Employee Segmentation: Probabilistic vs Rules-based approaches

In the last post I discussed how to conduct an employee segmentation based on needs-based data from employee feedback. In this article I want to explore how employee segments are likely to map to other common concepts such as generations.

Traditional HR segmentation

HR teams have for a long time performed some form of segmentation based usually on one variable - for example using seniority to provide a different compensation mix based on grade.

If we think about how these different differentiated offers are created it’s often by using one variable to split the population.

This type of segmentation might also be done not to provide a better offer, but instead simply to understand the population better. For example you might do analysis to look at an outcome variable (eg attrition) by gender or by age group. If you have a large-enough population you might combine these variables to provide a more granular set of splits - e.g. Women under 30 years old etc.

Mutually Exclusive, Completely Exhaustive

One of the likely characteristics of such an analysis or segmentation is that the groups are likely to be Mutually Exclusive & Completely Exhaustive (MECE). This implies that each employee will fit into one, and only one group.

A MECE based segmentation or analysis enables you to implement rules to allocate each employee to a segment. An employee can’t be in Grade C and Grade E, they can’t be Generation Z and Generation X. 

With a MECE-type segmentation we can use various ‘parts-of-a-whole’ visualisation like a pie chart (or stacked bar chart of proportions). There is no uncertainty which group in these charts they belong to, hence we can choose a visualization that doesn’t need to show uncertainty

Probabilistic segmentation

As an alternative, the needs based segmentation analysis that I described in the last article is probabilistic. It is unlikely that every employee can be allocated without any uncertainty to any single segment. If we are to perform a segmentation - i.e. to allocate them to a single segment - we need to understand which segment they’re most likely to be a member of.

The needs-based segmentation that I described in the last article used the topics that each person mentioned in their comment to create segments. We used a graph-based approach which looked at the cooccurence of themes, or more accurately the probability that the relationships between themes was unusually strong, and then used community detection as a way of creating segments. 

Of course it is possible that somebody’s comments might be represented by more than one segment if they talk about multiple things (it’s the fact that many do talk about multiple topics that enabled us to segment them together). At the level of the employee we are likely to see individuals who’s needs fit more than one of our groups. We could represent these relationships with a weighted graph of cooccurence between each of our communities. Some segments might have a moderate relationship, others will have a very weak one. It’s unlikely that two groups will have a strong relationship otherwise the segmentation algorithm would have merged them when creating the segments.

Describing segmentation membership

A typical need when using needs-based segmentations is to describe a typical profile of who is part of which segment. This is what you’re doing when you’re building personas.

When creating a persona based on a need-based segmentation it’s likely that the needs data will have been collected in a confidential manner, i.e. that the promise that you’ve made to employees is that their comment won’t be explicitly linked to them. Therefore the persona description when mapped to other variables - such as age, education, seniority, business group - is based on the likelihood that somebody with a certain set of needs is in a certain demographic group.

Probabilistic segmentations are likely to overlap

A persona for a needs based segment will describe the themes that somebody in this segment will use and then to describe the individuals it will use a relationship between the other variables. For example a persona about interest in the business might say that this group is 2X more likely than expected to be female, 3X more likely to be university educated and 5X more likely to be engaged.

Thinking about the needs of MECE groups

Another way of thinking about this is to instead think about the needs of a particular group - for example those within a certain ‘Generation’

There are lots of articles and papers about what each generation want. Serious research on generations shows different results depending on the context that the research is done. For example ideas and beliefs about societal issues does appear to be generation related and could be defined by significant events in formative years. However attitudes towards career related topics don’t seem to be generation related but instead could be more accurately described by other factors such as life-stages (e.g. whether the person has children or still lives with their parents).

When we look at generations and map them to needs based segments then we see a distribution of needs segments across each generation. There are likely to be some people of each generation sharing each different needs-based segment however it’ also likely that there might be 2X more GenX employees who fit into needs-group ‘1’ than GenY employees.

When thinking of MECE groups then it’s important that we don’t think of each group as having a certain need but instead think about the chance that a group has a this need or what proportion of a group has that need.

Push versus Pull action

One of the reasons we use segmentation and personas is to improve the individualisation of the employment offer for each employee. We want to increase the probability that we provide an offer or action most likely to be relevant for each employee. When we don’t match their needs either we waste resources (least bad) or annoy them (bad). At the worst case we’ve seen instances where a mismatch can drive employees to leave.

There are two main ways that we can implement a change. Either we can push it onto a population or we can offer it and ask the employee to self-select (pull). In most instances a pull-type solution is more likely to create a bigger benefit from the employee’s perspective. For example defining ‘menus’ of employee benefits based on a needs based segmentation and asking employees to choose one offer will provide a bigger benefit to employees on aggregate than to push them into one group (obviously each menu choice needs to cost the same!)

Sometimes however you need to push changes. If an action from a survey is that managers need to communicate better you can’t ask employees to self-select (and you certainly can’t ask managers as those with the strongest need are likely to be less likely to request support!). In these instances it is appropriate to push implementation to groups with the strongest needs, for example you might want to push resources to the French business but not to the German business. Here needs based segmentation followed by mapping to groups is valuable. Furthermore when a group has been identified (eg employees in France) then we can do further analysis of the comments to identify the specific needs of French employees.

Segment membership evolves

Whilst it’s tempting to think about needs-based segments as fixed, in truth they evolve over time.

Individuals’ membership of the segments will change over time as they evolve as employees and people. For example an employee who becomes a parent for the first time is likely to change segments as their life needs change. If the implementation of manager communication change in France is successful we’d expect less French employees to be allocated to this segment then next time the research is conducted.

Needs based segments also evolve over time, though in many instances this should be a slower process. If you’re creating needs based segments from a survey response it’s worth mapping segment membership over time and re-running the segmentation analysis each time.

We believe that segmentation based on needs is an essential part of any HR approach. Conducting this via the analysis of text feedback is a cost-efficient method which will create significant value for most organizations.