Reconsidering Learning Objectives in Unbiased Recommendation: A Distribution Shift Perspective