Segmentation analysis is a marketing technique that, based on common characteristics, allows you to split your customers or products into different groups. This in return gives the ability to create tailor-made and relevant advertisement campaigns, products or to optimize overall brand positioning.
Segmentation analysis supports the following:
- To create new products that satisfy the needs of different consumers groups in the market
- To successfully position yourself against your competitors when entering a market or launching a product
- To understand preferences and trends so you can target your audience with the right marketing messages
Why SKIM Segmentation Analysis
The SKIM segmentation solution utilizes advanced analytics and data fusion techniques to allow brands to better identify customers’ evolving preferences. We aggregate multiple data sets, such as survey and client data, for more robust and actionable recommendations.
For a good segmentation there are two important aspects: the characteristics of the data and the algorithms applied to analyze it. SKIM uses a two-stage process to ensure that both elements will be covered properly.
Stage 1: Data Reduction
During the data selection, based on your research objectives, budget and timeline, we align on the characteristics needed to answer your business questions. Additionally, given that the data might come from different sources or tackle different areas, a thorough data pre-processing stage prior to running of any algorithms is done. Some of the checks performed include data recoding, data scaling, correlation plots and variable merge via principal components analysis.
Stage 2: Ensemble segmentation
After we have our final data set, we proceed to apply the clustering algorithms. In some cases, given the nature of the data, some algorithms are not appropriate to identify the patterns of similarity, and in some others, two algorithms can lead to different segmentation results.
For instance, k-means would put consumer A and C in the same segment and B in another segment whereas latent class and hierarchical clustering would put consumer A and B in the same segment and not C.
Because of this, at SKIM, the backbone of our segmentation methodology focuses on capitalizing the strength of many of the available algorithms via a unique ensemble approach.
How does our unique ensemble clustering approach work?
The approach starts by creating micro-segments based only on certain a priori ad-hoc group of variables such as behavioral data, demographics, attitudes, drivers, barriers, etc.
By achieving the most granularity from each group of variables and ensuring that there are no variables with higher weight than others, these micro-segments ensure sample accuracy in the creation of overall segmentation.
In order to create the segments, our ensemble approach combines different techniques such as
- K-means: a method of vector quantization, that aims to partition n observations into k clusters in which each observation belongs to the cluster with the nearest mean, serving as a prototype of the cluster.
- Latent Class: relates a set of observed multivariate variables to a set of latent variables. A class is characterized by a pattern of conditional probabilities that indicate the chance that variables take on certain values.
- Hierarchical Clustering: a method of cluster analysis which seeks to build a hierarchy of clusters.
- Random Forests: A random forest sees prediction as a kind of tree. Instead of a tree splitting into branches, we split the data using attributes. These splits grow from each other, with a first level split and then a second split of that, creating a tree-like structure of sequential data splits. These splits are derived to build the most predictive tree and a random forest builds many of these trees by using random subsets of data and attributes.
The different solutions from the clustering algorithms are combined using a “voting” system. In this sense, each solution “votes” to which segment each observation belongs, assigning the observation based on majority voting. With this approach, we group observations that were consistently put together across different clustering algorithms, ensuring more robust segments.
For example, if consumer A is put into a segment with consumer B most of the time, but not with C, A and B will be put into the same (robust) segment whereas C would be placed in another segment.
What can you expect from Segmentation Analysis?
A good segmentation leads to meaningful, differentiating and actionable segments.
Our segmentation solution can include everything from the size and profile of segments, to identification and tagging tools for your customer database and workshops to socialize with internal stakeholders. To map prospects and current clients into the different segments, one uses Golden Questions and Database matching.
- After defining the segments, it is possible to come with a simplified set of questions in order to classify new observations into the segmentation.
- This special set of questions also known as “golden/magic questions” or “Typing Tool” allows to use the segmentation in future studies
- These questions also give continuity and actionability to the segmentation since it allows to classify respondents on the fly on other research studies
- Sometimes a segmentation study has the (partial) objective to classify a very large customer database. For all customers in that database, it is not possible to ask the “golden questions”. In such a case, database segment matching is necessary.
- This classification is more complex as it requires the use of only the database variables to classify the observations. The optimal approach balances between maximizing the accuracy of the classification with database variables while keeping the survey segments as intact and clean as possible.