Completed AutoML run example

This public example gives you a quick look at a finished Titanic survival classification run in Xalec AutoML Services. It shows the kind of detail a run should preserve: dataset shape, readiness checks, model comparison, selected-model metrics, explainability, model-review context, report excerpts, diagnostics, and saved outputs.

Job job_1780394224_fea89f17970b163f 1,309 rows 16 models trained 21 saved plots SHAP completed

View AutoML Discuss fit Open manifest

What this shows

A completed run should leave evidence behind.

This Titanic example shows the kind of trail Xalec AutoML Services should preserve after a run: the data that went in, the checks that happened before training, the model comparison, the selected-model evidence, the review context, and the reports or files someone can inspect later.

For analysts

See how the run moves from dataset preparation to EDA, model comparison, diagnostics, SHAP, and generated reports.

For reviewers

Review the selected model, selection metrics, class-balance caution, saved plots, and model-review context before using the output.

For technical readers

Open the manifest when you want the machine-readable evidence excerpt behind this completed run.

Open manifest →

Data Preparedness

The run starts with a defined dataset, target, and predictor set.

This page follows the order a user would naturally check: what data went in, what target was modeled, what predictors were used, and what outputs were saved.

Data split

The prepared dataset had 1,309 rows, split into 1,047 training records and 262 test records.

Target

The modeled outcome was passenger survival, using the saved target column survived.

Predictors

The raw predictors were pclass, sex, age, sibsp, parch, and fare, expanded into 12 encoded features for modeling.

Prepared outputs

The run preserved the machine-readable manifest, reports, plots, model files, preprocessing files, prediction outputs, diagnostics, logs, and metadata.

EDA

The platform checked the data before model training.

Before training, the platform checks whether the data is ready to model. The EDA step confirms the target, requested predictors, plot generation, missingness, class balance, and basic feature quality before the run moves into model comparison.

Readiness status

EDA passed. The report found the target column, requested features, numeric and categorical fields, plots, and statistics.

Target distribution

809 passengers did not survive and 500 survived, giving a 61.8% / 38.2% split.

Missingness

Overall missing fraction was 2.88%, visible in the saved missing-values diagnostic.

Feature quality

The checked fields were pclass, sex, age, sibsp, parch, and fare. No high-cardinality or near-constant feature warning was recorded; duplicate rows, missingness, and outlier signals were preserved for review.

Checks before training

The readiness profile kept the expected checks visible: stratified holdout, class-balance review, expected metrics, and expected plots such as target distribution, feature distributions, correlation heatmap, confusion matrix, ROC curve, and precision-recall curve.

Recommended models

The EDA stage recommended CatBoost, LightGBM, and RandomForest as strong starting points for this classification workflow.

Readiness caution

The run kept a class-imbalance caution visible: inspect minority-class recall and avoid treating headline performance as production-ready without review.

Service diagnostics

The diagnostic bundle was marked complete for the classification workflow, with 21 saved plots available across EDA, model diagnostics, threshold review, and SHAP explainability.

Target distribution chart for the completed AutoML example run — Target distribution used for class-balance review.

Missing values chart for the completed AutoML example run — Missing-values diagnostic from the EDA stage.

Correlation matrix from the completed AutoML example run — Correlation matrix generated before model training.

Age distribution from the completed AutoML example run — Age distribution used during EDA review.

Sex distribution from the completed AutoML example run — Sex distribution checked before modeling.

Fare distribution from the completed AutoML example run — Fare distribution preserved for feature review.

Rare category diagnostic from the completed AutoML example run — Rare-category diagnostic for categorical review.

Model Building & Selection

Sixteen candidate models trained successfully.

CatBoost was chosen using F1, ROC AUC, and PR AUC. Accuracy is still reported below, but it is supporting context rather than the selection rule.

85.88%Accuracy, CI 81.14% to 89.58%

80.83%F1 score, CI 75.13% to 86.43%

90.39%ROC AUC, CI 85.53% to 93.73%

87.24%PR AUC, CI 81.57% to 92.12%

Model	Accuracy	F1	ROC AUC	PR AUC
CatBoost	85.88%	80.83%	90.39%	87.24%
XGBoost	84.73%	79.17%	89.96%	84.62%
Keras CNN	83.59%	77.95%	87.01%	81.98%
Gradient Boosting	83.59%	77.72%	89.22%	85.46%
LightGBM	83.21%	77.08%	88.57%	84.84%

Model comparison chart for the completed AutoML example run — Model comparison plot generated by the completed run.

Training-time comparison chart for the completed AutoML example run — Training-time comparison is kept as supporting context, not as the model-selection rule.

Best Model Performance Plots

The selected model includes diagnostics and explainability plots.

The performance plots show how the selected CatBoost model behaved on the held-out test set. The threshold plot shows the tradeoff between sensitivity, specificity, precision, recall, and related metrics as the classification cutoff changes.

