data Preparation (generally, the most time-consuming phase): select data; clean data; construct data; integrate data; format data
  4. Modeling: select modeling technique; generate test design; build model; assess model
  5. Evaluation: evaluate results; review process; determine next steps
  6. Deployment: plan deployment; plan monitoring and maintenance; produce final report; review project

CRISP DM
CRISP DM Life Cycle (Image: Wikimedia Commons)

As one can see, the CRISP-DM process outlines the steps involved in performing a data science analysis from business need to deployment. In addition, as shown in the CRISP-DM diagram, it is an iterative process (in that the project “loops back” when needed).

CRISP-DM Tasks (in each phase)

While many people are familiar with the phases defined by CRISP-DM, not as many are aware that CRISP-DM also defines the tasks that need to be done in each phase as well as the expected outcomes from these tasks (see the chart below).

https://media-exp1.licdn.com/dms/image/C4D12AQEI5EGDrmLhng/article-inline_image-shrink_1000_1488/0?e=1596067200&v=beta&t=vVWCyV-3JM1RrFALfCxwqIje3hmJb5pTp7t30lf8wyc

CRISP-DM tasks in bold, and outcomes in italic (table from CRISP-DM Guide)

Key Actions to Consider to improve CRISP-DM Data Science Projects

With this in mind, when using CRISP-DM for data science projects, teams should consider the following items to help ensure a successful project:

1. Define a team process

  • The team process should define how the team communicates, prioritizes tasks and “loops back” to previous project phases.
  • Some argue that CRISP-DM suffers from the same weaknesses of waterfall frameworks, and encumbers rapid iteration (since CRISP-DM has a heavy on documentation; for example, CRISP-DM calls for 12 reports prior to data collection). This is a good example of why a data science project should not be thought of as a waterfall effort.
  • Some teams ignore the document aspect of CRISP-DM, but if this is done, there needs to be other mechanisms for the team to communicate and prioritize work. Hence, perhaps the largest drawback when using CRISP-DM is that it doesn’t really focus on how should the team collaborate, communicate, and effectively prioritize potential tasks? Teams can leverage the CRISP-DM phases, and then use a framework such as Scrum, Kanban or Data Driven Scrum to prioritize potential tasks.

2. Ensure multiple experiments 

  • When using the CRISP-DM framework, it should also be clear how the team will repeat / iterate through the life cycle multiple times within a project. For example, after exploring some data and deploying an analysis, additional data might be added to the analysis.
  • While it’s important to do multiple iterations, each team needs to determine how this would actually work for their project. So, the team needs to think through how iterations are defined and then evaluated (such as Scrum’s time boxed sprints, Data Driven Scrum’s functional iterations or Kanban’s focus on minimizing flow but not specific iterations).  For example, here is one way to merge CRSIP-DM and Scrum.

3. Define team roles  

  • CRISP-DM does not include the concept of roles (nor a team). Roles can include stakeholders / product owners (to ensure the insight is actionable). See 10 reasons why you might need a product owner / product manager for some reasons why this is so important.
  • In addition, there might need to be a team member more focused on process (in other words, a process master) or on an aspect of the analysis (such as data engineering). See the discussion on 8 roles in a data science team for more on the need for roles in a data science project.

4. Ensure actionable insight 

  • How does the team ensure actionable insight? If the team does not focus in the stakeholders / product owners (which is possible within the CRISP-DM construct), there is the risk that there will be interesting insights, but that those insights will not be actionable.

5. Define the subitems within each phase

  • Finally, the six phases of a project are fairly high level. There are many potential subtasks within each phase of the project. The team needs to be clear how those tasks (within a phase) are defined, and how to ensure tasks will not be forgotten. Many have defined lists of tasks for each phase. Others have addressed these concerns by also defining additional high level phases (e.g., Microsoft’s TDSP)
  • For example, some tasks that should be explicitly discussed include:
    • How/where does the team talk about potential bias (in the data or the machine learning algorithm) or if model transparency is required
    • How should missing data been handled
    • How to determine the appropriate level of accuracy for the business problem (sometimes, data scientists love to improve a model by 0.5% – without knowing the business value of that improvement)
    • How does the team know the results are valid (i.e., there was no error during the analysis)
    • If the model will be deployed in a production environment, how often will it need to be updated 
  • If you’re curious what might be appropriate to think about, take a look at this project, or this description of typical activities within each phase.

Potential Next Steps:

   To explore CRISP-DM in more depth, you can: