Page 139 - Proceeding of Atrans Young Researcher's Forum 2019_Neat

“Transportation for A Better Life:
Smart Mobility for Now and Then”

23 August 2019, Bangkok, Thailand

2.3.3.1. Confusion Matrix values, N c orrect is the number of samples correctly

Correlation matrix is popularly used to identify predicted, and N total is the total number of samples.
independent variables from the input space of
multivariable problems [57]. This matrix is helpful 3. Results and Discussion
in exploring the linear correlation between each pair
of variables in the input space on which various 3.1. Prediction Capability of AI models
correlation structures in a set of random variables In this section, the prediction capability of EDT
can be deduced and highlighted [58]. Basically, Bagged and SVM are evaluated. Firstly, the
correlation matrix is based on the Pearson linear accuracy of the constructed AI models was
correlation coefficient of two real-valued random investigated via a confusion matrix. A given random
variables U and V as follow [58]: dataset was selected and the predicted results of the
two proposed AI models for the training dataset are
n
  U -U  V -V  shown in Figure 3. It can be seen that in this case,
i
i
r  i=1 both AI models exhibited good performance. The
U,V 2 2 (6) correct responses for the training dataset were 96.6%
n n
  U -U    V -V  and 98.5%, whereas the false responses were 3.4%
i
i
i=1 i=1 and 1.5% for SVM and EDT Bagged algorithms,
respectively. For the testing part (Figure 4), the
where n is the number of values in U and V, U and corresponding accuracy slightly decreased such that
V infer average values of U and V, respectively. the correct responses were 76.1% and 80%, whereas
the false responses are 23.9% and 20% for SVM and
2.3.3.2. Root Mean Square Error, Mean Absolute EDT Bagged algorithms, respectively. More
Error and Accuracy precisely, true positive rate for the training part were
Root Mean Square Error (RMSE) is inferred as 96.2%, 98.0%, 100%, 97.7% and 100% for the
the squared difference error the predicted and actual alternatives 1, 2, 3, 4 and 5, respectively. For EDT
output values whereas Mean Absolute Error (MAE) Bagged algorithm, the true positive rate are 98%,
infers the absolute average differences between 100%, 100%, 100% and 100%, showing excellent
actual and prediction values [59]. Both are well- prediction capability of SVM over EDT bagged
known methods for overestimating with an model. As regard to the testing part, the true positive
acceptable margin of error for predictions of real rate of SVM is confirmed better than EDT Bagged
world problems [60]. Accuracy is well-known (83.5%, 72.7%, 33.3%, 46.2% and 100% compared
metric used to evaluate classification models [61]. with 84.9%, 56.9%, 66.7%, 31.6% and 100% for the
RMSE, MAE and accuracy can be calculated using five alternatives, respectively. The false negative
the following equations [61], [62]: rate from both training and testing part of EDT
Bagged as well as SVM are lower than that of the
n
2
RMSE   (t  0 t p ) / n (7) true positive rate. In the case of alternative 3 (travel
decision = shift to walk) the false negative rates were
i 1
low (e.g. 20% for EDT Bagged and 10% for SVM).
This might be influenced by the hesitation of people
1 n
MAE   t  t (8) on walking under hot weather and low-quality
n i 1 0 p infrastructure, especially for working trip purpose
with long travel distance. With an overall accuracy
N
Accuracy  c orrect (9) regarding the testing part (e.g. 76.1% for EDT
N total Bagged and 80.0% for SVM), it can be concluded

where n means the number of input data, t0 that both AI models exhibited promising ability for
infers the actual values, and tp infers the predicted predicting the travel decisions of transport users.

114