Nowadays, we have access to millions of data examples from various sources. However, not all of them are relevant to the problem we are trying to solve at hand. In order to train an effective machine learning model, one needs to select the most useful data to conduct learning and inference, which is Junfeng's research focus. In this talk, Junfeng will first provide an overview of his research in transfer learning, reinforcement learning and federated learning. Then he will showcase a few developed algorithms, including multi-source domain adaptation and personalized federated learning.