SymReg-MT: Iterative multi-task feature learning through weighted symbolic regression
|Title||SymReg-MT: Iterative multi-task feature learning through weighted symbolic regression|
|Booktitle||LMCE 2014 - 1st International Workshop on Learning over Multiple Context, Nancy, France, September 19, 2014|
|How published||Workshop Paper|
We present a novel approach to unify regression models learnedin parallel on differnet but related datasets using multi -task feature learning based on symbolic regression. The FFX framework (Fast Function Extraction) is used for symbolic regression. It relies on regularized linear regression instead of genetic programming, thus providing a scalable and deterministic framework for implementation. FFX provides a basis for an iterative multi-task feature learning approach. Models learned on separate tasks are coupled by iteratively promoting common terms and penalizing seldom occurring terms, leading to improved consistency and interpretability of models across tasks and improved stability onnew data. We conducted experiments on both real world datasets and synthetic datasets. The results show that already the use of this basic approach leads to models which share more features, show less complexity and still retain the same model performance when compared to a single symbolic regression run. Finally, we provide ideas and plans for future improvements based on our first implementation.