FEDOT is an open-source framework for automated modeling and machine learning (AutoML). The main idea of the FEDOT is to provide open, modular, extendable, and flexible instruments to the community to solve non-standard modeling tasks.
It can build composite modeling pipelines for different real-world processes in an automated way using an evolutionary approach. FEDOT supports classification (binary and multiclass), regression, and time series prediction tasks, as well as different data types and multi-modal cases. Also, sensitivity analysis of the pipelines, custom pipelines design as the initial assumption of optimization, domain-specific objective functions and models, and other interesting features are implemented. FEDOT can be evaluated with both CPU and GPU.
Current plans for the project development are:
- Support for the time series classification, hierarchical time series forecasting, etc;
- Better support of the multi-modal AutoML;
- Support for meta-AutoML to as initial stage for composing;
- Controllability and interpretability features;
- Additional benchmarking for different tasks and datasets;
- Implementation of reinforcement learning and hybrid evolutionary optimization;
- Integration with MLFlow and other MLOps tools;
- Automated design for other types of models using the core of FEDOT (PDEs, bayesian networks, etc) and a lot of other ideas.
FEDOT is maintained by the Natural Systems Simulation Lab of ITMO University, Russian. Also, any external software developers or AI researchers can use FEDOT or contribute to it. We are open to any feedback.
Also, several posts on towardsdatascience if you want to go deeper: