Paths Subjects Questions Quizzes Pricing Search
Advanced Open Pro

min_child_samples as a Regularizer

You are training a LightGBM model on a dataset with 500,000 rows. A colleague sets min_child_samples=1 arguing: "More specific splits give lower training error, which is always better."

  1. Explain why min_child_samples=1 is dangerous.
  2. What does this parameter physically prevent in the tree structure?
  3. How would you choose an appropriate value for a 500,000-row dataset?
  4. How does min_child_samples interact with num_leaves?

Share this question

← Back to LightGBM practice

We use cookies for product analytics to improve OmniAtlas. See our Privacy Policy.