Development of Accelerator-based Activity Recognition Technique Robust for Missing Data

这项研究考察了即使在部分传感器数据丢失的情况下也可以执行的活动识别技术。从现有的大多数活动识别技术中使用的模式识别技术的过程中，研究了缺失传感器数据的补齐、缺失特征值的补齐、剩余特征的分类。对于缺失传感器数据的补全，采用ARAR模型的序贯数据预测。对于缺失特征值的补齐，采用了多元回归和核回归的方法。对于具有剩余特征的分类，通过在每个传感器上人为地省略传感器数据来学习多个分类器，并在传感器丢失的情况下适当地选择分类器。一些实验表明，在很多情况下，对缺失的传感器数据或特征值进行补齐比处理分类器的选择具有更好的性能。此外，根据缺失传感器的位置不同，合适的方法也不同。结果表明，对缺失传感器的位置采用适当的定位方法，处理缺失传感器数据的性能几乎与没有缺失情况或f-度量退化约0.03个点的性能相当。

This study examined the activity recognition technique that can perform even when a part of sensor data is missing. From the process of pattern recognition technique, which is used in most existing activity recognition techniques, completion of missing sensor data, completion of missing feature value, classification with remaining features are examined. For the completion of missing sensor data, the sequential data prediction with the ARAR model is used. For the completion of missing feature value, multiple regression and kernel regression are exploited. For the classification with remaining features, multiple classifiers are learned with sensor data artificially omitted on each sensor, and are selected appropriately in the situation of the loss of sensor. From some experiments, the completion of missing sensor data or feature value achieves better performance than coping with the selection of classifiers in many cases. In addition, the appropriate method differs depending on the position of missing sensor. Then, it was found that using appropriate method for the position of missing sensor achieves almost same performance as no missing case or degradation of about 0.03 point in f-measure for coping with missing sensor data.