Abstract: Social media gains great prosperity in the last decade, facilitating users to create, share, and propagate information or ideas. Social media records users’ profiles and opinions, social relationship among users, and information cascades, allowing us to study human dynamics with an unprecedented breadth, depth and scale. Social media analytics aims to gather and analyze the data from social media to support business intelligence and decision-making. In this talk, I will introduce several practices on social media analytics from our team. First, I will talk about inferring interpersonal influence among users. User’s influence is one key factor for information propagation, yet is not directly observable. I will introduce how to infer users’ interpersonal influence from records of information cascades on social media. Second, based on the inferred interpersonal influence, I will present an efficient influence maximization algorithm that is applicable to viral marketing on large-scale social media. Finally, I focus on modeling and predicting popularity dynamics of information on social media. We aim to uncover universal mechanisms or rules governing information propagation, and to design effective methods to predict the final popularity of individual information given their early propagation cascades. In sum, these practices on social media analytics offer us useful tools for mining social media and leverage social media for business intelligence and decision-making. We also discuss some problem with potential applications on social media analytics.
Data Science is an interdisciplinary area between computing, mathematics, statistics, analytics, methods, machine learning, data processing and domain expertise. Mastering Data Science often will need to have clear understanding about the data, which methods suitable to deal with the research challenges, extract important results and explain fully and accordingly to the goals of research or business requirements. Similar methods can be applied to different disciplines if used appropriately.
This keynote presents the latest research outputs for Data Science and Analytics services including healthcare, finance, data center computing, social networks, security and weather studies. In particular, two services will be at the center of attention. First, the health informatics service allows us to understand more about tumor, genes and proteins, as well as interpret part of how our human body can function. This includes the study of malignant tumor and genes that are prone to certain types of cancers. Second, it is the financial deep learning. We can identify the correlation between the current and the past stock movement from the best matching scenarios. Various techniques and results will be discussed in details. These two real-life examples can represent one of the next generation of Data Science and Analytics Services.
His current reserach interest includes intelligent information processing, machine learning and theories in numerical modeling. He proposes the L(1/2) regulation theory, which serves as foundations for sparse microwave imaging. He also discovers and proves Xu-Roach Theom in machine learning, which solves several difficult problems in nueral networks and simulated evolutionary computation，and provides a general deduction criteria for machine learning and nonlinear analysis under Non-Euclidean fromwork. lastly, he initiates new modeling theories and methodologies based on visual cognition, and formulates series of new algorithms for clustering analysis, discriminant analysis, and latent viable analysis, which have been widely applied to science and engineering. He is owner of the National Natural Science Award of China in 2007，and winner of CSIAM Su Buchin Applied Mathematics Prize in 2008. He delivered a sectional talk at International Congress of Mathematicians (ICM 2010) upon the invitation of the congress committee.
Professor Xu currently makes several important services for government and professional societies, including Consultant Expert for National (973) Program in Key Basic Science Research and Development (Information group), Ministry of Science and Technology of China; Evaluation Committee Member for Mathematics Degree, Academic Degree Commission of the Chinese Council; Committee Member in Scientific Committee of Education Ministry of China (Mathematics and Physics Group); Vice-Director of the Teaching Guidance Committee for Mathematics and Statistics Majors, the Education Ministry of China; Director of the Teaching Guidance Committee for Mathematic Education，the Education Ministry of China; Member in the Expert Evaluation Committee for Natural Science Foundation of China (Computer Science Group), The National Committee for Natural Science Foundation of China; Vice-president of Computational Intelligence Society of China; Editor-in-chief of the Textbooks on Information and Computational Sciences, Higher Education Press of China; Co-editor of nine national and international journals.
Abstract: Deep learning (DL) has becoming a powerful, standard AI technology which
helps to yield increasingly breakthroughs of learning system applications. As a representative of data driven approach, it faces however many challenges like contradictions between standardization and personalization, versatility and efficiency, the difficulties in design, anticipation and explanation for the results, and the serious dependence upon the amount and quality of training samples. On the other hand, the model-driven approach provides another learning paradigm that bases on the physical mechanism and prior modeling, which has the characteristics of determinacy and optimality while meets with obstacle of impossibility of precise modeling. In this talk we propose and formalize a data & model dual-driven learning approach, which define then the model driven deep learning (MDDL).
The model driven deep learning start with construction of a Model Family (MF), which is a rough description of solution of the problem under consideration, followed then by the design of an Algorithm Family (AF) which is a collection of iterations whose limit give the solution of the model family. The Algorithm Family then unfolded into Deep Architecture (DA) with which
learning can be performed. We provide examples to substantiate the effectiveness and superiority of the MDDL over others. We particularly show the following advantages of MDDL: It recedes the requirement for precise modeling in model-driven learning, provides the sound methodology for the DL network design, making it easy to incorporate into prior knowledge to make DL more efficient, designable, predictable and interpretable, and also significantly reduce the number of samples needed for DL training. Based on this study, we conclude that MDDL has great potential in the future DL research and applications.