Transductive Experiment Design

摘要: This paper considers the problem of selecting most informative experiments x to get measures y for learning an inference model = f(x). We propose a novel concept active learning, transductive experiment design, overcome shortcomings existing design methods, e.g. insucient exploration available unmeasured data and poor scalability large sets. In-depth analysis clearly shows that method tends favor are hard predict meanwhile typical in representing remaining hard-to-predict data. Ecient solutions further developed through mathematical programming techniques. Encouraging results on toy problems real-world sets included highlight advantages proposed approaches.