IBM SPSS 建模

5 Modeling

As the given problem is to deal with forecasting on customer campaign feedback, the following 3

modeling tools are selected to create predictive model: Neuron Network , C5.0 and CART. During the

different stages of CRISP-DM process, a number of predictive models have been created and evaluated.

They will be discussed in detail in the following paragraphs:

Stage1 :

During this stage,

correlations

of all the fields were exhaustively analyzed. Those pares of fields which

has an absolute correlation of greater than 0.9 will be separated, and one member of such pare must be

eliminated. Besides that,

heuristic

also were used to do fields

elimination, eg: “For Future Tax Filer Use”

Stage 2:

After fields elimination finished,

anomaly

model

were used to eliminate potential outliers. We filter out

those outliers before model creation.

Stage 3:

To further improve model accuracy, we attempt to incorporate clustering methods for Census and Tax

filer data. To achieve that, census data were filtered and clustered. Then cluster number of each record

was appended back to original table. Same thing happened for Tax filer.

Model Evaluation

As the objective is to increase the response rate, model accuracy is not a good indicator for model

fitness.

Assume that company has a very large customer database. After the creation of predictive model,

campaign will only be held among those predicted responders. Under this assumption, fitness of the

model can be measured by the formula below:

Assume: A = Predicted responders B = Actual responders

Model Fitness = Count(A

∩

B) / Count(A)

We apply it to the models created in different stages of CRISP-DM process, the Model fitness table can

be created: