Non-IID Learning in Big Data

Learning from big data is increasingly becoming a major challenge and opportunity for big business and innovative learning theories and tools. Some of the most critical challenges of learning from big data are the uncovering of the explicit and implicit coupling relationships embedded in mixed heterogeneous data from single/multiple sources. The coupling and heterogeneity of the non-IIDness aspects form the essence of big data and most real-world applications, namely the data is non-IID. Most of classic theoretical systems and tools in statistics, data mining, database, knowledge management and machine learning assume the independence and identical distribution of underlying objects, features and values. Such theories and tools may lead to misleading or incorrect understanding of real-life data complexities. Non-IID learning in big data is a foundational theoretical problem in AI and data science, which considers the couplings and heterogeneity between entities, properties, interactions and contexts. In this tutorial, we present a comprehensive overview of the non-IID learning. We begin with the limitation of IID learning in handling big data and introduce abstract learning model and representation for non-IID learning, and then present frameworks and algorithms for non-IID metric learning, classification, clustering, ensemble clustering, outlier detection, feature selection, recommender systems, and text mining, and finally discuss open challenges and prospects. Link to tutorial: http://kdd2017tutorial-noniid.datasciences.org/