AutoML is a Domino extension that enables data scientists and domain experts to build, evaluate, and deploy machine learning models through a streamlined, no-code interface.Powered by AutoGluon, AutoML automates the end-to-end model training pipeline - from data profiling and feature engineering through model selection, hyperparameter tuning, and ensembling - so you can go from raw data to a production-ready model in minutes.AutoML provides two primary workflows:
Data Exploration: Upload a dataset, review column statistics and distributions, examine correlations, assess data quality, and apply transformations before training.
Model Training: Configure and run an AutoGluon training job that automatically trains multiple model types, ranks them on a leaderboard, and produces deployment-ready artifacts.
Exported models and notebooks should only be shared with authorized users.
You can access AutoML from the left navigation sidebar of any Domino project under the Extensions section.
In the left sidebar, scroll down to the Extensions section.
Click AutoML.
The AutoML landing page displays all existing training jobs with their name, type, status, best model, and creation date. From here you can start a new training job or explore your data.
Use the Explore Data button in the top-right corner of the AutoML landing page to profile and transform your data before creating a training job.
The Data Exploration tool lets you analyze data quality, distributions, and correlations, and prepare transformations before training a model.To open it, click Explore Data from the AutoML landing page.
Upload File: Drag and drop (or click to browse) a CSV or Parquet file. The maximum file size is 550 MB.
Domino Dataset: Select a file from a mounted Domino Dataset already connected to your project.
Once a file is loaded, Domino automatically profiles the data and takes you to the Data Exploration interface. The file name, column count, and row count are displayed at the top of the page. You can switch to a different file at any time by clicking Change File.
The Data Preview tab shows a sample of your raw data in table format. You can browse columns and rows to get a quick sense of the dataset’s structure and values. Use the Rows per page control to adjust how many rows are displayed at a time.
The Column Analysis tab provides detailed profiling information for each column in your dataset. Select a column from the list on the left to view its profile on the right.AutoML automatically detects each column’s semantic type, displayed as a colored tag beneath the column name. Detected types include:
Type
Description
Color
numeric
A column containing continuous numeric values (integers or floats).
Blue
monetary
A numeric column identified by name patterns related to financial values (e.g., price, cost, amount, revenue, salary).
Blue
binary
A numeric column with exactly two distinct values, commonly used as a classification target.
Purple
category
A column with a moderate number of distinct text or object values, or identified by name patterns (e.g., type, class, status, group).
Purple
categorical_numeric
A numeric column with fewer than 20 distinct values representing less than 5% of the total rows, treated as a categorical feature.
Purple
datetime
A column containing date or timestamp values, detected by data type or name patterns (e.g., date, time, timestamp, created, updated).
Green
text
A column containing long text strings (average length > 100 characters), or identified by name patterns (e.g., description, comment, note).
Orange
identifier
A column whose values are likely unique row identifiers, detected by name patterns (e.g., id, uuid, guid) or high cardinality. Identifiers are flagged with a warning and should typically be excluded from training.
Gray
boolean
A column containing true/false values, detected by data type or name patterns (e.g., is_, has_, flag).
Gray
unknown
A column whose type could not be confidently determined. Review these columns manually before training.
Gray
Additional types are detected but displayed with default styling:
email: Columns with email-related names
phone: Columns with phone-related names
url: Columns with URL-related names
latitude/longitude: Geographic coordinate columns
percentage: Columns with percentage-related names (e.g., percent, pct, ratio, rate)
name: Columns with name-related patterns (e.g., name, title, label)
count: Columns with count-related names (e.g., count, num, qty, quantity)
For each selected column, the following information is displayed:
Potential Issues: Warnings such as High cardinality – may be an identifier, High missing rate, or Constant or near-constant column.
Unique Values: The count and percentage of distinct values.
Missing: The count and percentage of null or missing entries.
Data Type: The underlying Pandas data type (e.g., int64, object, float64).
Statistics: For numeric columns: min, max, mean, median, standard deviation, skewness, and kurtosis.
Distribution: A histogram for numeric columns, or a bar chart of top values for categorical columns showing the count and percentage for each category.
The Correlations tab displays a correlation matrix heatmap for all numeric columns in the dataset. Correlation values range from −1 (strong negative correlation) to +1 (strong positive correlation), with color coding to highlight the strength and direction of each relationship.Use the Min correlation slider to filter out weak correlations and focus on the strongest relationships. This view is helpful for identifying multicollinearity between features and understanding which variables are most associated with your target column.
