Using ML to Understand Behaviours
Problem
A social learning platform was experiencing significant variation in levels of user engagement across its client base — and that variation was having a direct impact on the key metrics that evidenced the value each contracting organisation was receiving from their use of the service.
The company had usage data, but it wasn't structured in a way that lent itself to straightforward analysis. The challenge was compounded by the fact that each client organisation had implemented the platform differently, making it difficult to draw meaningful comparisons or identify patterns. On the surface, it appeared almost impossible to determine what was driving the disparity — let alone what needed to change to resolve it.
Solution
Using Google's machine learning tools, I ran a regression analysis across the available data to identify which variables had the strongest relationship with negative engagement behaviours. The goal was to find characteristics that could serve as early indicators of disengagement — signals that appeared upstream of the problem, before engagement had already declined.
Identifying those leading indicators would open up two distinct possibilities: either directly influencing the characteristic itself to shift behaviour in the right direction, or using it as a basis for segmenting the user base and tailoring interventions to the groups most at risk — increasing the likelihood of better outcomes down the line.
Result
Using machine learning analysis, I identified a causal link between a decline in user engagement and the onboarding experience of a specific user segment — a connection that had been obscured by noise in the wider experience data.
This finding gave us the focus needed to run a targeted programme of user interviews, through which we uncovered the root causes behind the behaviours we were observing. The insights from those interviews directly informed a series of changes across the platform — spanning the timing of email prompts and the clarity of onboarding messaging around expected user behaviours.
The process demonstrated how machine learning can surface meaningful relationships within large volumes of unstructured data — relationships that would be difficult to detect manually — and translate them into focused, evidence-based action.