Information Technology Reference
In-Depth Information
(K=3). When K=3, the mode of the outcomes of the three retrieved cases is consi-
dered. This was repeated for each similarity combination,
S
TD
,
S
SP
and
S
var
. The num-
ber of correct retrievals and the success rate with respect to the different similarity
combinations,
S
TD
,
S
SP
and
S
var
are shown in
Table 1 for both situations, i.e., when the
target cases consist of known accident cases and when the target cases consist of
known non-accident cases, generated as explained above. The case base consists of
both accident and non-accident cases.
Table 1.
Average number of correct retrievals and success rate of accident and non-accident
cases
S
TD
S
SP
S
var
S
All
K=1 K=3
K=1
K=3
K=1
K=3
K=1
K=3
Target cases = Accident cases
Number of correct retrievals
0
0
25
30
21
24
28
27
Success rate
0
0
64% 77% 54% 62% 72% 69%
Target cases = Non-accident cases
Number of correct retrievals
0
0
25
27
26
26
20
23
0
0
Success rate
78% 84% 81% 81% 63% 72%
Total success rate
0
0 71% 81% 68% 71% 68% 71%
As can be seen from the table, the best retrieval success rate is obtained when the
spatio-temporal, single-point similarity combination
S
SP
is used with K=3. As expected,
computing the similarity based on spatio-temporal attribute differences has the strongest
prediction power for both accident and non-accident cases as target cases. Counter intui-
tively, it was found that the time-day category attribute is not capable of differentiating
between accidents and non-accidents and on its own is not sufficient to retrieve a case
with the same outcome as the test cases. It is possible that the time and day information
is already included in the flow data. For instance, during rush hour we assume that the
flow of vehicles is large and the speed is small. Also, it is difficult to define accurately
the time-day categories, which are currently arbitrarily defined. The success rate im-
proves in general when three cases are retrieved and the mode of the retrieved cases'
outcomes is used as the solution. By retrieving more than one case, we can reduce the
risk of retrieving an uncharacteristic case that happens to have a large similarity with the
target case. The retrieval success rates obtained with
S
SP
, in particular, look promising
and can be considered to be a good starting point for further research.
5
Conclusion and Future Work
The relatively high success rate of 81% shows that the CBR system under develop-
ment is capable of differentiating between accident and non-accident cases based on
