*5.1. Data Specification*

For the experiment, we collected 948 min of data from 25 different volunteers for 10 activities. Subjects were asked to wear a wrist-wearable device and have a smartphone, performed activities that they wanted to perform, and tagged the activity they were doing on the smartphone when the new activity started. They were also asked not to perform more than one activity simultaneously to collect accurate sensor data for each class. If they performed another activity that were not supposed to be collected, such as moving to another place or getting a phone call, collection was temporarily stopped. To collect as much real-life data as possible, we did not request them to come to a certain place; instead, we went to where they lived while performing their daily activities and collected the data. When a self-tagging was difficult, like for a baby or the elderly who are not familiar with a smartphone, we observed and tagged their activities simultaneously. Each subject performed, at most, four different activities and each activity was prolonged for, at most, 20 min to prevent a small number of subjects from dominating most of the data. A specific distribution of each item is shown in Table 6, and indices of activities and jobs are shown in Table 7. We attempted to balance the gender of the subjects, and chose the list of activities by referencing Activities of Daily Livings (ADLs) which is known as a proper method describing the functional status of a human, performing an important role in a healthcare service [19]. 'Etc' in the job includes a four-year old baby. An eating activity consists of 47.27% (448 min out of 948 min), so the data is well-balanced in terms of the eating activity.

Table 8 shows a brief comparison of the collected data with other popular open data for HAR: Opportunity dataset [20] and Skoda dataset [21]. Note that as our approach is supposed to recognize various real eating activities with people with various contexts, we focused on collecting the data from

a sufficiently large number of subjects, so the length of collected data for each subject is relatively small, which is supposed to capture short intervals of daily life, mainly including eating activities. Additionally, note that we tried to use very limited sensors and devices, which are supposed to only include low-power sensors that are easy to use in daily life.


**Table 6.** Data specification.

**Table 7.** Index of activities and jobs.


**Table 8.** Comparison of our dataset with another open dataset for HAR.