The Data Quality tab gives you an at-a-glance assessment of your dataset’s readiness for model training. It includes the following sections:
Missing Values by Column: A horizontal bar chart showing the number and percentage of missing values for each column that has any. Columns are color-coded by severity: green for less than 5% missing, orange for 5–20%, yellow for 20–50%, and red for more than 50%.
Missing Value Pattern: A compact visual representation of where missing values occur across columns, helping you identify whether missingness is concentrated or scattered.
Warnings: Issues that could affect training quality. For example, Small dataset – consider using best_quality preset for better results.
Recommendations: Actionable suggestions to improve model performance. These are tagged by priority (high or medium) and category (Target, Preprocessing). Examples include:
The Transformations tab lets you define preprocessing steps that will be applied to the dataset before model training. AutoML analyzes your data and suggests recommended transformations based on detected issues.Recommended transformationsAutoML may recommend transformations for columns with the following issues:
Identifier columns: Columns where all values are unique and are likely row identifiers (e.g., PumpId). Recommended action: drop the column.
High cardinality: Columns that may be identifiers due to a very large number of unique values (e.g., ManufacturerModel). Recommended action: review and potentially drop.
Missing values: Columns with a significant percentage of null values (e.g., OperatingYears at 19.9% missing). Recommended action: fill missing values.
Extreme outliers: Numeric columns with extreme values detected (e.g., FlowRatePSI with 6.9% outliers). Recommended action: clip or remove outliers.
High missing rate: Columns with a very high percentage of missing data (e.g., WellSector at 77% missing). Recommended action: drop the column.
Click Add next to any recommended transformation to include it. You can also create custom transformations by selecting a column and a transformation type (e.g., Fill Missing Values) from the dropdown controls at the bottom of the page.All selected transformations appear in the Selected Transformations list. Transformations are included in the exported notebook and can be reviewed and modified in code.
At any point during data exploration, you can click the Export Notebook button in the top-right corner to download a Jupyter notebook that contains all of your data profiling results and any transformations you have selected.This notebook can be opened in a Domino Workspace for further analysis or custom preprocessing.
AutoML model training uses AutoGluon to automatically train, tune, and rank multiple machine learning models. The training job wizard guides you through four steps: selecting your data source, choosing a model type, configuring training parameters, and reviewing your settings before launch.To begin, click New training job from the AutoML landing page.
Select the dataset you want to use for training. As with Data Exploration, you can either upload a CSV or Parquet file (up to 550 MB) or select a file from a mounted Domino Dataset.Once your file is loaded, click Continue to proceed.
Controls the trade-off between model quality and training speed. Options include Medium Quality (Faster), Best Quality, and others. Medium Quality (Faster) is the default.
Time Limit (seconds)
Maximum wall-clock time (in seconds) for the entire training run. Default: 3600 (1 hour). AutoGluon will train as many models as possible within this limit.
Evaluation Metric (optional)
The metric used to rank models on the leaderboard. Set to Auto-detect by default, which chooses an appropriate metric based on the problem type (e.g., accuracy for classification, RMSE for regression).
Experiment Name (optional)
An optional Domino Experiment name for tracking this run. If left blank, a name is auto-generated.
Cache data in memory for faster training. Enabled by default.
Tabs available for tabular models only
ModelsSelect or exclude specific model families from the training run.
Setting
Description
Excluded Model Types
Click to exclude model types from training (highlighted in red). Available models: LightGBM, CatBoost, XGBoost, Random Forest, Extra Trees, K-Nearest Neighbors, Linear Regression, Neural Network (PyTorch), Neural Network (FastAI).
Bagging Folds
Number of folds for bagging (2–10). Set to Auto by default.
Stack Levels
Number of stacking levels (0–3). Higher values increase ensemble complexity.
Fraction of data reserved for validation (0.01–0.5). Typically 0.1–0.2. Set to Auto by default.
Inference Time Limit (s/row)
Maximum inference time per row in seconds. Leave blank for no limit.
Calibrate Probabilities
Calibrate predicted probabilities for better reliability. Useful when probability outputs will be used for decision-making.
Refit on Full Data
After training, refit the best models on the full dataset (training + validation) for maximum performance.
