Using meta‐learning to predict performance metrics in machine learning problems

Machine learning has been facing significant challenges over the last years, much of which stem from new characteristics machine problems, such as streaming data or incorporating human feedback into existing datasets and models. In these dynamic scenarios, change time models must adapt. However, do not necessarily mean patterns. The main goal this paper is to devise a method predict model's performance metrics before it trained, in order decide whether worth train not. That is, will model hold significantly better results than current one? To address issue, we propose use meta-learning. Specifically, evaluate two different meta-models, one built for specific problem, another based on many meant be generic meta-model, applicable virtually any problem. paper, focus only prediction root square error (RMSE). Results show that possible accurately RMSE future models, event scenarios. Moreover, also reduce need re-training between 60% 98%, depending problem threshold used.