Approaching segmentation and its challenges

Back to homepage

Interpretation and its ambiguity

On the face of it, segmentation is simplistic in concept – divide the marketplace and utilise the differences between groups to meet a specific business objective. However, as simple as this may seem on the surface the possible techniques that underpin this process are much more open to the user than ever. With a whole array of different machine learning algorithms to utilise and developments in clustering techniques constantly evolving – the notion of ‘how to segment’ can be interpreted very differently depending on the researcher/analyst. This is not only due to the ever expanding data scientist’s toolbox but also due to the different methodological approaches that can be adopted. Raising the question which method is best to adopt and how do we decide on the underlying technique used to carry out a segmentation analysis?

Getting closer to the objective is key

Historically segmentation was largely based on demographic data using variables such as region, gender and socio-economic grade, this approach suffers from the over-generalisation of large sub-groups of the population and as a result segmentation has shifted towards methodologies based on attitudinal and behavioural data. The advantage is these behavioural traits are directly actionable and can be incorporated into a model that separates consumers on more utilisable data, informing how to target consumers efficiently.

However if we want to understand how to drive sales based on product strategy, we must ensure that the entire segmentation process is tailored specifically to this core business objective - rather than defaulting to purely attitudinal or demographic segmentations. In this example making sure that the segmentation is based on product needs to understand what consumers want and why they purchase products is vital, using this information can then be utilised to tailor offers or product features to improve sales.

Rise of machine learning

There are many machine learning algorithms available; broadly there are two fundamentally different approaches. The first is hard clustering which assigns each consumer to one segment only, the alternative is soft clustering where consumers can be allocated to multiple segments, with a probability assigned for membership of each segment.

The advantage of soft-clustering is its flexibility – as this technique is able to capture different elements of a consumer’s personality. A consumer’s behaviour can change depending on factors such as mood or time of day affecting how consumers make ‘in the moment’ decisions. This technique is therefore well-suited for products that are bought on a regular basis as the models flexibility captures fluidity present in real-world decision making.

When considering infrequent high-value purchases for example buying a car, an algorithm that provides hard cluster assignment can provide a more reflective model of a consumer’s behaviour. The binary, infrequent nature of these purchases means buying behaviour is more consistent making hard clustering highly predictive in this scenario.

Final remark

All these techniques have their place in segmentation but what is imperative when designing a project is considering how a sub-set of the myriad different methodologies and techniques available can combine to create the most effective solution. With such variety in set-up what seems like a simple concept in principle, can under the bonnet quickly become a very complex and intricate task.

Get in touch

Steven Pesarra
Statistical and Insight Analyst

t: 01434 611160
f: 01434 611161
e: Steven Pesarra