Confusion matrix for the completed AutoML example run — Confusion matrix for the selected model.

ROC curve for the completed AutoML example run — ROC curve for discrimination review.

Precision-recall curve for the completed AutoML example run — Precision-recall curve for class-balance review.

Threshold performance chart for the completed AutoML example run — Threshold performance plot; the saved cutoff is 0.4412 by Youden's J for the Survived class.

Feature importance chart for the completed AutoML example run — Feature importance from the selected model.

SHAP summary bar chart for the completed AutoML example run — SHAP summary bar plot for global review.

SHAP beeswarm chart for the completed AutoML example run — SHAP beeswarm plot for distributional review.

Signed SHAP effects chart for the completed AutoML example run — Signed SHAP effects for reviewing direction and strength.

Sex

39.96% feature-importance share in the saved global ranking.

Fare

21.43% feature-importance share in the saved global ranking.

Passenger class

16.01% feature-importance share after grouping class indicators.

Age

14.21% feature-importance share in the saved global ranking.

Model Review

Before reuse, someone can review the fitted model.

The example shows the run results. In the live AutoML platform, a completed job also has a model review panel so a reviewer can inspect the fitted model, leave notes, and record the decision.

What the review checks

The review panel checks job state, model result rows, the selected model, selection metrics, dataset readiness, saved files, and billing evidence.

Review states

Runs can be left as draft, moved into review, approved, marked as needing changes, or rejected. That keeps approval separate from raw training success.

Reviewer notes

Notes and history stay with the job, so later users can see why a fitted model was accepted, held back, or sent for changes.

What this example does not claim

No approval decision is invented for this public excerpt. The page shows the saved details that a reviewer would use.

AI Reports

The run also produced two written reports.

This completed AutoML run compared 16 classification models on a Titanic survival dataset with 1,047 training records and 262 test records. CatBoost was selected on the strength of its F1, ROC AUC, and PR AUC evidence: 80.83% F1, 90.39% ROC AUC, and 87.24% PR AUC. Its 85.88% accuracy is reported as supporting context.

SHAP explainability completed successfully. The strongest descriptive drivers were sex, fare, passenger class, age, and parch. These signals help reviewers understand model behavior, but they should not be read as causal claims.

What the manuscript-style report adds

Methods context for the retrospective supervised classification setup.
The full model-comparison table with confidence intervals and training times.
An explainability table ranking sex, fare, passenger class, age, parch, and lower-ranked encoded features.
Limitations, future-work notes, references, and a methods appendix drawn from saved run outputs.

Report Outputs

Two formats for two audiences.

Executive brief

Short summary for the run outcome, selected model, and practical interpretation.

Journal manuscript

Manuscript-style report with methods, model comparison, confidence intervals, explainability tables, limitations, references, and figures.

Methods appendix

Compact reproducibility notes covering the job id, service type, dataset fingerprint, selected models, split plan, and explainability status.

Report Excerpt

Short excerpts from the generated reports.

The full report package is long, so this section keeps only the parts a reviewer is most likely to scan first.

Executive brief

In short: the best model was CatBoost; the target outcome was survival; the leading features were sex, fare, passenger class, age, and parch.

This is the short summary a manager or reviewer would scan first.

Manuscript results

In short: CatBoost was the strongest observed model in the saved package, with 0.808 F1, 0.904 ROC AUC, and 0.872 PR AUC.

Use this for technical review, not as a causal claim.

Methods appendix

In short: the job used the Titanic CSV, a balanced training preset, 16 selected models, a 0.2 test split, and SHAP explainability.

The dataset fingerprint starts with c1b78e7b57a3.

What stays out

The full manuscript includes availability statements, references, figures, and long tables. Those are better kept in the AutoML workspace or manifest package, where reviewers can open them when needed.

Open manifest →

Additional Outputs

The page shows highlights; the manifest lists the full package.

The updated job includes more than the visible charts on this page. The manifest keeps those outputs grouped so a reviewer can see what exists without turning this page into a file browser.

Report package

Executive brief, manuscript HTML/Markdown, methods appendix, references, manuscript input, prompt package, metadata, and report JSON.

Diagnostic package

Service diagnostics report 21 saved plots and a complete classification diagnostic bundle for the expected plot groups, including threshold performance.

Reproducibility package

The reproducibility manifest records the dataset fingerprint, selected models, training preset, package versions, and output inventory.

Model package

The private workspace keeps model binaries, prediction outputs, encoders, preprocessing files, optimal cutoffs, and logs. The example names those artifacts without exposing private files.

What is included

A public example, not the full workspace.

The run generated reports, prediction outputs, plots, SHAP files, model files, preprocessing files, diagnostics, logs, and metadata. This page shows the summary; the complete workspace will live inside the AutoML platform.