Pervasive computing based home-monitoring has attracted increasing interest over the past years, especially regarding applications in the growing population of older adults. Applications include safety, monitoring chronic conditions like dementia, or providing preventive information about changes in health and behavior. Commonly used components of such systems are inexpensive and low-power passive infrared motion sensing units, usually placed in distinct locations of an older adult’s apartment. To efficiently analyse the resulting data the majority of procedures expect the resulting sensor data to be encoded in a vector space. However, most common vector space encodings are based on orthogonal representations of the sensor locations and thus lead to loss of information as the sensors are placed in a 3D-space. In this work we introduce an embedding of sensor-locations in a 2D-space based on multidimensional scaling, without knowledge of the physical position of the sensors. We evaluate this embedding, using two different algorithms and compare it to commonly used baselines in different tasks. All evaluations are carried out on a real-world home-monitoring data-set.
