Problem
LinkedIn sales reps had to dig through offline data to find which accounts would renew, grow or leave, and which products to offer. An ML score per account helped, but focus groups showed that a bare score was not enough. Reps wanted to know why account A scores higher than account B and to check the reasons against their own knowledge. Standard feature-importance lists did not help: feature names are technical, and related features are not grouped.
Example from the article: a model predicts the upsell of job slots for an account, and the rep sees "Viewers per job changed from 200 to 300 (+50%) in the last month" next to the score instead of a raw feature list.
Solution in brief
Train XGBoost regressors for upsell and churn with overlapping label windows and monthly rescoring. CrystalCandle then computes per-account SHAP/LIME importances, groups features into a hierarchy, fills narrative templates and shows the top few deduplicated sentences in the CRM.
Questions this case answers
- How to build labels for a propensity model when the outcome can change several times inside one contract cycle
- How to make a tabular model's scores trusted and actionable for non-technical operators who check them against domain knowledge
- How to convert per-sample feature importances into a few readable, non-redundant sentences at scale
- How to predict which existing B2B accounts will upsell or churn at renewal, when labels are blurred by mid-cycle add-ons?
Tools and methods
- XGBoost regressors — predict whether and by how much an account upsells or churns at the next renewal.
- SHAP / LIME / K-LIME — per-account feature importance inside CrystalCandle.
- Feature hierarchy (feature, super feature, ultra feature, category) — groups features so one narrative covers each super feature.
- Narrative templates — sentences with blanks filled from feature values.
- Apache Spark — runs the explainer at scale.
Google or a one-time email code. No password.