Evaluating Efficacy of MAML based Approach on Regression using Astronomical Imbalanced Dataset

Although meta learning based approach has gained huge popularity in classification tasks, its efficacy was experimented less in the context of regression. This article investigates and evaluates efficiency of a meta learning based approach namely Model Agnostic Meta-Learning (MAML) coupled with neural network model (MLP) on a regression task. The regression task experimented in this article is to estimate distance of galaxies from earth using intrinsic properties of galaxies. This application deals with highly skewed target variable that imposes a greater challenge in obtaining accurate estimate from predictor. To tackle this issue, instead of directly applying a skewed target to the model, square root transformation has been applied and subsequently through task distribution approach of model agnostic meta-learning (MAML), entire dataset was divided into multiple tasks that help in overall reduction of error. Empirical results show satisfactory performance for regression metrics such as Mean Absolute Error (0.078), RMSE (0.179) and R2 score (0.887) for highly imbalanced data. Additionally, comparative performance with the traditional neural network has been demonstrated to indicate MAML performs better in terms of reporting overall low Mean Absolute Error(MAE), NMAD and bias compared to traditional Neural network.