Structural Learning of Activities from Sparse Datasets - F. Albinali, N. Davies, A. Friday
@inproceedings{Albinali:2007:SLA:1263542.1263715,
author = {Albinali, Fahd and Davies, Nigel and Friday, Adrian},
title = {Structural Learning of Activities from Sparse Datasets},
booktitle = {Proceedings of the Fifth IEEE International Conference on Pervasive Computing and Communications},
year = {2007},
isbn = {0-7695-2787-6},
pages = {221--228},
numpages = {8},
url = {http://dl.acm.org/citation.cfm?id=1263542.1263715},
doi = {10.1109/PERCOM.2007.33},
acmid = {1263715},
publisher = {IEEE Computer Society},
address = {Washington, DC, USA},
}
Key Points
- Sample size was increased using Efron's bootstrapping
- Metrics which had a low collinearity yielded better classification results due to less noise from cross over
Using sensor data from a home deployment, a Bayesian classifier was used to detect Activities of Daily Living (ADLs). Each ADL is made up of many different data features which need to be extracted. In order to reduce cross contamination of features in different ADLs, features which had a very low collinearity were chosen to classify the ADL.
To increase the sample size for training, Efron's bootstrapping was used to extend the data. 100 instances of an activity were used for training, and 20 were used for validation. Results were good with a mean classification accuracy of 85%. Adding extra features, but with high collinearity decreased the overall accuracy.