ULTech to Observe Elderly’s Behavior Changes over Time in SH
135
behavior change period. Activities in this period has been mapped and compared
with normal life pattern.
In the domain of at-home activities, anomalies can be classified as point,
collective and contextual anomalies. The point anomaly [12–14] considers each
activity independently and decides whether it is normal or not with respect to
the normal behavior. The collective anomaly [15] considers groups of activity
instances together to determine whether the group is normal or not. The contextual anomaly [16,17] considers activities under a particular context (e.g., day
of week, person under medication, etc.). In our work, we focus on detecting
point anomaly (i.e., missing activity or activity with an unusually long/short
duration).
For the activity “going to the toilet” which occurs several times per day, we
model and compare frequencies between a routine day and the observed day in
the behavior change period. A frequency is provided by the normal behavior
pattern. This will be illustrated in the following section.
4 Use Case
In this section, we present in 4.1 the dataset used as use case. We detail in 4.2
the learning steps for building the normal behavior pattern. By plotting the
daily score along time in 4.3, we exhibit the period when the score changes over
time and then identify daily anomalous activities found in the behavior change
period.
4.1 Dataset
The dataset used for our analysis is provided by Washington State University’s CASAS program
1 [18]. CASAS (Center for Advanced Studies in Adaptive
Systems) aims to provide aid to residents using smart home technology. They
therefore collect and use real-time data from sensors to analyze and monitor
residents’health and behavior to improve future smart home living.
We use one public data set (named HH120) [18] which was used in other works
like [19]. It includes one unique subject, covering a total of 63 days. All data used
in this paper was handled in an anonymized way. The set of activities includes
“sleeping”, “eating meals”, “taking a shower”, “going outside” and “going to the
toilet”. The data sets do not provide any medical information. For training the
normal behavior model, we use the first month while the rest of the available
data is used to test the effectiveness of our proposals.
4.2 Learning for Building the Normal Behavior Pattern
For building the normal behavior pattern from the training data set, we follow the different learning steps presented in the Sect. 3.2. We then apply the
1 http://casas.wsu.edu/.
135
behavior change period. Activities in this period has been mapped and compared
with normal life pattern.
In the domain of at-home activities, anomalies can be classified as point,
collective and contextual anomalies. The point anomaly [12–14] considers each
activity independently and decides whether it is normal or not with respect to
the normal behavior. The collective anomaly [15] considers groups of activity
instances together to determine whether the group is normal or not. The contextual anomaly [16,17] considers activities under a particular context (e.g., day
of week, person under medication, etc.). In our work, we focus on detecting
point anomaly (i.e., missing activity or activity with an unusually long/short
duration).
For the activity “going to the toilet” which occurs several times per day, we
model and compare frequencies between a routine day and the observed day in
the behavior change period. A frequency is provided by the normal behavior
pattern. This will be illustrated in the following section.
4 Use Case
In this section, we present in 4.1 the dataset used as use case. We detail in 4.2
the learning steps for building the normal behavior pattern. By plotting the
daily score along time in 4.3, we exhibit the period when the score changes over
time and then identify daily anomalous activities found in the behavior change
period.
4.1 Dataset
The dataset used for our analysis is provided by Washington State University’s CASAS program
1 [18]. CASAS (Center for Advanced Studies in Adaptive
Systems) aims to provide aid to residents using smart home technology. They
therefore collect and use real-time data from sensors to analyze and monitor
residents’health and behavior to improve future smart home living.
We use one public data set (named HH120) [18] which was used in other works
like [19]. It includes one unique subject, covering a total of 63 days. All data used
in this paper was handled in an anonymized way. The set of activities includes
“sleeping”, “eating meals”, “taking a shower”, “going outside” and “going to the
toilet”. The data sets do not provide any medical information. For training the
normal behavior model, we use the first month while the rest of the available
data is used to test the effectiveness of our proposals.
4.2 Learning for Building the Normal Behavior Pattern
For building the normal behavior pattern from the training data set, we follow the different learning steps presented in the Sect. 3.2. We then apply the
1 http://casas.wsu.edu/.