Top 10 Machine Learning Gotchas
summary
Uses a 1,000‑row UC Irvine German credit set with 17 features (≈70% repay) and trains a C5.0 decision tree that yields 13.9% error on holdout, then ensembles trees to a boosted ensemble → 3.4% error. Demonstrates tools and frameworks: R and Python examples, scikit‑learn style APIs, and warns that deep learning (TensorFlow/CNTK/Theano/Torch) needs large data and can be fooled — 0.7% added noise flips a panda prediction from 57.7% to 99.3% gibbon. Lists concrete gotchas: information leakage, resampling bias, collinearity, cherry‑picking, parameter tweaking, and human‑loop overfitting.
AI-written summary · about
chapters (8)
- 0:00Introduction and welcome
- 1:27What is machine learning
- 6:27Live demo: credit risk model
- 14:56What is overfitting
- 18:28Top 10 machine learning gotchas
- 31:59Deep learning and overfitting
- 40:16Kaggle and model accuracy
- 43:28Resources and closing
related talks
30mMachine Learning: Beyond the Tutorial
45m"Breaking black-box machine learning"
32mThe Dangers of Mindless Forecasting
48mCommon Data Science Mistakes
42mMachine Learning Crash Course
01h 33mUnderstanding and diagnosing your machine-learning models
52mMachine Learning Crash Course
39mMachine Learning in Financial Credit Risk Assessment