Learning embedding layers (for classes, words, items, etc.) is a key component of lots of applications, ranging from natural language processing, recommendation systems to electronic health records, etc. However, the frequency of real-world items follows a long-tail distribution in these applications, causing naive training methods perform poorly on the rare items. A line of previous works address this problem by transferring the knowledge from the frequent items to rare items by introducing an auxiliary transfer loss. However, when defined improperly, the transfer loss may introduce harmful biases and deteriorate the performance. In this work, we propose a harmless transfer learning framework that limits the impact of the potential biases in both the definition and optimization of the transfer loss. On the definition side, we reduce the bias in transfer loss by focusing on the items to which information from high-frequency items can be efficiently transferred. On the optimization side, we leverage a lexicographic optimization framework to efficiently incorporate the information of the transfer loss without hurting the minimization of the main prediction loss function. Our method serves as a plug-in module and significantly boosts the performance on a variety of NLP and recommendation system tasks.