Abstract
This paper develops a model for multi-task machine learning that incorporates per-task parametric and nonparametric effects in an additive way. This allows a practitioner the flexibility of modeling the tasks in a customized manner, increasing model performance compared to other modern multi-task methods, while maintaining a high degree of model explainability. We also introduce novel methods for task diagnostics, which are based on the statistical influence of tasks on the model’s performance, and propose testing methods and remedial measures for outlier tasks. Additive multi-task learning model with task diagnostics is examined on a well-known real-world multi-task benchmark dataset and shows a significant performance improvement over other modern multi-task methods.