Overview

SurveyAnalytica provides a complete machine learning lifecycle within the platform, enabling you to create, train, version, evaluate, and deploy ML models without writing code. The ML capabilities span tabular data classification and regression, image classification with AutoML and transfer learning, and custom model training with dedicated compute resources.

Navigating to Models

Access the ML model management interface from the Data section. When viewing a dataset, the Predictions view within analytics shows trained models and their results.

Creating a Model

To create a new model, provide the following information:

• Name: A descriptive name for your model.
• Category: The type of ML task — classification, regression, clustering, or forecasting.
• Model Type: The algorithm to use. Options include:
  • Tabular models: Logistic regression, boosted tree, deep neural network, random forest, XGBoost, ARIMA time series, K-means clustering
  • Image models: AutoML vision, transfer learning models
  • Text models: Language models, AutoML text
  • AutoML: Automated tabular and classification models
• Workspace: The workspace this model belongs to.
• Data Entity: The dataset to train on.
• Input/Output Schema: Define the expected input features and output predictions.

Training a Model

Starting Training

Once a model is created, start training by configuring:

• Dataset: The data source for training — tabular data or image data.
• Target column: The column to predict (for supervised learning).
• Hyperparameters: Algorithm-specific settings such as learning rate, batch size, epochs, number of trees, etc.
• Compute resources: The level of compute resources allocated for training.
• GPU acceleration: Optional GPU resources for image and deep learning models.
• Budget hours: Maximum training time allowed.

Training Options

Tabular Data Training

For structured data, training runs on an optimized backend designed for fast, scalable ML on tabular data. This is the fastest option for structured datasets and supports all standard ML model types.

Image and Custom Model Training

For image classification, transfer learning, and custom models, training runs on dedicated compute resources with GPU acceleration. Supported training paths:

• AutoML Tabular: Automated feature engineering and model selection.
• AutoML Image: Automated image classification model training.
• Transfer Learning: Fine-tune pre-trained models on your data. Configurable epochs, learning rate, batch size, and image size.
• Custom Models: Train custom deep learning models with full control over architecture.

Monitoring Training

Training progress is tracked in real-time through:

• Live progress stream: Real-time progress updates pushed to the UI showing stage (initializing, training, complete), percentage, and messages.
• Status checking: Poll-based status checking with automatic detection and recovery of stale statuses.
• Activity tracking: All training activities are logged with timestamps, costs, and metrics.

Grid Search (Hyperparameter Tuning)

For systematic hyperparameter optimization, use Grid Search to automatically train multiple model variants:

1. Define a parameter grid specifying the values to try for each hyperparameter.
2. Set maximum parallel jobs (default: 5).
3. The system generates all parameter combinations and submits each as a separate training job.
4. Results are tracked and compared, showing metrics for each combination.

Example parameter grid:

{
  "learning_rate": [0.001, 0.01, 0.1],
  "batch_size": [16, 32, 64],
  "epochs": [10, 20, 50]
}

This would generate 27 combinations (3 x 3 x 3) and train each one.

Model Versioning

Each model supports multiple versions, allowing you to iterate and compare:

• Add version: Each training run creates a new version with its own hyperparameters, metrics, and artifact references.
• Live version: Promote a specific version to “live” status for deployment and inference.
• Archive version: Archive old versions to clean up while preserving history. Optionally clean up cloud artifacts.
• Delete version: Permanently remove a version and its artifacts.
• Refresh metrics: Re-fetch metrics for all versions, recovering any stale training statuses.

Version Metadata

Each version stores:

• Hyperparameters used for training
• Training configuration (target column, dataset, model type)
• Training job reference
• Metrics (accuracy, precision, recall, F1, AUC, loss, etc.)
• Runtime minutes and estimated cost in USD
• Model artifact references (for deployment)

Model Deployment

Trained models can be deployed as API agents for real-time inference:

• Tabular models: Deployed with automatic model reference resolution. The system verifies the model exists before deployment.
• Image/custom models: Deployed with exported model artifacts. The system ensures the export is available before creating the agent.

Deployment creates a containerized service that exposes the model as a REST API. See the AI Agents article for more details on agent management.

Downloading Models

Download trained model artifacts in various formats via secure, time-limited download links. Select the model, version, and format to generate a download link.

Activity Timeline

The ML Activities timeline shows a chronological history of all ML operations in your workspace, including training runs, grid searches, deployments, and their statuses (running, completed, failed). Each activity includes timestamps, costs, and links to the associated model and version.