Recent years have witnessed a prospering of data-driven systems, such as e-commerce, social networks, online learning, digital health, and sharing economy applications. These systems have accumulated a large amount of user-generated data, which help to personalize the user preferences, understand their information needs, and provide satisfactory experience for the users. However, the data can come in very heterogeneous, dynamic, and extremely unstructured forms, such as free-texts, ratings, click series, images, or videos, which make it a difficult task to profile the users for personalized services. In this talk, I will introduce data-driven techniques for personalized recommendation systems, which include 1) Leveraging sentiment analysis on textual reviews for explainable recommendation; 2) Modeling the shifting of user preferences for dynamic recommendation; 3) Unified representation learning from heterogeneous data sources for multi-view preference modeling; and 4) The economic nature of online systems. As a conclusion of the talk, I will also provide my future vision on personalization theories for broader and emerging application scenarios, such as personalized education, personalized healthcare, personalized smart home devices, NLP for recommendation, and privacy-preserving recommendation systems.