Skip to main content
You create a training job to train a Machine Learning (ML) model against a feature-engineered table. You can then use the trained model to make predictions on new data, while NexusOne tracks the training results in MLflow.

​
Prerequisites

Before starting this task, ensure you have the following:
  • Appropriate permissions: nx1_automl_admin and nx1_crew_admin or nx1_automl_user and nx1_crew.
  • A previously created feature table.

​
Create a training job

Perform these steps to create a training job.
  1. Log in to NexusOne.
  2. On the top navigation bar, hover your mouse over AI and then select AutoML > Training.
  3. Click New training job.
You can also create a training job from an existing feature table using the following step:
  1. On the top navigation bar, hover your mouse over AI and then select AutoML > Feature engineering. You are going to see a list of feature tables.
  2. Select the feature table you want to create a training for.
  3. Click Train model.
  1. Enter a run name to identify this training job.
  2. Select a DataHub domain from the previously created feature table.
  3. Select a source training table schema and table. This is the feature-engineered table NexusOne trains the model on.
  4. Select one of the following problem types to tell NexusOne what kind of prediction you want the model to make:
    • Anomaly detection: Flags records that look out of place compared to the rest of your data. No label column required.
    • Binary classification: Predicts one of two outcomes, such as yes or no, or churn or retain.
    • Contextual bandit: Decides the best action to take given a situation, learning from the outcomes of past actions.
    • Multiclass classification: Predicts one of three or more categories. For example, low, medium, or high risk.
    • Ranking: Orders items by relevance or preference, such as search results or product recommendations.
    • Regression: Predicts a number, such as a price, a score, or a temperature.
  5. Select one of the following algorithms to determine the machine learning technique used to train the model:
    The listed algorithms depend on the problem type you selected.
    • Classifiers:
      • LightGBM classifier: Trains faster on large datasets. Supports SHAP explainability.
      • Vowpal Wabbit classifier: Designed for very large datasets. Learns from data one example at a time.
      • XGBoost classifier: A reliable alternative to LightGBM. Doesn’t support SHAP explainability.
    • Regressors:
      • LightGBM regressor: Trains faster on large datasets. Supports SHAP explainability.
      • Vowpal Wabbit regressor: Designed for very large datasets. Learns from data one example at a time.
      • XGBoost regressor: A reliable alternative to LightGBM regressor. Doesn’t support SHAP explainability.
    • Rankers:
      • LightGBM ranker: Trains faster on large datasets.
      • XGBoost ranker: A reliable alternative to LightGBM ranker.
    • Isolation forest: Detects anomalies by isolating unusual records from the rest of your data.
    • Vowpal Wabbit contextual bandit: Learns the best action to take in each situation from reward signals.
  6. Select one of the following presets to control how the model trains:
    • Balanced: Balances training speed and model quality. This is the default.
    • Best quality: Trains longer with more iterations for the highest accuracy. Best for production.
    • Fast: Trains quickly with fewer iterations. Best for experimentation.
  7. If you selected a problem type other than anomaly detection, then select a label column. The model uses the selected label column as the prediction target during training and predicts its value for new data after training.
  8. If you selected binary classification, then enter a positive class value. It represents the outcome value the model should treat as the positive prediction.
  9. Optional: Select a time column from your training table. NexusOne uses this time column to split training and evaluation data when you select a train cutoff date. NexusOne also drops this time column from the model’s features so the model doesn’t train on it. Without a time column, NexusOne uses an 80/20 random split instead.
  10. Optional: If you selected a time column, then enter a train cutoff date. Rows before this date go to the training data, while rows on or after go to the evaluation data. Requires a time column.
  11. Optional: Select feature columns. If you leave this empty, then NexusOne uses all feature columns.
  12. Optional: Enter an MLflow experiment name to group a training run under a named experiment in MLflow. If it’s left empty, then NexusOne logs the run under a default experiment.
  13. Optional: Enter a registered model name to register the trained model in the MLflow model registry under that name.
  14. If you selected a LightGBM algorithm, then select Run TabularSHAP to explain which features influenced the model’s predictions and by how much.
  15. Select Trigger immediately to start training as soon as you submit the job. When not selected, NexusOne uploads the job but keeps it in a READY state until you trigger it manually.
  16. Click Submit training job to create the training job.

​
Additional resources

For a high-level overview of training jobs and how they’re used in NexusOne, refer to Training.