The Machine Learning Process refers to the structured approach used to apply machine learning to real-world problems. One of the most widely accepted process models for this is the CRISP-DM model (Cross Industry Standard Process for Data Mining).
Since Machine Learning is similar to Data Mining (except for the goal), CRISP-DM can be effectively applied to ML workflows as well.
CRISP-DM Model – 6 Phases of the Machine Learning Process
1. Business Understanding
- The first step is to understand the goals and objectives of the business.
- It includes:
- Identifying the problem clearly.
- Understanding business needs.
- Framing it as a machine learning problem.
- Usually, one algorithm is enough to begin solving the problem.
Example: A retail company wants to predict customer churn — this is defined as a classification problem.
2. Data Understanding
- In this step, data is collected and its characteristics are analyzed.
- It includes:
- Understanding the structure, types, and patterns in the data.
- Formulating hypotheses and verifying them using statistical tools.
- Helps identify whether the available data is suitable for modeling.
3. Data Preparation
- Raw data is cleaned and transformed into a usable format.
- It involves:
- Handling missing values, duplicate records, and incorrect data.
- Feature selection, encoding, and formatting data for training/testing.
- Proper data preparation is critical for the model’s success.
Missing values can severely impact classification accuracy and need special handling strategies.
4. Modeling
- At this stage, a suitable machine learning algorithm is applied.
- Focus is on:
- Training the model with the prepared data.
- Selecting proper hyperparameters and tuning them.
- Output is a trained model or pattern.
Example: Using a Decision Tree classifier on customer churn dataset.
5. Evaluation
- The model’s performance is assessed using:
- Accuracy, precision, recall, F1-score, etc.
- Visualization tools and domain knowledge.
- Evaluation determines whether the model solves the business problem accurately.
- If performance is poor, go back to data preparation or modeling.
Example: If an email classifier incorrectly flags many good emails as spam, model refinement is needed.
6. Deployment
- Final stage where the model is implemented in the real-world system.
- It can be used to:
- Make predictions
- Improve existing workflows
- Trigger automated decisions
Example: Deploying a fraud detection model in an online payment system.