The Need
I have an energy client who needed to make predictions based on customer usage patterns. Given the number of categorical fields—and based on feedback from my Machine Learning Model Picker’s flowchart for prediction tasks—I decided to go with the XGBoost model.
My Test Drive
While I was waiting for access to their data, I decided to take a test drive with the Titanic dataset—to see how well it did predicting survival rates.
Data Exploration Dashboard
I started with an exploration into the dataset and created a dashboard around some of the key dimensions and metrics, including some categories I created, that I thought might aid in predicting survival rates. I used Plotly for the charts and stored them in a Flask web app (which you can take for a test drive here).
Model Performance Dashboard
The Model Performance Dashboard transforms complex AI metrics into an interactive, educational experience that helps non-technical stakeholders understand exactly how a predictive model works and why it makes certain decisions. Through strategic use of hover tooltips, visual formatting, and plain-English explanations, it bridges the communication gap that contributes to lost opportunities with AI pilots, turning confusion matrices and ROC curves from intimidating black boxes into tools for meaningful business conversations.
Model Info Dashboard
The Model Info Dashboard provides a transparent view into the XGBoost model’s configuration, displaying key parameters like the number of decision trees (100), tree depth (3), and learning rate (0.1) that power the Titanic survival predictions. This technical transparency helps stakeholders understand not just what the model predicts, but how it was built and trained, demystifying the AI process.
Analysis
For a full analysis of the dataset, check out the blog post.