Linear SVM Models for Online Activity Recognition - Smartphone-Based Human Activity Recognition - page 98

Table 6.2 Experiments with different feature subsets and conventional linear SVM models
Acc Gyro Time Freq Feature groups N. Features MC-L2-SVM (%) MC-L1-SVM (%)
0 1 1 0 GT 108 78.02 -

0 1 1 1 GTF 213 81.04 -

1 0 1 0 AT 164 90.62 -

1 0 1 1 ATF 348 91.23 -

1 1 1 0 AGT ( D 2 T ) 272 96.30 96.61

1 1 1 1 AGTF ( D 2 ) 561 96.54 96.17

linear SVM model (MC-L2-SVM), trained on the corresponding subset of features.
Table 6.2 presents the results.
These results suggest that the more relevant features are added into the system, the
higher classification accuracy is achieved. The whole set of features (AGTF) provides
the best performance. However, from this analysis, it is also noticeable that frequency-
related inputs are only producing a very small improvement in this application as they
do not largely affect recognition performance when compared with the AGT subset.
This indicates they are not strictly necessary for the classification as we already have
a significant number of features in the time domain. Moreover, frequency domain
features requires an extra effort for their derivation such as the calculation of the FFT
for each window sample (Sect. 4.3.2 ) . The accuracy difference between the AGT and
AGTF subsets is only 0.24% which is a negligible value.
Results also allow to gather some evidence of the benefits that the incorporation
of gyroscope signals brings into the HAR system that counterbalance the limited
slowdown in prediction due to the presence of these extra features. It can be observed,
for instance, the visible accuracy difference between the ATF and AGTF feature
groups which diverge only on the use of the gyroscope (5.31%). It is also noticeable,
on the other hand, that themodels trainedwith sets using only gyroscope features (GT,
GTF) have a lower performance. This suggests that the use of the gyroscope by its
own is not appropriate for its application in HAR, despite enhancing the recognition
when exploited concurrently with accelerometers.

6.4.3 Dimensionality Reduction with L1-SVM

In this sectionwe explore the use of theMC-L1-SVMinstead of the conventionalMC-
L2-SVM which does not perform any dimensionality reduction (Khan et al. 2012 ).
This fact is desirable in some practical applications to highlight relevant features as
well as to reduce the computational burden of performing the classification of new
samples. Table 6.3 presents the confusion matrices resulting from the classification
of the D 2 T datasets with this SVM algorithm. Moreover, in Table 6.4 we

display the comparison of MC-L2-SVMandMC-L1-SVM, both in terms of accuracy

D 2 and