Use Bag Holdout
Use a separate holdout for bagged models, which can improve ensemble quality.
Hyperparameter Optimization (HPO)Enable and configure hyperparameter tuning. When enabled, AutoGluon searches over hyperparameter combinations for each model family.
Setting
Description
Enable HPO
Toggle hyperparameter optimization on/off.
HPO Scheduler
Local (single machine) or Ray (distributed across multiple workers).
Search Algorithm
Auto (recommended), Random Search, Bayesian Optimization, or Grid Search.
Number of Trials
More trials = better results but longer training (1–100). Default: 10.
Max Iterations per Trial
Maximum training iterations for each trial. Leave blank for auto.
Grace Period
Minimum iterations before early stopping can occur (ASHA scheduler).
Reduction Factor
Factor by which to reduce the number of trials at each rung. Typically 3.
Per-Model Hyperparameters: Override default hyperparameters for specific model types (LightGBM, XGBoost, CatBoost, Neural Network) using JSON format.ThresholdFor binary classification tasks, optimize the decision threshold.
Setting
Description
Enable Threshold Calibration
Find optimal decision threshold instead of using the default 0.5.
Optimization Metric
Metric to optimize: Balanced Accuracy (default), F1 Score, Precision, Recall, or Matthews Correlation Coefficient.
Thresholds to Try
Number of threshold values to evaluate (10–1000). Default: 100.
ImbalanceConfigure how AutoGluon handles class imbalance.
Setting
Description
Imbalance Strategy
None (default), Oversample (duplicate minority class), Undersample (reduce majority class), SMOTE (synthetic minority oversampling), or Focal Loss (down-weight easy examples).
Sample Weight Column
Name of a column containing sample weights for weighted training.
FoundationOptions for foundation model-based approaches.
Setting
Description
Use Tabular Foundation Models
Include TabPFN and other foundation models in training.
Foundation Model Preset
None (use with other models), Zero-shot (instant predictions without training), or Zero-shot + HPO (optimize foundation model parameters).
Dynamic Stacking
Use dynamic stacking for adaptive ensemble configurations.
Pseudo-Labeling
Enable semi-supervised learning with unlabeled data. Requires specifying an unlabeled data path.
Drop Unique Features
Automatically drop high-cardinality unique features (like IDs).
AdvancedAdditional low-level AutoGluon parameters for experienced users.
Setting
Description
Enable Distillation
Transfer knowledge from the ensemble to a single faster model for deployment.
Distillation Time Limit
Time allocated for distillation (seconds). Leave blank for auto.
Include Only Specific Models
Whitelist specific model types to include (overrides excluded models). Selected models appear in green.
Bagging Sets
Number of complete bagging sets (increases diversity).
Use Refit as Best
Use the refitted model as the final predictor.
Tabs available for time series models only
Time SeriesConfigure time series-specific settings.
Setting
Description
Frequency
Data frequency: Auto-detect, Daily, Weekly, Monthly, Hourly, Minutely, Quarterly, or Yearly.
Target Scaler
Scaling method for target values: Default, Mean Absolute, Standard, Min-Max, or No Scaling.
Use Chronos
Enable Amazon’s Chronos foundation model for time series forecasting.
Chronos Model Size
Model size when Chronos is enabled: Tiny (8M params), Mini (20M), Small (46M), Base (200M), or Large (710M). Larger models are more accurate but require more memory.
Enable Ensemble
Combine multiple time series models into an ensemble. Enabled by default.
Tabs available for multimodal models only
MultimodalConfigure multimodal-specific settings for combined image, text, and tabular data.
Setting
Description
Text Backbone
Pre-trained text model (e.g., google/electra-base-discriminator).
Image Backbone
Pre-trained image model (e.g., swin_base_patch4_window7_224).
Max Text Length
Maximum token length for text inputs (32–2048). Default: 512.
Image Size
Input image size in pixels (32–512). Default: 224.
Batch Size
Training batch size. Leave blank for auto.
Max Epochs
Maximum training epochs. Leave blank for auto.
Learning Rate
Model learning rate. Leave blank for auto.
Fusion Method
How to combine modalities: Late Fusion (combine at prediction) or Early Fusion (combine at feature level).
Review all your selected settings on the summary screen. If everything looks correct, click Start Training to launch the job.You will be taken to the training run’s overview page where you can monitor progress in real time.
