conclusion.tex

\section{Conclusion}

We conclude that the \hoeffdingtree, the \mondrianforest, and the
\naivebayes data stream classifiers have an overall superiority over the
\FNN and the \mcnns ones for \har.  However, the prediction performance
remains quite low compared to an offline \knn classifier, and it varies
substantially between datasets. Noticeably, the \hoeffdingtree and the
\mcnns classifiers are more resilient to concept drift that the other ones.

Regarding memory consumption, only the \mcnn and \naivebayes classifiers
were found to have a negligible memory footprint, in the order of a few
kilobytes, which is compatible with connected objects. Conversely, the
memory consumed by a \mondrianforest, a \FNN or a \hoeffdingtree is in the
order of 100~kB, which would only be available on some connected objects.
In addition, the classification performance of a \mondrianforest is
strongly modulated by the amount of memory allocated. With enough memory, a
\mondrianforest is likely to match or exceed the performance of the
\hoeffdingtree and \naivebayes classifiers.

The amount of energy consumed by classifiers is mostly impacted by their
runtime, as all power consumptions were found comparable. The
\hoeffdingtree and \mondrianforest are substantially slower than the other
classifiers, with runtimes in the order of 0.35~ms/element, a performance not compatible 
with common sampling frequencies of wearable sensors. 

\revision{
Our results show that even though data stream classifiers were developed
to minimize memory footprint, they were not developed to work with a given
memory budget. For instance, there is no guarantee that the \mondrianforest or
the \hoeffdingtree would make the best split selection on a device with a
small amount of memory. Future research could focus on developing algorithms that
can guarantee a peak memory consumption and ensure an optimal response within a
memory budget.  In particular, we are planning to explore tree sampling methods
based on memory and performance criteria in \mondrianforest.  In
addition, deploying classification algorithms on actual connected objects might
highlight other relevant research directions.
}

\section*{Acknowledgement}
This work was funded by a Strategic Project Grant of the Natural Sciences
and Engineering Research Council of Canada. The computing platform was
obtained with funding from the Canada Foundation for Innovation.

% vim: tw=80 ts=2