We used J48 decision trees to recognize these contexts. Using features total are 63: six Average Acceleration, one Average Difference Acceleration between Devices, two Average Resultant Acceleration, 48 Bands of Frequency Power, and three Maximum Bands of Frequency Power and then three Frequency. We performed a long-term experiment to evaluate how accurately it can recognize 24 contexts and whether it can recognize these contexts when using data gathered in the past as training data. By performed a experiment over long term, we can gather data that have little affect of a learning effect because participants forget about a experiment. Therefore, we can evaluate how accurately it can recognize these contexts using themselves data without a learning effect. In this experiment, we performed this experiment twice (1st and 2nd rounds) over approximately five month. Twelve male …show more content…
\figref{fig:opepos}) and then gathered acceleration data of the smarphone and smartwatch. Firstly, we asked the participant to wear a smartwatch on his left wrist. After wearing the watch, we also asked the participant to perform contexts that were showed the smartphone's display in a randomized order for approximately ten seconds or until we give the participant a sign. These contexts are described by using document with the instructions. When the participants perform a'\,--\,j', we asked the participant to perform while walking between signs: the diameter between the signs is 10\,meters (approxiametely 10.94\,yd). Moreover, if participants can operate a smartphone, we asked the participant to operate a image viewer application by using swipe or tap by intervals of approximately one second. The participant was asked to perform contexts again if the participant perform the different order context. In each session, he performed the 24 contexts once in a randomized order. The participant took a break of five minutes between two