Analysis overview and configuration
| Parameter | Value | _row |
|---|---|---|
| confidence_level | 0.95 | confidence_level |
| test_size | 0.3 | test_size |
| classification_threshold | 0.5 | classification_threshold |
| positive_class | completed | positive_class |
| laplace | 1 | laplace |
This analysis applies Naive Bayes classification to identify which student characteristics predict test preparation completion. The model was trained on 701 observations and tested on 299 to evaluate its ability to distinguish between students who completed versus did not complete test preparation, using three numeric predictors.
The Naive Bayes model demonstrates moderate predictive capability but with notable limitations. The model is more conservative in predicting completion (
Data preprocessing and column mapping
| Metric | Value |
|---|---|
| Initial Rows | 1,000 |
| Final Rows | 1,000 |
| Rows Removed | 0 |
| Retention Rate | 100% |
This section documents the data preprocessing pipeline for a binary classification model predicting task completion. It shows that no data loss occurred during cleaning, which is critical for understanding whether the subsequent model performance (AUC=0.673, Accuracy=0.629) reflects true predictive capability or is constrained by data quality issues.
The perfect retention rate suggests the dataset arrived clean and required no corrective preprocessing. However, this raises a subtle concern: the model's moderate performance (Kappa=0.226, Precision=0.484) may not stem from data quality issues but rather from weak predictive signal in the three numeric predictors themselves. The class imbalance (64.2% "none" vs. 35.8% "completed") was preserved unchanged, which appropriately reflects real-world distribution but may explain the lower sensitivity (0.579) for the minority class.
The train/test split details are not
| Finding | Value | Assessment |
|---|---|---|
| AUC | 0.673 | Acceptable |
| Accuracy | 62.9% | Moderate |
| Sensitivity | 57.9% | Moderate |
| Specificity | 65.6% | Moderate |
| Precision (PPV) | 48.4% | Moderate |
| F1 Score | 0.528 | Moderate |
| Kappa | 0.226 | Moderate |
This section synthesizes the Naive Bayes classification model's performance across all evaluation metrics to assess whether the model achieves acceptable predictive capability for distinguishing "completed" from "none" outcomes. Understanding these results is critical for determining whether the model is ready for operational deployment or requires further refinement.
The model demonstrates acceptable but limited predictive utility. With an AUC of 0.673
ROC curve showing Naive Bayes model discrimination ability
The ROC curve visualizes the Naive Bayes model's ability to discriminate between "completed" and "none" classes across all possible classification thresholds. This section evaluates whether the model performs meaningfully better than random chance (AUC = 0.5) in ranking positive and negative instances, which is essential for understanding overall predictive reliability.
The AUC of 0.673 indicates the model has acceptable but limited discrimination ability. This aligns with the overall accuracy (62.9%) and moderate kappa (0.226), suggesting the three numeric predictors provide meaningful but incomplete separation between classes. The optimal threshold of
Confusion matrix heatmap showing prediction accuracy by class
The confusion matrix quantifies how the Naive Bayes classifier performed on the test set (n=299), breaking down predictions into four categories: correct and incorrect classifications for each class. This section is essential for understanding not just overall accuracy, but the specific types and frequencies of errors the model makes, which reveals whether misclassifications are balanced or skewed toward one class.
The model demonstrates asymmetric error patterns. While it correctly identifies negative cases (TN=126), it struggles with
Distribution of predicted probabilities by actual class
This section visualizes how well the Naive Bayes model separates predicted probabilities between the two classes. Good discrimination occurs when "completed" observations cluster at high probabilities and "none" observations cluster at low probabilities. This distribution directly reflects the model's ability to rank-order cases by likelihood of completion.
The modest separation between distributions indicates the three numeric predictors provide moderate discriminative power. The 57.9% sensitivity-specificity balance (near 60-65% range) suggests the model performs slightly better at identifying non-completions than completions. This aligns with the overall AUC of 0
Conditional feature means by class showing predictive power of each feature
Feature profiles reveal the conditional distributions (mean and standard deviation) of each predictor within each class, representing the core parameters learned by the Naive Bayes classifier. Features exhibiting larger mean differences between "completed" and "none" classes possess greater discriminative power for predicting task completion. This section directly illuminates which predictors drive the model's 0.673 AUC performance.
All three numeric predictors show consistent upward shifts in the "completed" class relative to "none," confirming they contribute positively to classification. The modest mean separations (5–10 points) align with the model's moderate performance metrics (accuracy 0.629, F
Comprehensive classification performance metrics
| metric | value | interpretation |
|---|---|---|
| AUC | 0.6731 | Acceptable |
| Accuracy | 0.6288 | Moderate |
| Sensitivity (Recall) | 0.5794 | True positive rate |
| Specificity | 0.6562 | True negative rate |
| Precision (PPV) | 0.4844 | Positive predictive value |
| NPV | 0.7368 | Negative predictive value |
| F1 Score | 0.5277 | Harmonic mean of precision & recall |
| Kappa | 0.2259 | Moderate agreement |
| Optimal Threshold | 0.4538 | Youden J optimal cutoff |
This section evaluates the Naive Bayes classifier's overall discriminative ability and predictive reliability on the test set (n=299). These metrics collectively assess whether the model generalizes beyond training data and whether its predictions are trustworthy for the "completed" vs. "none" classification task.
The model demonstrates acceptable but suboptimal performance. While AUC suggests the model can distinguish between classes across thresholds, the low F1 score and Kappa reveal that predictions lack reliability, particularly for identifying "completed" cases (precision only 48%). The imbalanced dataset (64% negative class) inflates accuracy; the model performs better at