Claudio Bruderer (Product Manager at Modulos)
Modulos AutoML version 0.4.4 is available! With this version, we are promoting the Time Series Forecasting feature introduced in v.0.4.2 to a Beta feature. This release also contains a preview of the REST API interfaces for the Solutions, which will be fully available soon. And lastly, among many other enhancements and refinements, we are including new ML models and objectives.
Time Series Forecasting Upgraded to a Beta Feature
A multitude of business use cases such as forecasting supply, demand, growth, and the evolution of prices inherently depend on time. To allow you to tackle these, we introduced Time Series Forecasting as an alpha feature in Modulos AutoML v.0.4.2. This feature was initially restricted to forecasting a single time step into the future and was still a bit rough around the edges.
We have significantly expanded the functionalities of this feature with this and the previous release (AutoML v.0.4.3). You can now use it to also do multi-step forecasting, which extends the range of use cases you can implement. Additionally, we have added more plots giving you insights into the performance of your Solution. We have also made the workflow creation and the feature itself more robust by fixing minor early-life bugs.
With these improvements, the Time Series Forecasting feature is now mature enough to be up from an alpha to a beta feature. It is now enabled by default when setting up the platform. Before creating a ML workflow, you can choose whether you want to create a Time Series Forecasting workflow or a standard regression or classification workflow.
With this release, we are also adding a new target objective for Time Series Forecasting: the Mean Absolute Scaled Error (MASE). It is a popular objective, as it is independent of the scale of the data (i.e. performs equally well for small and large prediction values). It also penalizes negative and positive deviations equally and thus addresses weaknesses of other useful objectives for Time Series Forecasting such as the Mean Absolute Percentage Error (MAPE).
Preview of the REST API for the Solution
The Solution, which consists of the trained ML model and feature extractor combination, is the principal output of any workflow trained with Modulos AutoML. The key characteristics of the generated Solutions are that they are fully independent and not tied to the platform. They are also modular, allowing you to easily replace an already deployed, existing Solution with a better performing, new Solution to facilitate model DevOps practices.
We are currently working on extending the deployment options of our Solution. By adding standardized REST API interfaces, we aim to make the Solutions easier to deploy and integrate into your services. This will allow you to call the Solutions using your own language of choice. Furthermore, Solutions can easily be deployed on a server and queried from another service.
Other New Features
New and More Data Science Objectives and Models
Besides the aforementioned Mean Absolute Scaled Error (MASE), we have also added another ML model. Support Vector Machine (SVM) models are now available for your ML workflows. SVM models are used for both classification (predicting a category) and regression (predicting a number) use cases. They are versatile and a popular ML model choice. Our implementation was developed for smaller datasets specifically. Due to the systematic and unbiased search strategy of our platform, Modulos AutoML will identify when SVM models are indeed the best performing models for your use case.
In addition to SVM models, we are also adding a Random Forest implementation for regression tasks. This extends the range of use cases when Random Forest models are applicable. The full and updated list of Data Science modules (ML models, feature extractors, objectives, and optimizers used to find the best ML Solutions) is available here.
Better platform configuration & setup
Another focus area of our development has been to refine how platform administrators set up the platform in their environments. We have streamlined the installation and setup procedure with a comprehensive installation script. In addition, we have cleanly separated the editable configuration entries (e.g. containing networking configurations) from the settings exclusively used by Modulos for debugging purposes. Lastly, we have added more details to the output of the “automl diagnosis” command introduced in AutoML v.0.4.3. This will help the platform administrator to assess the current state of the platform more thoroughly.
OTHER IMPROVEMENTS AND BUG FIXES
We additionally have made a range of other improvements and fixes to AutoML, which include:
- Renamed the “Tutorials” section with the more appropriate term “Documentation”.
- Various improvements and minor bug fixes to the creation process. Among other changes, each step now links to a corresponding documentation entry for help and further information.
- Extended the metadata on workflows to explicitly communicate the specific trigger conditions leading the platform to automatically pausing a workflow.
- Fixed a bug of a persistent page counter when leaving the overview page of individual ML workflows.
- Replaced the text “No Scores!” with dashes (“-”) to more clearly indicate workflows without any valid and trained ML Solutions.
- Added additional checks for slashes (“/”) in feature and component names during the upload process. This avoids issues when querying these features during the training process of ML Solutions.