Classifying Sales Professionals on LinkedIn
You are building a machine-learning system that automatically classifies LinkedIn members who are likely to be sales professionals. The classifier will power product features such as Sales Solutions targeting, onboarding flows, content recommendations, and analytics.
Constraints & Assumptions
-
Define "sales professional" clearly before labeling.
-
Use only permissible data and respect privacy and regional policies.
-
Handle class imbalance and threshold tuning by use case.
-
Include model evaluation, deployment, and monitoring.
Clarifying Questions to Ask
-
Which roles count as sales: SDR, BDR, account executive, sales manager, sales operations, business development, or account management?
-
Is precision or recall more important for the product use case?
-
What profile, behavior, graph, and product-interaction data is allowed?
-
How will predictions be used and explained?
Part 1 - Features and Data Sources
Which features would you engineer and what data sources would you use?
What This Part Should Cover
-
Include profile text, title, headline, skills, industry, company, seniority, groups, endorsements, network graph, content interactions, product usage, and job history.
-
Use text embeddings or keyword features for titles and descriptions.
-
Include graph and behavioral signals with privacy safeguards.
-
Avoid leakage from the target product if it creates circular labels.
Part 2 - Labels and Imbalance
How would you create ground truth and handle class imbalance?
What This Part Should Cover
-
Use self-declared roles, profile labels, manual annotation, Sales Solutions usage, job titles, or high-confidence heuristics.
-
Audit label noise and ambiguous roles.
-
Use stratified sampling, class weights, threshold tuning, PR-AUC, precision/recall, and calibration.
-
Evaluate by segment and use case.
Part 3 - Modeling and Deployment
What modeling approaches would you start with, and how would you evaluate and deploy?
What This Part Should Cover
-
Start with interpretable baselines, then compare tree models, linear text models, embeddings, graph features, or neural models.
-
Use offline validation, holdout sets, human review, and online product metrics.
-
Monitor drift, fairness, privacy, and feedback loops.
-
Define thresholds for targeting versus analytics.
Follow-up Questions
-
How would you reduce false positives among recruiters or marketers?
-
How would you explain the classifier to product partners?
-
How would you update labels as members change roles?