Information Technology Reference
In-Depth Information
•
The target case is an accident case. The retrieved case is an accident case.
•
The target case is a non-accident case. The retrieved case is a non-accident case.
4.1
Experimental Setup
The available test data comprises the traffic flow data and accident data for accidents.
The case base contains 71 historical flow data cases, for which the outcome is known,
i.e., the flow data in a case was either followed by an accident or it was not. Creating
accident cases was relatively straight forward by using the incident data obtained
from the WTTC for the year 2010. Currently, there are 39 accident cases, with suffi-
cient information to correlate them to corresponding flow data. The non-accident
cases had to be artificially generated by selecting for each non accident data from the
existing flow data over the same time
T
and distance
L
. To take into consideration
differences arising from road locations or timing, we have chosen the non-accident
data at the same locations and same date and time at which the accidents considered
occurred but in a different year. In other words, currently the non-accident data con-
sists of the locations and relevant flow data taken from the accident data (2010) but in
a different year (2008). The time
T
and distance
L
prior to an accident (or non-
accident) are empirically chosen to be
T
= 10 minutes and
L
= 1km.
For assessing the prediction capability of each of the possible models, leave-one-
out cross validation (LOOCV) technique was chosen. In LOOCV, one case of the test
set (the target case) is chosen in each iteration, and the remaining cases of the test set
with the exception of the chosen case is used as the case base.
For each target case, the retrieval mechanism is run, the similarity between the tar-
get case and the cases in the case base is computed and the most similar case (K=1) or
the three most similar cases (K=3) are retrieved. For each retrieval, we determine if
the most similar retrieved case (or cases) is of the same type (accident or non-
accident) as the target case. When
K=
3, i.e., the three most similar cases to the target
case are retrieved, the mode of the three retrieved case solutions is used. For instance,
if for a given target case at least two cases out of the three retrieved ones are of the
type 'accident', then the prediction for the target case is 'accident'. Each case in the
case base is consecutively made the target case with the remaining cases constituting
the case base. A retrieval is considered to be successful if the outcome of the retrieved
case or cases, is the same as the known outcome of the target case, i.e. either “acci-
dent” or “non-accident”.
The average success rate of the retrieval mechanism is computed using
100
(9)
where
r
c
is the number of correct retrievals for the target cases and
n
is the total num-
ber of cases in the case base.
4.2
Results
The experiment was carried out first by retrieving the top most similar case to the
target case (K=1) and then by retrieving the three most similar cases to the target case
