If 50% of your data is close to the mean with a variation of 1, Mean Squared Error (MSE) might be the better choice, provided the remaining data is relatively clean and free from significant outliers. MSE penalizes larger errors more heavily, which can lead to faster convergence and finer adjustments in the model. However, if the rest of your data includes outliers, Mean Absolute Error (MAE) would be preferable due to its robustness, treating all errors equally and ensuring that outliers do not disproportionately affect the model's performance.