Once a training job is launched, its results page provides comprehensive information about the run’s progress and outputs.The results page is organized into several tabs: Overview, Leaderboard, Diagnostics, Metrics, Outputs, and Logs.
Overview
Model Leaderboard
Diagnostics
Metrics
Outputs
Logs
Forecast (Time Series Only)
The Overview tab displays the training job’s metadata and real-time progress. While the job is running, a progress bar shows the estimated completion percentage and a Cancel button is available to stop the run.The metadata table includes:
Field
Description
Run ID
A unique identifier for this training run.
Model Type
The predictor type (e.g., Tabular).
Problem Type
The detected or specified problem type (e.g., Binary).
Target Column
The column being predicted (e.g., Failed).
Preset
The quality preset used (e.g., Medium Quality Faster Train).
Time Limit
The configured time limit in seconds (e.g., 3600s).
Created
The date and time the job was launched.
Duration
Total elapsed time for the training run.
Status
Current status: Running, Completed, Failed, or Cancelled.
The Leaderboard tab ranks all trained models by their validation score. The top-ranked model is the one AutoGluon recommends for deployment.Models are displayed in a table with the following columns:
Column
Description
Rank
Position on the leaderboard (1 = best).
Model
Name of the model or ensemble (e.g., WeightedEnsemble_L2, LightGBM, CatBoost).
Score
Validation score using the selected evaluation metric.
Fit Time
Time taken to train the model.
Pred Time
Time taken to generate predictions on the validation set.
Click the expand arrow next to any model to view additional details:
Performance: Validation score and marginal fit time.
Model Info: Stack level, whether the model can perform inference (Can Infer), and other metadata.
Use the All Models dropdown to filter the leaderboard by specific model families.
The WeightedEnsemble models combine predictions from multiple base models and typically achieve the highest scores. They may have longer prediction times, so consider this trade-off for latency-sensitive applications.
The Diagnostics tab provides visual insights into model behavior. The available sub-tabs depend on the problem type.Feature Importance: A horizontal bar chart ranking features by their contribution to model predictions. Features with the highest importance have the greatest influence on predictions. Use this chart to understand which variables drive your model’s decisions and to identify potential features for engineering or removal.Model Comparison: Side-by-side comparison of all trained models showing model names with color-coded performance bars, validation scores for each model, and a visual comparison helping you understand how different algorithms performed on your data. The best-performing model is highlighted at the top of the comparison.Classification Metrics (binary and multiclass classification only):This view provides comprehensive classification-specific metrics.
Component
Description
Confusion Matrix
A grid showing predicted vs. actual class labels. Cells are color-coded by frequency, helping identify where the model makes correct predictions and where it confuses classes.
Accuracy
Overall proportion of correct predictions.
Precision
Weighted average precision across all classes (proportion of positive predictions that were correct).
Recall
Weighted average recall across all classes (proportion of actual positives that were correctly identified).
F1-Score
Weighted average F1-score (harmonic mean of precision and recall).
ROC Curve
For binary classification only. Plots the true positive rate against the false positive rate at various threshold settings. Includes the AUC (Area Under Curve) score.
Regression Diagnostics (regression only):This view provides regression-specific metrics and visualizations.
Component
Description
R² (R-Squared)
Coefficient of determination indicating the proportion of variance explained by the model. Values closer to 1.0 indicate better fit.
RMSE
Root Mean Square Error: average magnitude of prediction errors, in the same units as the target variable.
MAE
Mean Absolute Error: average absolute difference between predicted and actual values.
MSE
Mean Square Error: average squared difference between predicted and actual values.
Predicted vs Actual Plot
Scatter plot comparing predicted values (Y-axis) against actual values (X-axis). Points close to the diagonal line indicate accurate predictions.
Residuals Plot
Scatter plot showing residuals (prediction errors) against predicted values. Ideally, residuals should be randomly distributed around zero with no visible patterns.
The Metrics tab provides two comparison charts and a summary statistics panel:Validation Scores: A horizontal bar chart comparing every model’s validation score. The best-performing model appears at the top.Training Times: A horizontal bar chart showing how long each model took to train. This helps you evaluate the cost-performance trade-off for each model.Training Statistics: A summary panel displaying key metrics for the overall run.
