OASIS Online Active SemI-Supervised Learning.pdfVIP

OASIS: Online Active SemI-Supervised Learning Andrew B. Goldberg Xiaojin Zhu and Alex Furger and Jun-Ming Xu Arcode Corporation University of Wisconsin–Madison, Computer Sciences Department Bethesda, MD 20814 Madison, WI 53706 goldberg@ {jerryzhu,furger,xujm }@ Abstract with an online algorithm that updates its decision function based on incoming unlabeled data. Many real-world learn- We consider a learning setting of importance to large scale ing tasks fit nicely into this framework, such as classifying machine learning: potentially unlimited data arrives sequen- images collected by a surveillance camera, or categorizing tially, but only a small fraction of it is labeled. The learner cannot store the data; it should learn from both labeled and blog posts and tweets in real-time as they emerge on the unlabeled data, and it may also request labels for some of social Web. We present a novel, fully Bayesian algorithm the unlabeled items. This setting is frequently encountered capable of online active semi-supervised learning (OASIS). in real-world applications and has the characteristics of on- The learning setting we consider is similar to that in Gold- line, semi-supervised, and active learning. Yet previous learn- berg, Li, and Zhu (2008), but with an optional active learn- ing models fail to consider these characte