At Databricks, AutoML is our low-code/no-code model training API that empowers customers to create quality machine learning (ML) models with their data on our platform. However, data scientists want to focus on delivering insights, not managing clusters or navigating the complexities of data and model governance. They also seek access to state-of-the-art model architectures to achieve the highest quality predictions.
- To address these challenges, we are excited to announce a powerful new capability in Mosaic AI Model Training: Time Series Forecasting. This new AutoML feature brings enhanced flexibility, governance, and performance to help businesses unlock the predictive power of their time series data. Here’s what’s new and how it can make a difference for your forecasting needs:
Serverless Experience for Simplified Model Training
Gone are the days of configuring and managing clusters for training forecasting models. With our serverless solution, you can train models—whether using custom training code or AutoML—without the need to handle complex infrastructure. Databricks automatically optimizes both performance and cost with autoscaling, delivering the best user experience while reducing the operational burden of training and serving time series models. This means more time for you to focus on insights, not infrastructure.
Unified Governance with Seamless Integration
With our new capability, the best model is automatically registered to the Unity Catalog. This integration eliminates the need for customers to maintain a separate set of data governance policies for their models. Prediction results are also automatically stored as Unity Catalog tables. You can now manage models and data under a single governance framework, ensuring better consistency, security, and compliance across your organization.
Higher Quality Models Out of the Box
We are introducing DeepAR, a deep neural network model-based algorithm, to our portfolio of time series forecasting tools. DeepAR delivers up to a 50% improvement in prediction error rate, according to our benchmarks, see the below comparison graph. This new algorithm is enabled by default. Customers can benefit from cutting-edge model performance without the need for additional tuning, making it easier than ever to get high-quality forecasts right out of the gate.
Benchmark datasets: rossmann, walmart, wind, cinema
Improved Usability with New Features
We’ve introduced a host of new features designed to make time series forecasting more customizable and effective:
- More Customization in Data Splits: Now, you can tailor model evaluations with custom Train/Validate/Test data splits that align with the unique patterns and trends in your data. This ensures more accurate assessments and fine-tuning of models.
- Weighted Evaluation for Better Accuracy: Users can assign different weights to individual time series during evaluation, allowing for a focus on the most critical or impactful series in the dataset. This ensures the selected model delivers the best accuracy where it matters most.
- Enhanced User Interface: Our improved UI offers a one-click experience to serve the best model through batch inference or real-time endpoints. This intuitive design makes it easier to deploy models to production, helping you derive value from your forecasts faster.
Get Started Today
Whether you’re forecasting sales to increase revenue, or predicting user trends to enhance engagement, our tool automates the heavy lifting, allowing your team to focus on leveraging insights rather than building complex models from scratch.
Check out the documentation to get started.
Source link
lol
