Figure 3. Classification accuracy plot providing an overview of
the individual (grey bars) and cumulative (red line and circles)
contribution of each feature (in the in which they were selected in the
stepwise forward selection (SFS) process).
2.5 Feature
visualization
Above, under “Feature selection” we already mentioned the three
objectives with feature selection: improving interpretability, reducing
overfitting, reducing computational requirements. Visualisation of the
features can further assist in deciding on the features to use in the
ultimate behaviour classification model, yet its main use is in deciding
if any behaviour types should be combined to ultimately improve
behaviour classification performance. Alternatively, the visualisation
may also lead to considering splitting up existing behaviour types into
multiple behaviours. In other words, this visualisation aids in
evaluating the behaviour set.
The rabc package offers three ways to visualize features. The first two
visualise the features in isolation whereas the third is an integrative
approach where entire feature domains are analysed collectively. The
first of the visualisation methods, plot_feature, draws individual
values of features ordered by behaviour (Fig. 4). The second,
plot_grouped_feature, produces a boxplot of a selected feature for all
behaviour types, as demonstrated for the ODBA feature in Fig. 5. In the
case of the White Stork dataset it suggests clear differentiation of
behaviours by ODBA with a trend of ODBA decreasing going from active
flight via walking to passive flight, standing and sitting. The third
and most important, integrative approach uses Uniform Manifold
Approximation and Projection (UMAP).