Skip to content
Musab Edriss
← All case studies
MLEngineering

Data foundations that lift model accuracy: Pathfinder SNA and a shared feature store

A subscriber-level social-network and behaviour feature table, plus a model-training feature store in build, so every model starts from richer, consistent signals — lifting the inactivity model's precision.

Role
Function lead — Data Science, AI & Analytics
Where
stc Bahrain
19.6% → 21.3%Inactivity model precision (P@5%)LivePathfinder SNAIn developmentFeature store
SQLPythonSocial network analysisDPI dataAirflow

Some specifics are anonymised to respect commercially sensitive information.

Context

Every model the team builds — inactivity, churn, recommendations, next-best-offer — needs the same kind of inputs: who the customer is, how they use the network, how they pay, and who they are connected to. Each project was assembling those inputs for itself.

Problem

Rebuilding features per use case was slow, inconsistent between models, and left the strongest signals — relationships between customers and their app and web behaviour — mostly unused. Time to market and model quality were both paying for it.

What I did

I set up two shared data foundations.

Pathfinder (SNA 2.0) is a subscriber-level social-network feature table, enriched with app and web behaviour from DPI data. It gives every model relationship and behaviour signals without rebuilding features per use case.

The model-training feature store combines all of our data sources into reusable training assets: customer profile and demographics, usage, billing and payments, complaints and tickets, device and add-on purchases, the Pathfinder network features, SMS/app/web behaviour, and fintech and insurtech data. It feeds the inactivity model, the recommendation engine, next-best-action and next-best-offer, and future models.

Result

Pathfinder is live and already lifting model quality. On the prepaid inactivity model, precision in the top 5% of scored customers rose from 19.6% to 21.3% — about an 8% relative improvement. The feature store is in development, aimed at faster time to market and consistent features across models.

What I learned

A richer feature set can make a model worse before it makes it better. The first run with Pathfinder features actually dropped precision to 14.2%; only after refining the feature set — including restoring a financial-behaviour signal — did it beat the original. Every feature addition has to be measured against the model it is meant to help, not assumed to help.