Identifying User's Smartphone Possession Location and Behavior Including Untrained Classes by Using Deep Metric Learning
Rui Kitahara, Lifeng Zhang · 2022
In our previous study,(1) we confirmed that it was possible to classify the user’s smartphone possession location and behavior simultaneously using Mdk-ResNet.(2) However, there was a problem that when untrained classes were given as input, it was impossible to output them as unknown classes using normal classifications as in our and other’s previous studies. Therefore, this time, we aimed to output untrained data as an unknown class by learning the distance between images using deep metric learning, and not only classifying the learned classes. In this study, we used a smartphone’s accelerometer to acquire the data of experimental participants and created a trained model by learning through deep metric learning. We classified the user’s smartphone possession location and behavior, including the unknown class of unlearned data, by calculating the cosine similarity between the image vector prepared in advance and the image vector output from the trained model during inference using the trained model. As a result, we confirm not only that the untrained data can be output as the unknown class, but also that the accuracy is comparable to the previous study.