Real-world complex systems such as social media, e-commerce platform, cyber-physical system, or chemical synthesis, are often associated with heterogeneous (multi-modal) data, including structural relation or graph/network, unstructured text or image, temporal context, and so on. The heterogeneous data provide opportunities for researchers and practitioners to understand complex systems more comprehensively but also pose challenges to discover knowledge from them. Besides the difficulty of extracting and representing useful information from such complex data, it is hard to fuse the extracted knowledge in a unified and customized manner so as to facilitate various applications.In this thesis, we develop the methodologies and algorithms for learning from heterogeneous data through fusion, which have been deployed and validated in a variety of real-world applications, e.g., recommender system, relevance search, graph embedding, anomaly detection. Extending from the fusion learning, we further investigate the principles and methodologies for label-efficient fusion learning from heterogeneous data, with the emphasis on the applications of web personalization and knowledge graph reasoning.