This interface translates your churn-modelling outputs into a presentation layer suitable for hosting. It focuses on model comparison, top risk-driver patterns, imbalance strategy results, and feature-selection trade-offs.
This dashboard is designed as a dissertation-friendly implementation layer. It does not retrain models on the server. Instead, it presents the outputs from your analysis notebook in a way that is easier for supervisors, markers, and future stakeholders to review.
PHP is a practical choice for hosting because it can render CSV-driven result pages on standard shared hosting. That makes it easy to deploy without requiring a Python application server in production.
| Model | ROC-AUC | PR-AUC | Precision | Recall | F1 | Brier | Precision@10% | Lift@10% |
|---|---|---|---|---|---|---|---|---|
| XGBoost | 0.694 | 0.169 | 0.215 | 0.278 | 0.242 | 0.065 | 0.215 | 2.922 |
| Logistic regression | 0.668 | 0.143 | 0.176 | 0.258 | 0.209 | 0.066 | 0.177 | 2.414 |
| Random forest | 0.653 | 0.137 | 0.146 | 0.370 | 0.209 | 0.067 | 0.153 | 2.087 |
| Dummy baseline | 0.500 | 0.073 | 0.073 | 1.000 | 0.137 | 0.068 | 0.064 | 0.871 |
Why this matters: PR-AUC and Precision@10% are especially valuable for churn because the class is rare and outreach is usually targeted at the highest-risk segment rather than the full customer base.
Why this matters: the top features help you explain which anonymised variables carry the strongest signal. In the hosted dashboard, they can be shown as evidence of model behaviour without over-claiming business meaning.
| Strategy | PR-AUC | Recall | Precision | F1 | Brier |
|---|---|---|---|---|---|
| Class weight balanced | 0.093 | 0.515 | 0.094 | 0.160 | 0.186 |
| SMOTENC | 0.092 | 0.730 | 0.087 | 0.155 | 0.176 |
| No rebalancing | 0.088 | 0.650 | 0.089 | 0.156 | 0.068 |
Why this matters: imbalance handling changes the recall–precision trade-off. The dashboard should keep these metrics visible together so future users do not optimise only one number and ignore decision costs.
| Model setup | ROC-AUC | PR-AUC | Precision | Recall | F1 | Precision@10% | Lift@10% |
|---|---|---|---|---|---|---|---|
| XGBoost - full screened features | 0.703 | 0.178 | 0.164 | 0.417 | 0.236 | 0.219 | 2.976 |
| XGBoost - Top 40 selected features | 0.694 | 0.169 | 0.215 | 0.278 | 0.242 | 0.215 | 2.922 |
Why this matters: this section helps justify whether the production dashboard should use a full screened feature set or a leaner subset. Here, the full screened feature configuration has the edge on PR-AUC.
XGBoost is the strongest dashboard candidate because it leads on PR-AUC (0.169) while also delivering the highest Lift@10% (2.922), which means the riskiest decile is enriched with churners relative to the base rate.
XGBoost - full screened features performs best in the feature-selection experiment. This suggests that keeping a broader screened feature set preserves more useful signal than the compact Top 40 subset alone, although the smaller subset remains attractive for a leaner dashboard view.
Class weight balanced achieves the best PR-AUC in the imbalance comparison, but the dashboard should still surface recall and precision together because higher-recall strategies may be desirable when the business wants to catch more potential churners.
The Orange variables are anonymous. The dashboard therefore shows feature identifiers such as Var217 and Var202 rather than invented business labels, which keeps the interface honest and methodologically defensible.
orange_php_dashboard folder to your hosting account.public_html or htdocs.your-domain.com/orange_php_dashboard/.data folder whenever you export fresh results from your notebook.