Linear SVM Models for Online Activity Recognition - Smartphone-Based Human Activity Recognition - page 104

Table 6.7 Accuracy, feature selection and grouping ability for the different approaches with D 2 T
Method

Algorithm % Accuracy % ˁ % ˁ % ˃

L1 SVM SMLP 96 . 61 61 . 76 19 . 24 0 . 0

L1-L2 SVM ʻ = 0 . 001 EX-SMO 96 . 78 61 . 40 19 . 42 0 . 0

L1-L2 SVM ʻ = 0 . 005 EX-SMO 96 . 74 62 . 87 20 . 34 0 . 0

L1-L2 SVM ʻ = 0 . 01 EX-SMO 96 . 81 66 . 18 21 . 38 10 . 8

L1-L2 SVM ʻ = 0 . 05 EX-SMO 96 . 91 72 . 06 24 . 88 20 . 4

L1-L2 SVM ʻ = 0 . 1 EX-SMO 96 . 74 74 . 63 27 . 45 60 . 5

L1-L2 SVM ʻ = 0 . 5 EX-SMO 96 . 71 91 . 54 39 . 58 90 . 6

L1-L2 SVM ʻ = 1 EX-SMO 96 . 30 100 . 00 91 . 67 100 . 0

L2 SVM SMO 96 . 30 100 . 00 91 . 67 100 . 0

SVMs with (very) small values of are preferable, but this could not be, in gen-

eral, the best choice. The advantage of a very flexible solver that copes with L1,
L1-L2 and L2 SVMs, as the one presented in this chapter, consists in the possibil-
ity of identifying the best application-dependent trade-off between performance and
dimensionality reduction, at the expense of a very small implementation effort.

ʻ

6.5 Summary

In this chapter, we showed the benefits of adding gyroscope information into a HAR
system based on smartphone technology. We verified that a set of common BAs can
be accurately classified when this sensor is used along with the accelerometer. We
also explored four SVM algorithms including linear (L1-SVM, L1-L2-SVM and the
conventional L2-SVM) and non-linear (GK-SVM) approaches on the

D 2 dataset.
We found similar performance between them in terms of classification accuracy, but
our selection criterion was subject to prediction speed and the possibility of applying
them in devices with limited resources to provide less computational complexity and
energy consumption.
Linear approaches exhibited the best trade off between accuracy and prediction
speed, conferring distinctive benefits to the MC-L1-SVM, which provides itself a
reduction of the effective number of features needed for the prediction of BAs. Fur-
thermore, the study between different feature domains lead us to disregard frequency
domain features as they were not only marginally contributing to the recognition per-
formance but also adding expensive computations for their estimation. The ideal set
of features selected for our application was the AGT, which only takes into account
time domain features from the accelerometer and the gyroscope. Additionally, we
showed in experiments with theMC-L1-L2-SVMalgorithm, which combines L1 and
L2 norms, that it is provided with interesting characteristics that can be exploited in
order to fine tune different aspects of the learned model such as accuracy, dimension-
ality reduction and grouping ability that can be convenient in different applications.