TY - GEN
T1 - Where am I? Using mobile sensor data to predict a user’s semantic place with a random forest algorithm
AU - Lex, Elisabeth
AU - Pimas, Oliver
AU - Simon, Jörg
AU - Pammer-Schindler, Viktoria
PY - 2013/1/1
Y1 - 2013/1/1
N2 - We use mobile sensor data to predict a mobile phone user’s semantic place, e.g. at home, at work, in a restaurant etc. Such information can be used to feed context-aware systems, that adapt for instance mobile phone settings like energy saving, connection to Internet, volume of ringtones etc. We consider the task of semantic place prediction as classification problem. In this paper we exploit five feature groups: (i) daily patterns, (ii) weekly patterns, (iii) WLAN information, (iv) battery charging state and (v) accelerometer data. We compare the performance of a Random Forest algorithm and two Support Vector Machines, one with an RBF kernel and one with a Pearson VII function based kernel, on a labelled dataset, and analyse the separate performances of the feature groups as well as promising combinations of feature groups. The winning combination of feature groups achieves an accuracy of 0.871 using a Random Forest algorithm on daily patterns and accelerometer data. A detailed analysis reveals that daily patterns are the most discriminative feature group for the given semantic place labels. Combining daily patterns with WLAN information, battery charging state or accelerometer data further improves the performance. The classifiers using these selected combinations perform better than the classifiers using all feature groups. This is especially encouraging for mobile computing, as fewer features mean that less computational power is required for classification.
AB - We use mobile sensor data to predict a mobile phone user’s semantic place, e.g. at home, at work, in a restaurant etc. Such information can be used to feed context-aware systems, that adapt for instance mobile phone settings like energy saving, connection to Internet, volume of ringtones etc. We consider the task of semantic place prediction as classification problem. In this paper we exploit five feature groups: (i) daily patterns, (ii) weekly patterns, (iii) WLAN information, (iv) battery charging state and (v) accelerometer data. We compare the performance of a Random Forest algorithm and two Support Vector Machines, one with an RBF kernel and one with a Pearson VII function based kernel, on a labelled dataset, and analyse the separate performances of the feature groups as well as promising combinations of feature groups. The winning combination of feature groups achieves an accuracy of 0.871 using a Random Forest algorithm on daily patterns and accelerometer data. A detailed analysis reveals that daily patterns are the most discriminative feature group for the given semantic place labels. Combining daily patterns with WLAN information, battery charging state or accelerometer data further improves the performance. The classifiers using these selected combinations perform better than the classifiers using all feature groups. This is especially encouraging for mobile computing, as fewer features mean that less computational power is required for classification.
UR - http://www.scopus.com/inward/record.url?scp=84945984099&partnerID=8YFLogxK
U2 - 10.1007/978-3-642-40238-8_6
DO - 10.1007/978-3-642-40238-8_6
M3 - Conference paper
SN - 9783642402371
T3 - Lecture Notes of the Institute for Computer Sciences, Social-Informatics and Telecommunications Engineering
SP - 64
EP - 75
BT - Mobile and Ubiquitous Systems:Computing, Networking and Services
A2 - Jiang, Hongbo
A2 - Zheng, Kan
A2 - Li, Mo
PB - Springer Verlag Heidelberg
T2 - 9th International Conference on Mobile and Ubiquitous Systems: Computing, Networking and Services
Y2 - 12 December 2012 through 14 December 2012
ER -