First conceived of in late 1996 and later developed by a consortium consisting of Daimler-Benz, SPSS, and NCR, CRoss Industry Standard Process for Data Mining (CRISP-DM) 1.0 was released in 2000. CRISP-DM is currently one of the most popular and effective frameworks for advanced analytics and data science (data mining) projects. It is especially well suited for organizations that lack experience in these areas.
The CRISP-DM approach defines a business problem and gets an understanding of the available data. The data is then prepared and analytically modeled until a result is achieved that can be evaluated and deployed. The approach, which is iterative and repeatable by design, consists of six phases:
- Business understanding: Determine business objectives, assess the situation, determine your analytic goals, and produce a project plan.
- Data understanding: Collect, describe, explore, and verify the quality of the data available.
- Data preparation: Select, clean, construct, integrate, and format the required data.
- Modeling: Select modeling technique(s), generate a test design, build the model, and assess the model.
- Evaluation: Assess results and the process in order to determine next steps.
- Deployment: Plan the deployment, along with monitoring and maintenance processes, finalize the processes, and deploy.
When it comes to Digital Decisioning and decision modeling, there are three aspects of CRISP-DM that play key roles: business understanding, IT engagement, and deploying and delivering business value. Let’s delve into each one.
Before the analytical work begins, it’s critically important to have a firm grasp on the business objective and scenario. The numerous deliverables generated during this first phase define these. Deliverables can include background information, a glossary, risks and contingencies, business objectives and success criteria, requirements and assumptions, an inventory of resources, a cost-benefit analysis, data-mining success criteria and goals, and more.
Project teams using CRISP-DM need to work toward capturing the requirements, assumptions, constraints, and available resources to make the data-mining success criteria clear and to match those to the business objective. The ultimate goal is to develop models that assist with decision-making. Decision modeling involves describing each decision in terms of the specific question that need to be answered and linking it to previous decisions. It also requires input data and knowledge sources like policies, regulations, and best practices. The model is predicated on three things: the business objectives impacted by the decisions, the organizations or internal groups making the decisions, and the processes used in decision-making.
In the case of CRISP-DM, decision models generally include analytics or data science outputs—such as regression models, neural networks, and decision trees. Building decision models in the business-understanding phase helps with the following:
- Clarifying the business problem that analytics aim to solve.
- Focusing on business decision-making.
- Showing what role the analytics play in decision-making.
- Connecting analytics to business objectives and processes.
Decision models created during this phase offer multiple benefits that ripple out in the subsequent phases of the project.
Business and IT Engagement
n effective decision model requires engagement of both business and IT professionals.
As we mentioned earlier, analytic model development is an iterative process. Data scientists are constantly looking at the data in new ways. They try different algorithms. They apply multiple analytical techniques. And they combine data in a variety of ways. All of these technical tasks are carried out by the analytic team throughout other phases of a CRISP-DM project.
There is, however, a danger in this highly siloed approach. The business and IT teams are often left out of the loop. A decision model helps prevent this because the analytics always keeps the decision-making process top of mind with each iteration. With a decision model, the analytic team often needs to engage with the business team. For example, the analytic team may discover that it cannot build the analytical model that was originally envisioned and may find that the analytics can be applied to predict or describe something else entirely. They then collaborate with the business and IT teams—every step of the way—to see if changes can be made to take advantage of the analytical model they were able to come up with.
There’s another potential danger as well—what we call the “shiny object” syndrome. The plethora of advanced analytic tools available today may prove distracting and take the focus off the business problem that needs to be solved. Both business and analytic professionals frequently want to try out the coolest, newest approaches, but they may not always be the right tools for the project.
A decision model allays this concern because it has clearly defined guidelines for the project. For any new technology that the project team may be interested in, they are always compelled to ask themselves: Will this new technology or data help us achieve our business objectives? If the answer is “yes,” the decision model can justify adoption of new tools. If the answer is “no,” the tools likely should not be used.
Deploying and Delivering Business Value
A key advantage of CRISP-DM
is the fact that it includes evaluation and deployment. You embed analytics
into operational workflows and take action based on those analytics.
Additionally, in a CRISP-DM project, analytical success criteria is attuned to business success criteria. An analytic model is valuable only if it truly adds business value. In this highly iterative process, analytic teams try multiple modeling approaches, refining or even rejecting some as they go along. There is always the danger that the analytic team can become over-focused on accuracy and lose sight of the business objective. While developing a more accurate model is a lofty goal, the analytic team may end up consuming valuable time and resources in the process. They may also find that the data may not support a very accurate model and then apply themselves to extracting as much accuracy as possible.
This is where the decision model comes in. A few scenarios may play out. The analytic team may realize that their model is accurate enough, given the business parameters that are defined. Alternatively, the analytic team may need to go back to the business understanding phase and work with the business team to determine whether the business case itself makes sense or needs to be revised.
Once a model is complete, it needs to be deployed so that the organization can extract value from it. This is where the IT team’s involvement is required. They are responsible for embedding the analytic model and the algorithm developed into a production environment. This involves coordinating business rules, analytics, and artificial intelligence (AI) components into a decision service that supports different application contexts.
After deployment, the business and operations teams need to make sure that behavior changes are actualized in order to leverage the new analytic approach.
Developing a decision model at the outset of project has a number of advantages. Decision models help by:
- Striking a balance between the analytic model and business rules resulting from policies or regulations that must be enforced when the decision is made.
- Delineating business requirements by showing how business rules and analytics can work together to result in data-driven, compliant decisions.
- Showing what data is used where in the decision.
- Serving as a training and implementation framework for everyone involved and providing a roadmap for organizational change.
- Assisting with performance management and decision monitoring by tracking how the decision was made and then linking this decision-making data to the business outcomes.
In summary, decision models and CRISP-DM will help ensure you get business value from your analytic investments.