Enhancing the Probability Models for Inference of Significant Activities Using a Real-time Learning Machine in Smartphone

Keqiang Liu, Ruizhi Chen and Tianxing Chu, Yunjia Wang

Peer Reviewed

Abstract:	This paper extends an initial study on significant activities inference by using Bayesian theorem with contextual tuples. The real-time contextual tuples are constructed with local time, significant locations, location-dwelling states and user states. In the initial study, a Naive Bayes classifier with empirical probability distributions was adapted and reached about 80% inference accuracy. The empirical distributions are universal yet not appropriate for everyone. The best way to improve the performance and personalize the classifier is training on personal data set. However, conventional training methods are operated on full data set. This makes it inefficient with duplicate computation and inconvenient for incremental data. This paper presents a Real Time Learning Machine (RTLM) for Naive Bayes to personalize classifier on real-time. Then it conducted a live experiment that compared labeled truth activities and inferred activities by three testers during three days at TAMUCC campus and all the total tuples number are over 80,000. From the experiment results, it can be seen that training classifier has higher inference accuracy than empirical one. And RTLM reaches a similar accuracy to traditional training methods after 40,000 tuples training while the RTLM can supply training distributions and accuracy on real-time.
Published in:	Proceedings of the 28th International Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS+ 2015) September 14 - 18, 2015 Tampa Convention Center Tampa, Florida
Pages:	2055 - 2059
Cite this article:	Liu, Keqiang, Chen, Ruizhi, Chu, Tianxing, Wang, Yunjia, "Enhancing the Probability Models for Inference of Significant Activities Using a Real-time Learning Machine in Smartphone," Proceedings of the 28th International Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS+ 2015), Tampa, Florida, September 2015, pp. 2055-2059.
Full Paper:	ION Members/Non-Members: 1 Download Credit Sign In