Consequently, unlike [4], instead of simply feeding
the input data into a single model, they were first into
K clusters, using K-mean clustering [32], with K set
to 2. Each cluster would be used as training dataset
of separate ANN. Prior to predicting the unseen data;
they would be classified first into one of these clusters,
based on their averaged factors similarity and then
assigned to the respective ANN. With this modification,
it was ensured that a proper model was chosen
to predict the data with similar characteristics. The
results are illustrated in Fig 15. If unseen prediction
was made by the model trained using data with
similar characteristics, the accuracy are 69.53, 75.85
and 69.89 percents, in 2011, 2010 and 2006 respectively.
Unless proper classification and hence model
was assigned, the results was no better than flipping
a coin.