Databases Reference
In-Depth Information
Open an interactive ANN regression modeler for the trnValHomes dataset
selecting price as the prediction column.
As before, the ANNmodeler uses just the training data to complete one epoch.
At the end of this epoch, R
2
with respect to the training data is computed and then
again with respect to the validation data. Both are displayed above the training
progress plot. At this point after one epoch, the model is very unlikely to be
trained. Given the initial random neuron weights, R
2
is likely to be negative.
Click “Decrease” to reduce the epochs per step to 1.
Click “Slower” about seven or eight times.
The purpose of the above two steps is to slow the training process down,
allowing you, the user, to react in time to changing states in the model. If after
training a few epochs, you feel able to respond fast enough, you may want to
adjust the training speed upward.
The overall objective of the process is to find a training state with the best
possible validationR
2
. The challenge of finding this state ismostly a trial and error
process. The best location and direction of movement on the training grid is
dependent on the attributes and values of the dataset. It will change fromdataset to
dataset. The best strategy is to try moving in different directions on the grid.
Whenever, you get to a point where the training R
2
is steadily increasingwhile the
validation R
2
is decreasing, the ANN is overtrained. Go back to a previous
location and try a different direction. Keep doing this until you feel that you have
found a state that is difficult to improve upon. In testing that we have conducted
using this dataset, we have been able to push the validation R
2
up to about 0.52
while at the same time R
2
for the training set has reached 0.79. Of course, R
2
for
the training set could be pushed even higher at the expense of the validation R
2
.
Model Application
If our objective in building the model is to locate homes for investment purposes
that have a good potential for appreciation, our approach is to use the model to
generate predicted prices for all homes in the initial dataset. Once they are
generated, we can compare actual asking price to predicted price, selecting the
best bargains for additional investigation. Obviously there are characteristics
and conditions of homes not captured by the dataset that would require a
physical inspection to assess. However, using the model is a good starting point
in identification of the best candidates for inspection.
Keep your newly created, best performing ANN regression model open.
Open the previously created SelectedHomes.csv dataset.
