Geoscience Reference
In-Depth Information
un-matched roads is dependent on the size of candidate road sets and the cross-
effects of other input variables.
For these roads that are in the candidate road set (here, the label is False) but not
the matched roads, the connectivity results into a percentage of 0.465 and 0.535 for
connected and unconnected road segments, respectively. This means the uncon-
nected roads are more possible (0.535 vs. 0.465) to be un-matched road than
matched roads, indicating the connectivity correlation will bene
t to matching the
road correctly.
In order to evaluate the performance of the proposed BBN model, we use the
results according to the indicators of the correctly classi
ed instances (CCI),
incorrectly classi
ed instances (ICI) and Kappa value (Kappa). The ICI, being the
other round of CCI, labels the percentage of incorrectly matched data. The kappa
statistic measures the agreement of prediction with the true class [
14
]. Here, the
value of 0 and 1 signi
es incomplete and complete agreement, respectively.
A higher value of Kappa indicates a better performance of the model.
Table
2
presents the details of the prediction accuracy. It is found that the
accuracy of correctly classi
ed instances 87.02 %. This means the BBN model has
a good performance in matching the GPS data. The level of Kappa for BBN model
is 0.577, which is also satis
ed in this context.
Table
3
shows the results of hit ratio for the two classes of the output variable in
the BBN model. Hit ratio is a measure of business performance traditionally
associated with sales. It is normally a matrix which includes all the combinations of
different classes in the prediction result. Here, the hit ratio shows how accurate a
road was matched. The higher is the value, the more accurate are the matched
results. Since we set the variable into 2 levels, one is the right road segment, and the
other is the found road segments but not the right road. We found for both classes,
the accuracy percentages are 56.19 and 95.64 %, respectively. These ratios are
calculated based on the whole candidate road segments, which means that the more
number of candidate road, the lower of the level of con
rmed ratio, because only
one road among the candidate roads can be labeled as the right road. This is why the
ratio of con
rmed no, because the number of
unmatched roads in the candidate road set is large.
rmed yes is lower than that of the con
Table 2 Prediction accuracy
and model performance
CCI
ICI
Kappa
All samples
87.018 %
12.983 %
0.577
Table 3 Results of the hit
ratio
Con
rmed yes
Con
rmed no
Hit ratio
56.19 %
95.64 %
Search WWH ::
Custom Search