Metric
Description
Models Trained
Total number of models trained during the run (e.g., 12).
Best Score
The highest validation score achieved across all models (e.g., 0.860).
Total Train Time
Total wall-clock time for all model training (e.g., 25s).
Avg Train Time
Average training time per model (e.g., 2.1s).
The Outputs tab provides two ways to export your trained model:Deployment Package: A complete, self-contained package with everything needed to deploy the model as an API endpoint. You specify an output directory (default: /domino/datasets/local/<project>/<job_name>) and optionally enable Optimize for Inference. Click Create Package to generate the deployment artifacts.The deployment package contains:
File
Description
model/
Trained model artifacts generated by AutoGluon.
inference.py
A Python inference script that loads the model and returns predictions.
requirements.txt
Python dependencies needed to run inference.
Dockerfile
A container configuration file for building a deployment image.
Training Notebook: Click Download Notebook to get a Jupyter notebook with the full training configuration, evaluation code, and deployment instructions. This notebook serves as a reproducible record of the experiment and a starting point for further customization.
The Logs tab displays the real-time console output from the AutoGluon training process.This includes detailed information about each model being trained, hyperparameters used, validation results, and any warnings or errors.Logs are useful for debugging failed runs and understanding the training process in detail.
The Forecast tab is available only for completed Time Series training jobs. It allows you to generate future predictions using your trained model.Job InformationDisplays key configuration details from training:
Target Column: The variable being forecasted.
Time Column: The timestamp/date column used for time ordering.
ID Column: The identifier column for multi-series forecasting (if configured).
Generating a Forecast
Set Prediction Horizon: Enter the number of future time steps to predict (1-365). This defaults to the prediction length specified during training.
Click Generate Forecast: The model produces predictions for the specified number of time steps into the future.
Forecast ResultsAfter generation, results are displayed in a table format showing:
Column
Description
Time Step
The future time period index (1, 2, 3, etc.).
Predicted Value
The forecasted value for each time step.
Confidence Intervals
Lower and upper bounds at various quantiles (if available), indicating prediction uncertainty.
For multi-series forecasting (when an ID column was specified), predictions are grouped by series identifier.
Navigate to the Outputs tab of your completed training run.
Verify the output directory path under Deployment Package.
(Optional) Check Optimize for inference to reduce model size and improve prediction latency.
Click Create Package.
The generated package is saved to the specified directory in your Domino project’s dataset storage. It includes the model artifacts, inference script, requirements file, and Dockerfile.
To register your model in Domino’s Model Registry for versioning, governance, and deployment:
Click the Register button in the top-right corner of the training results page.
In the Deploy to Model Registry dialog, enter a Model Name.
Select the Model Type (e.g., Tabular).
Optionally add a description to document the model’s purpose and training context.
Click Register to save the model to the registry.
Once registered, the model appears in the Models section of your Domino project and can be deployed as an API endpoint, used in batch scoring, or shared across teams.
Review data quality before training. Use the Data Exploration tool to inspect missing values, outliers, and column types. Address high-priority recommendations (especially dropping or imputing columns with more than 30% missing data) before starting a training job.
Choose the right preset. The Medium Quality (Faster) preset is a good starting point for rapid iteration. Once you have identified a promising dataset and target, switch to Best Quality for production models, especially on smaller datasets where the additional training time yields meaningful improvements.
Set an appropriate time limit. A longer time limit allows AutoGluon to train more models and explore more hyperparameter configurations. For initial exploration, 600–1800 seconds is often sufficient. For production runs, consider 3600 seconds or more.
Exclude identifier columns. Columns with all unique values (flagged as identifiers) do not contribute meaningful signal and should be dropped before training. AutoML will flag these in both Column Analysis and Recommended Transformations.
Examine feature importance. After training, review the Feature Importance chart on the Diagnostics tab. If unimportant features dominate, consider removing them and retraining to reduce noise and improve generalization.
Consider the prediction time trade-off. Ensemble models (e.g., WeightedEnsemble_L2) typically achieve the best validation scores but may have higher inference latency. For real-time applications, compare leaderboard scores with prediction times and consider selecting a simpler model that meets your latency requirements. Use the exported notebook for reproducibility. Download the Training Notebook from the Outputs tab to preserve a complete record of your training configuration. This notebook can be re-run in a Domino Workspace and serves as the starting point for any custom modifications.
