API Workflow
The core steps in the TimeDetect workflow are:
- Upload historical approved registration: Establish a baseline for model training
- Train models: Train models based on historical data to find anomalies in new registrations
- Create predictions: Generate predictions (identify anomalies) for new registrations
- Get results: Retrieve prediction results (anomalies)
- Upload new approved registrations: Continuously update the historical data with new approved registrations
- Update models: Maintain model accuracy by updating or retraining models
Optimizing your workflow involves the strategic scheduling of various jobs. Below are recommendations for how to efficiently handle data uploads, training, and predictions.
Data upload
New approved registrations should be uploaded daily. This practice keeps your dataset current and maximizes the relevance of your predictions.
After each upload, wait for the validation job to finish, confirming that the data is ready for the next step.
Initiating training
Full Retrain: Conduct a full training with rebuildModels=True once per week. This comprehensive approach accounts for any significant changes or trends in the data.
Update Models: Initiate training with rebuildModels=False if there is any new approved registrations, e.g. every evening. This configuration allows for the inclusion of new data points without undergoing the full training process, optimizing model freshness and performance.
Prediction Strategy
Real-Time Prediction: For immediate feedback, you can ask for real-time prediction for every new registration.
Batch Predictions: Ask for batch predictions on all unapproved registrations nightly. This approach gives you a comprehensive overview, capturing trends and insights from the aggregated results.