ZEMCH 2015 - International Conference Proceedings | Page 253
Figure 7: TA approach with timeslice variations from 30 seconds to 10 minutes.
Dataset1 House C.
This might seem contradictory with the fact that the longer the timeslice, the more information
about real sensors readings is lost. However, as we are working with unbalanced datasets, there is
a chance that the samples which are going to be lost earlier (as we increment timeslice duration),
are those labelled as the less frequent activities. Consequently, fewer classes are to be classified
and the models perform better with less infrequent states.
Regarding overall model performance, the experimental results show that the SVM approach
yields the most accurate performance among the three models evaluated, reaching values above
80% accuracy in two out of three scenarios. We tried different kernels using several values for
the kernel parameter, following the suggestions given by the libsvm developer (Chang and Lin,
2011). However, when introducing value variations within reasonable limits, the results do not
present large variations. The overall SVM best parameters were the RFB as kernel with gamma=1/
num_features.
For House A and B the results are over the 80% of accuracy. However, results from House C were
significantly less accurate. All of our experiments, in addition to the results reported by this dataset team, are always under 50% of accuracy in House C. The poor accuracies suggest that this data
does not contain representative enough information, or there could have been issues with labelling or sensor deployment. Another different factor from this House C scenario is that it contains
a larger area from all three, thus it has the less sensor density and it performs up to 16 activities (3
more that House B and 6 more than House A.) All this factors seem to have had a negative impact
on the performance of the classifiers for that scenario as all the results for the other scenarios are
much more consistent.
Moreover, results show that the multiclass SVM-based approach outperforms other approaches.
As we can see in Table III, SVM outperforms HMM and kNN significantly in all scenarios and SVM
results are better than all the previous generative models proposed (NB, HMM and HSMM); and
also above the discriminative one (CRF) in House B scenario. The averaged accuracy for the three
scenarios shows that our SVM approach outperforms all the previous models proposed, with a
71.64%.
Modelling occupant activity patterns for energy saving in buidings
251