IEEE Computational Intelligence Magazine - February 2022 - 40

GravityAcc-min-X ≤ 0.096
True
LAYING
False
BodyACCJerk-bandsEnergy ≤ -0.974
True
False
True
GravityAcc-energy-Y ≤ -1
True
False
STANDING
SITTING
angle(Y,gravityMean) ≤ 0.14
False
BodyGyro-min-X ≤ 0.845
False
True
STANDING
SITTING
BodyACCMag-energy ≤ -0.525
True
GravityAcc-arCoeff-Z ≤ 0.564
False
True
False
BodyGyro-correlation-Y ≤ 0.276S
True
WALKING DOWNSTAIRS
False
WALKINGWALKING UPSTAIRSWALKING UPSTAIRS
(a)
1.44%
2.93%
69.4%
22.57%
81.65%
21.73%
(b)
28%
35%
28%
84%
28% 99%
(c)
FIGURE 6 Example model interfaces using (a) Decision tree (b) non-pictorial prototype representation (PType), and (c) PIP pictorial representation
(PIP) for UCI-HAR datasets.
observe that x 01 .= and e 00 . 8=
perform well across diverse
datasets. Similarly, we terminate weight updates after observing
10 to 20 consecutive epochs of decreasing L values. The
hyperparameters are summarized below.
❏ Learning rate {, ,,}e eee3 2113 332- --❏
Batch size {16, 32, 64}
❏ Encoding size {32, 64, 128}
❏ Weight learning rate ()e {. ,. ,. ,. }
❏ Weight threshold ()x {. ,. }
05 01 0080 05
01 005
❏ Updating weight period {10, 20, 30}
We employed a grid search to find the set of hyperparameters
that are most effective across multiple datasets. The reason for
selecting the same hyperparameters for each reported experiment
is to study the effect of the number of prototypes on PIP's
accuracy. We optimized the cross-entropy loss using Adam [34]
with base learning rate
= 3e ,3batch
size = 32 , encoding size
= 64 , e 008= .,x 01= ., and updating weight period = 20 .
The number of prototypes selected is equal to the number of
TABLE II Average Likert responses for raw time series data.
STEM
3.1
1.3
2.7
3.7
classes to increase the interpretability. As we show later, a larger
number of prototypes will in some cases increase the accuracy of
the model.
V. Results and Analysis
To validate PIP's performance, we ran a user experiment to
measure the interpretability of models generated by PIP. Moreover,
we assessed the user's trust to employ PIP for their
application. Lastly, we compared PIP's accuracy to other time
series classifiers.
WALKING VS. WALKING UPSTAIRS
WALKING VS. LAYING
MALFUNCTION VS. NO MALFUNCTION
NUMBER-ONE VS. NUMBER-THREE
CLINICIAN
3.4
1.1
2.4
3.7
OTHER
2.0
1.5
3.1
3.8
A. Interpretability of Raw Time Series Data
The results of this experiment are summarized in Table II.
These results reflect that raw time series data (Raw) do not
provide adequate interpretability in most cases, as shown in
Figure 7. Participants were asked to differentiate between multiple
class pairs. These include walking vs. walking upstairs
(UCI-HAR), walking vs. laying (UCI-HAR), malfunction vs.
no malfunction (UCR-FordA) and
number-one vs. number-three (UEAArabic).
The results reveal that only in
cases such as walking vs. laying, where
the difference between two signals is
noticeable, the overall averaged participant
response is close to Extremely Easy
(1). This contrasts with the other pairs,
where the overall averaged participant
OVERALL
3.0
1.2
2.7
3.7
40 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | FEBRUARY 2022
IEEE Computational Intelligence Magazine - February 2022

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - February 2022
