Paper: Informedia at PDMC, Physiological Data Modeling Contest White Paper Lin, W-H., Proceedings of the 21st International Conference on Machine Learning (ICML'04), Alberta, Canada, July 4-8, 2004 Abstract:   Our Digial Human Memory project (Lin & Hauptmann, 2002) aims to collect and index every aspect of human daily experiences in digital form. By wearing a spy camera, microphones, and a BodyMedia armband, the wearer can collect rich records in a unobtrutive fashion, and many applications can build on top of such multimodal collections. For example, digital human memory can serve as a memory prosthesis to help the wearer recall past events; the habits or anomalies of the wearer can be analyzed from digital human memory. The physiological recordings recorded by a Bodymedia armband provides complementary dimensions of the wearer’s experiences, and play an important role in identifying wearer’s context and activities. In this year Physiological Data Modeling Contest, we build a baseline system that models the gender and context tasks as simple binary classification problems using only unambiguous annotations. In addition, we explore two issues. First, instead of ignoring ambiguous annotations and unlabeled data, we attempt to disambiguate them into positive and negative annotations such that the learner can incorporate them in the training phase. Second, we exploit sequence relationship because context activities do not appear randomly but usually in the consecutive minutes as a cluster. A conditional model is built to incorporate the sequence information. Close