Exposing the risks of customer churn using a connected graph

Australian businesses face an uphill battle and escalating costs when it comes to maintaining customer loyalty.

Customer churn is a persistent threat to growth and revenue in today's competitive landscape.  

The need to replace lost customers is not only costly but also diverts resources from innovation and expansion. 

One of the challenges Australian businesses face is that some of the channels they traditionally relied on for early warning of customer dissatisfaction are starting to run dry. For example, customers have stopped responding to sentiment surveys, instead staying silent about experiences, both good and bad, according to one survey. This has left business leaders "in the dark about changing consumer behaviours or the reasons behind churn, making issues even harder to address."

Another key challenge is the recognition that churn is now a complex and multi-dimensional problem. 

Traditional churn models often rely heavily on static customer attributes and on analysing individual data points and customer profiles in isolation. This paints an incomplete picture of the customer journey.

But churn rarely occurs due to a single bad experience or factor. Instead, it's almost always the result of a series of connected events and relationships, interactions, and influences. 

Hidden relationships between these different data points influence whether a customer will stay or leave. For example, a customer might not churn because of their own issues with a service, but because a close peer in their social network has left. 

The real difficulty today lies in tracking these indirect influences on churn: things like peer behaviour, social connections within a service, or dependencies on other services. 

From a technical perspective, this kind of data is often siloed across different systems, leading to fragmentation and incomplete insights. As a result, interventions are often delayed, reactive, or even misdirected, making them less effective. 

To turn this around, leading businesses are modelling customer behaviour as a connected graph. 

By visualising the relationships between complex data points, businesses are able to access a much richer and more accurate understanding of churn drivers, allowing them to predict who is most in danger of leaving and why, and providing early points of intervention to make things right with the customer before the relationship passes the point of no return.

Five means of analysis

To assemble an accurate picture of customer churn using knowledge graphs, a design pattern for storing, organising, and accessing interrelated data entities, businesses must first ingest data from various sources, such as customer relationship management, billing, and support ticket systems, as well as less-structured data from web and application logs, which can provide insights into sentiment and influence.

Once this diverse data is ingested, it's transformed into nodes and relationships that are visually mapped and represented. Each node represents an entity, such as a customer, a product, a support ticket, or a marketing campaign. Each relationship represents an interaction or connection between these entities. The connected graph becomes a living, breathing network of this valuable information. 

By applying various algorithms to these nodes and relationships, businesses can uncover hidden connections and unlock the predictive power of graph technology. In particular, five classes of algorithms have been shown to be significantly effective in detecting customer behaviour patterns that are indicative of churn or of a deteriorating customer relationship.

First, community detection algorithms are useful to identify group behaviour. For example, a group of customers who frequently interact on a support forum and are experiencing similar product issues; or a collective decline in customer engagement in a certain region after a service change. Both circumstances could increase the risk of a churn event occurring.

Second, similarity algorithms can be used to find customers exhibiting similar browsing patterns or purchase patterns to recently churned customers, enabling the business to proactively intervene before one set of customers leaves, or to better control the service for the next cohort of customers that sign up to it.

Grouping customers based on service usage habits can also reveal otherwise undetectable changes in behaviour. Even if a customer's overall activity is moderate or normal, they may suddenly stop using a particular feature of the service, indicating they're more likely to churn. This is a subtle yet important indicator that may otherwise be overlooked.

Third, topological link prediction can be used to predict which customers are likely to disconnect from a specific service or community based on the weakening of connections between nodes over time. 

Fourth, centrality algorithms such as degree centrality, closeness centrality, between centrality, and PageRank can be used to pinpoint highly influential customers within a social network, or top-ranking support staff, whose actions could trigger churn among their connections. The availability and performance of particular customer support staff, for example, may directly impact customer satisfaction. Staff retention initiatives may be needed to keep the most valuable support staff engaged.

Finally, graph embeddings can be used to highlight predictive patterns you would not think to look for. The signals they surface can be used to move the business forward. For example, a TV streaming service may observe that a customer frequently uses a niche feature such as a watch party plugin. Traditional churn models may not easily connect the usage of this specific feature to overall churn risk, as it's not a direct billing or engagement metric. But using graph embeddings, the streaming provider could identify a subtle behavioural change, where use of that plugin by the subscriber, and customers who would normally join the watch party dip, but overall hours watched remain moderate. Graph embeddings can identify this weakening social connection within the circle of friends and factor it into churn risk accordingly.

Customer churn remains a complex challenge facing businesses today, but connected graphs enable businesses to identify high-risk user cohorts with a higher degree of accuracy.

The opportunity to proactively detect and remediate issues will help businesses retain customers and improve customer lifetime value. This, in turn, ensures the long-term health and profitability of Australian service businesses