Google·Data Scientist·Technical Phone Screen·Senior

Senior

Jun 2026

Summary

Google Data Scientist interview with a meaty ML modeling question centered on customer retention for a B2B product. One question but it had a lot of layers, probably 30+ minutes of back and forth on it.

Questions Asked (1)

Walk through how you'd build a model to predict whether an enterprise customer will renew their contract. What features would you engineer, when would logistic regression be enough versus something like gradient boosting or a neural net, and how would you measure whether the model is actually working?

Product Analytics & MetricsTechnical Trade-offsData Modeling

Author's notes

I started with feature engineering which felt like the right move, talked about usage signals, contract age, support ticket volume, things like that.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer around the end-to-end ML lifecycle: problem framing, data and feature engineering, model selection with trade-offs, and evaluation. Emphasize that the choice of model depends on data size, interpretability needs, and performance requirements, and that offline metrics must connect to business impact. Conclude with a monitoring plan to ensure the model remains effective over time.

Pro tip: Frame the problem as a business decision first: quantify the cost of false positives (wasted retention offers) versus false negatives (lost revenue) to justify your metric and threshold choices. This shows you think like a product-minded data scientist, not just a modeler.

1. Define the problem and success metrics

Clarify the prediction target (renewal vs. churn), the prediction horizon, and how the model will be used (e.g., prioritize retention outreach). Define business success metrics (e.g., retained revenue, lift over baseline) and translate them into ML metrics (e.g., precision@k, recall, AUC).

2. Data collection and feature engineering

Identify data sources: contract details, product usage, support tickets, billing history, firmographics, and customer interactions. Engineer features like usage trends, engagement scores, support sentiment, contract value, tenure, and seasonality. Handle missing data and ensure features are available at prediction time.

3. Model selection and trade-offs

Start with logistic regression for interpretability and speed, especially with limited data or when explainability is required. Move to gradient boosting (e.g., XGBoost) for better performance with complex interactions and larger datasets. Consider neural nets only for very large datasets with unstructured data (e.g., text) and when interpretability is less critical.

4. Evaluation and validation

Use time-based splits to mimic real-world deployment. Evaluate with metrics like AUC, precision-recall curves, and calibration. Perform cost-sensitive analysis to choose thresholds. Validate with holdout data and compare against a baseline (e.g., rule-based or logistic regression).

5. Deployment and monitoring

Deploy the model via an API or batch scoring, integrate into CRM or retention workflows. Monitor performance over time (data drift, concept drift) and set up alerts for degradation. Plan for periodic retraining and A/B testing to measure business impact.

Key Points to Mention

Feature engineering: usage trends, engagement metrics, support interactions, contract value, tenure, and firmographics.
Model trade-offs: logistic regression for interpretability and small data; gradient boosting for performance with complex interactions; neural nets for large-scale unstructured data.
Evaluation metrics: AUC, precision@k, recall, F1, and calibration; use time-based validation to avoid leakage.
Business alignment: cost-sensitive threshold selection and measuring lift over baseline in terms of retained revenue.
Monitoring: track data drift, concept drift, and model performance; set up retraining pipelines.
Interpretability: use SHAP or coefficients to explain predictions to stakeholders and build trust.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.