*5.3. Error Case Analysis*

Figure 7 shows the proportion of each activity to the whole error case, and Figure 8 shows the error rate of each activity. The index of each activity is shown in Table 7. Eating with dinnerware shows the highest proportion (40%), followed by sedentary work (30%) and conversation (10%). However, due to the proportion of eating with dinnerware being far greater than that of sedentary work, the error rate is much larger with respect to sedentary work (0.424). As sedentary work and conversation generally show similar patterns in the amount of movement of the hand, and usually happens indoors, the same as with the eating activity, the two activities show a higher error rate than any other activities. However, in the case of walking, as it is typically a dynamic activity easily distinguished from the eating activity, it showed a very low error rate (0.004%; 174 lines out of 39,822 lines). For driving and subway activities, differences of movement and spatial properties make those activities' error rates low.

**Figure 7.** Proportion of the error case.

Figure 9 shows the specific case, which is the eating activity of a left-handed person, who wore the wrist-wearable device on the right wrist and mainly used the left hand to eat, but also used the right hand for moving food, using a smartphone, gesturing in conversation, and so on. Compared to the right-handed person (Figure 1), the accelerometer shows a different pattern, such as a much lower and steady value for the *x*-axis and a higher and irregular pattern of the *y* and *z*-axis, as they used their right hand for various purposes in addition to eating. As a result, the probability of using dinnerware shows very low and high deviance. However, as the person ate in a normal environment like other subjects, the spatial property compensating the final recognition and overall eating probability shows acceptable results. This means that the proposed BN could approximately recognize the complex eating activity when one of the contexts or sensor values has a very different pattern or is even omitted. Note that the proposed method might approximately recognize these cases without incorporating information of which hand the person uses and applying different algorithms. This is important since, in the real world, the person might use different hands for various situations; one might prefer to use the left hand to drink coffee, while using the right hand to eat chicken.

**Figure 9.** Eating activity of a left-handed person.
