Katkat
Back

Model y from x and interpret distributions

RedditMachine Learning Engineer · Onsite
MediumMachine Learning EngineerOnsiteMachine Learning

Last updated

This question evaluates modeling judgment in machine learning, covering problem framing (regression versus classification), baseline model and metric selection, interpretation of feature or class distribution differences, and cold-start strategies for users, items, and regions.

Scenario

You are given a dataset with one input feature x and a target y. The interviewer asks: “How would you model this?”

Later, you are shown a plot with two distributions (e.g., distribution of a feature for two groups/classes, or train vs. production) and asked to interpret what it implies.

Finally, you are asked several cold-start questions.

Tasks

  1. Explain how you decide whether this is regression vs classification, what baseline models you try first, and what evaluation metrics you use.
  2. Given a plot with two distributions, explain how you would:
  • Describe what you see (separation/overlap, shift, variance, multimodality)
  • Diagnose potential issues (label leakage, covariate shift, class imbalance, thresholding)
  • Decide next steps (feature engineering, calibration, sampling, monitoring)
  1. Describe practical cold start strategies for:
  • New users
  • New items (videos)
  • New regions/languages

Assume you care about both predictive quality and production robustness.