CRISP-DM provides a useful structure for analytics writeups, whether those writeups be short (e.g., a blog, medium, or Kaggle post) or long (e.g., an academic paper).

Study the circular graphic on the wikipedia page for CRISP-DM. Note its stages. These stages become report sections. While the process is circular, the report describes a snapshot in time, so the report can flow somewhat linearly, without looping back.

If the report is written as a jupyter notebook, then outputs etc can easily be integrated into the report, interspersed with markdown.

1. Business Undestanding

Describes the general business problem that the analyst, and that the model, set out to address.

2. Data Understanding

Describes the source dataset.

• level of analysis of raw data
• meaning of each column
• descriptives, such as datatypes, number of unique values, modes, averages, etc.
• optionally, descriptive-analytics-type plots of features from the raw data.

3. Data Preparation

Describes all work done by the analyst to the data before the modeling step.

• all cleaning steps taken
• aggregation – level of analysis?
• engineered features – justification, description sufficiently detailed that someone else could replicate. Repeat of descriptives as for raw features

4. Modeling

Describes the selection of modeling algorithms. This section does not evaluate the performance of the models.

Describes all modeling algorithm settings, including hyperparameterization, if any.

Describes training approach – cross-validation? Splitting strategy?

5. Evaluation

Evaluates the predictive performance of each of the models described in the previous step.

• f1-score
• accuracy

Include plots! For each type of figure, plot all models on one figure.

5.x Select a model to recommend for deployment

Explore Model Interpretation – for the selected model, which features were the most predictive? Partial dependencies, etc.

6. Deployment

Make it possible for someone else to use the model without having to train it themselves.

Options include:

• putting a pickled pipeline on github, and add a Usage README at a minimum.
• uploading it as an ML endpoint to a service such as AWS Sagemaker