Multiobjective optimization in machine learning
Machine learning problems commonly have to deal with implicit or explicit conflicting objectives, thus leading to trade-offs. One of the most traditional is the model complexity vs. error rate on training trade-off. However, more complex is the model, more capable to adapt to data, thus reducing error; and vice-versa. But there are many other conflicting goals such as: error rate on each class; error rate vs. interpretability; error rate vs. fairness metrics.
Previous research investigated the impact of optimization methods in machine learning, finding interesting impacts of multi-objective optimization in model selection, ensemble diversity, and knowledge sharing for classification problems, and esulted in applications in biology, medicine, logistics, and power transmissions; and also have theoretical contributions to machine learning, operations research, and multi-objective optimization.
This project has three main goals: 1) Create a software framework to ease modeling and running conflicting objectives in machine learning; 2) Investigate theoretical consequences of modeling machine learning problems as multi-objective ones; 3) Creation of models to be employed in a wide sprectrum of applications. As a result, we expect to create a helpful framework that would machine learning experts to model tasks with conflicting objectives.