Professional Documents
Culture Documents
Guide To Build Better Predictive Models Using Segmentation
Guide To Build Better Predictive Models Using Segmentation
Guide To Build Better Predictive Models Using Segmentation
Introduction
We use linear or logistic regression technique for developing accurate
models for predicting an outcome of interest. Often, we create
separate models for separate segments. To judge their effectiveness,
we even make use of segmentation methods such as CHAID or CRT.
What is Segmentation ?
I could take you down in the depth of marketing to explain this
concept. But, why to complicate things if they could be made simpler?
Once the product is created, the ball shifts to the marketing team’s
court. As mentioned above, they make use of market segmentation
techniques. This ensures the product is positioned to the right
segment of customers with high propensity to buy.
Techniques of Segmentation
There are two broad set of methodologies for segmentation:
Objective (supervised) and Non-Objective (unsupervised)
segmentation methodologies. As the name indicates, a supervised
methodology requires the objective to be stated as the basis for
segmentation.
Non-Objective Segmentation
If you have thoroughly followed this article till here, we are now good
to delve into the methodology adopted for creating segments. Of
course, with the sole objective of building separate models for each
segments.
Usually, one uses the target (or ‘Y’ known as dependent variable) that
has been identified for model development to undertake an objective
segmentation. Remember, a separate model will be built for each
segment. A segmentation scheme which provides the maximum
difference between the segments with regards to the objective is
usually selected. Below is a simple example of this approach.
The commonly adopted approach would suggest that one should build
separate models for each of the terminal (or end) nodes, which have
been depicted in green in Fig-1. But, is this the best approach from a
modeling perspective? To answer that question, we need to find out a
measure for evaluating a segmentation scheme from a modeling
perspective.
Let us find out why this may be the case. It should be noted
that, when one is developing a linear model, the lift in Adjusted R
Square should be considered instead of lift in Gini.
While building the overall model (Model-1), one can always use
appropriate dummy variables to represent the segmentation. For
instance, one can use the following dummies (it should be noted that
due to degree of freedom constraint there will be one less than all
possible number of dummies)
The predictive power of the model will be even better if one uses
dummies to replicate the segmentation tree
However, this is not the complete explanation behind the low lift in
predictive power. It is also important to consider the set of
segmented models or the child models. Fig-2 provides the list of
variables in the child models.
It can be observed that the variables in each of the five child models
are quite similar, though the relative ordering of the variables is
slightly different. This implies that the drivers of response are similar
across segments. In addition, if one considers the predictive pattern
of a particular variable across segments, one can observe something
even more interesting.
It should also be noted that in this case the information value of the
variables are also similar across segments. In case of a linear model,
partial R square can be used instead of information value. If this is
the case for most of highly predictive variables, then segmentation
would add limited value to the overall predictive power.
Fig-3: Predictive Pattern of the Variable “Number of
purchase in the last 12 months” Across the Five Segments
This will help one to create a scenario where the predictive power of
the segmented models is higher than the predictive power of the
overall model. Fig-4 provides an alternate segmentation scheme for
the problem described earlier.
Therefore the impact of the variable in the overall model is quite very
different as compared to the segment wise impact. In other words, it
means that there is significant interaction effect between the
segmenting variables and the predictor variable “No. of different
categories of purchase made in last 18 months”.
Each tree in this case consists of a few nodes and ensures that it does
not over fit the data. In reality, each of these tree is expected to
capture the interaction effect instead of fitting too closely to the
target at hand. One can relate this philosophy (at a broad level) with
the idea behind creation of segments for developing models, wherein
the objective of the segmentation is not to achieve a closer fit with
the target but to identify interaction effects.
End Notes
In this article, we learnt the following aspects:
1. While performing segmentation for the purpose of model
development, one should strive to achieve a lift in predictive
power and not merely differentiation with regards to the
target of interest.
2. Segmentation capture interaction effects, whereby the
relative importance of the predictors or the predictive
pattern changes across segments.
3. The process of building such segmentations involve
developing child models for each segment and investigating
the variables that are selected in each of the child models.
4. A graphical examination of the highly predictive variables
across segments can also yield valuable insights. Therefore,
the process requires iteration and careful analysis of the
child models. Sole reliance on a CHAID or CRT algorithm
may not yield the appropriate result. One must remember
that the ultimate objective is to enhance the predictive
power of the segmented system.
Did you find the article useful ? Have you used any of these
techniques in your market segmentation process? Do share your
opinions and suggestions in the comments section below.
If you like what you just read & want to continue your
analytics learning, subscribe to our emails, follow us on
twitter or like our facebook page.
You can also read this article on Analytics Vidhya's Android APP
Related Articles
In "Business Analytics"
In "Deep Learning"
Top 10 Pretrained Models to get you Started with Deep Learning (Part
1 - Computer Vision)
In "Deep Learning"