Problem
Tele2 sells products meant for several people at once: a tariff with several SIM cards where one person pays, and a router that a family or an office can share. Marketing wants to offer them to groups of customers who talk to each other, but the base has no friendship or household labels. The only proxy for sociability was call history, a weak signal because most people moved to messengers and mobile data.
Example from the article: a graph of 3.5 million customers and about 5 million call edges, where the team picks one person per group for a mailing about the multi-SIM tariff.
Solution in brief
Build an undirected graph where customers are vertices and edges are calls weighted by total duration. Run Louvain community detection to split it into groups, re-cluster groups larger than 10, and mark the most talkative customer in each group as its center. Send the offer to the centers.
Questions this case answers
- How to find groups of connected units from their interaction history and choose one representative per group to receive an offer designed for several users
- How to cluster a graph of several million nodes on a single machine with a community detection method
- How to find groups of connected customers from their interaction history
- How to pick the right customer inside a group to receive an offer for a shared-use product
Tools and methods
- Louvain community detection (NetworkX) — weighted graph clustering that maximizes modularity
- Modularity — measure of how much denser the groups are than a random graph with the same degrees
- Iterative re-clustering — rerun Louvain on communities larger than 10 members
- Group features — group size and weight, degree in group and in graph, center flag
- A/B test — mailing to group centers versus a random mailing or no mailing
Google or a one-time email code. No password.
