Katkat
Back

Design a Skills inference system

LinkedInMachine Learning Engineer · Onsite
MediumMachine Learning EngineerOnsiteML System Design

Last updated

This question assesses the ability to build an end-to-end ML pipeline for extracting and normalizing professional skills from unstructured user-generated text, such as job descriptions or member profiles.

Problem Statement

Design a system to infer professional skills for LinkedIn members based on their profile data (e.g., job titles, summary, work experience). A member's profile is often filled with free-form text, abbreviations, or missing information. Your goal is to map these unstructured inputs to a standardized, canonical skill taxonomy (e.g., "Python" vs. "PyData" vs. "Scripting").

Requirements

  1. Functional Requirements:

    • Extract potential skill candidates from raw text.
    • Normalize extracted entities to a canonical skill taxonomy.
    • Provide confidence scores for each inferred skill.

  2. Non-Functional Requirements:

    • Scalability: Process millions of profiles daily.
    • Latency: Support both batch processing for offline recommendations and low-latency lookups for real-time profile completion.
    • Accuracy: Handle synonymy (e.g., "ML" vs "Machine Learning") and ambiguity (e.g., "Java" as a language vs. "Java" as a location).

What to Cover

  • Data Pipeline: How do you clean input text and handle noise?
  • Model Selection: Discuss candidates for entity extraction and mapping (e.g., CRF, NER, BERT-based embeddings, or Graph Neural Networks).
  • Taxonomy Management: How do you handle new, emerging skills that don't exist in the current taxonomy?
  • Serving Architecture: How do you store the mapping and serve predictions efficiently?
  • Monitoring: How do you measure model drift or quality degradation over time?

Hints

Hint 1

Consider using an embedding-based approach (e.g., Sentence-BERT or FastText) to compute semantic similarity between extracted phrases and the canonical skill taxonomy.

Hint 2

Think about building a heterogeneous graph where nodes represent skills, job titles, and companies to leverage co-occurrence patterns for better inference.

Hint 3

Implement a feedback loop where user edits to their inferred skills serve as noisy labels to retrain the model.

Follow-up questions

  • How would you handle cold-start problems for new skills or users with very short profiles?
  • If the taxonomy changes frequently, how would you design your model to remain robust without requiring a full retraining cycle?
  • How would you address potential bias in the model, such as over-inferring skills based on specific industries or demographics?