Mining Your Own Business in Retail Using DB2 Intelligent Miner for Data

Chapter 4. How can I characterize my customers from the mix of products that they purchase? 89

4.6.3 What does it all mean? — Mapping out your business

The question we posed was which clustering technique should we use to

describe our customers? The answer is that we use both. In fact we can now

clearly see that we have a way of combining not only the two cluster results but

also our original business rule segmentation into a single view of our business.

The results of course show us what we knew all along, that we don’t really have

distinct groups of customers but a continuum of different types of customer.

However, we can now create through the power of data mining, a completely new

way of visualizing our business. We can now picture not only where our

customers are in relation to segments to which they have been assigned, but

also in relation to each other and our other segments. It is also possible to

position individual customers onto this map, and so we can see exactly where

any specific customer is positioned.

We can also see that by increasing the number of neural clusters that we

generate we have an easy way of increasing the resolution of our picture. The

technique just gives us a better granularity as we increase the number of

clusters. So why do we need to combine the neural and demographic cluster

results? Where the neural clustering technique lays down the map the

demographic clustering technique provides a way of putting the contours onto

the map. Since demographic clustering uses the similarity between our

customers to group them together in an appropriate way, this produces contours

of similarity and gives us the scaling that we can compare our different groups of

customers.

Continuing the map analogy by combining the two types of cluster result we can

identify isolated hills on our map where we have areas of niche customer

behavior that may require special attention in comparison to the large plains

where we have more homogenous customer behavior. The areas between the

different groups are also interesting, particularly where there are valleys and

ridges of opportunity (for example, Clusters 20, 21, 25, and 26 in Figure 4-17).

Customers in these regions are almost certainly the ideal candidates for

cross-selling.

When we map out our customers in this way, we can immediately see the

benefits of data mining to understand our customers better. It gives us a new

view on what they are doing and this immediately prompts us to start asking

more questions about what is going on.

90 Mining Your Own Business in Retail Using DB2 Intelligent Miner for Data

Although we have not shown an example here, in the same way that you map

your business rule segments onto the neural segments, you can do the same

thing for other types of customer classification. As an example, you could

segment your customers in terms of profitability or any other business measure,

and then map this onto the neural clusters and see if there are regions where

particular type of customer behavior gives rise to particular categories of the

business metric. You will have to try this out for yourselves to see what we mean.

Data derived segments from transaction level data

In this section we have concentrated on segmenting customers who we can

identify using loyalty card information and for whom we can aggregate

transactions to produce the CLA data model. The main reason for doing this was

to show how to map the predefined business rule segments onto the data

derived segments. If you only have point of sale transaction level data, but no

means of performing customer level aggregation, then can you still perform data

derived segmentation?

If you perform clustering using the TLA model, then you will also produce a

segmentation that reflects the characteristics of your customers in terms of the

combinations of products they purchased during a single transaction. In this case

the picture is somewhat different from the CLA model results, as shown in

Figure 4-18.