An online brokerage firm's data scientist reports that their Amazon SageMaker Linear Learner model training sessions are consistently failing to converge and produce meaningful results, even though data normalization is enabled. How should they address the consistent failures and lack of convergence in these training attempts?