The art of categorical encoding for Tabular data problems
summary
Compares concrete encoders on the Titanic dataset with code and a notebook, using category_encoders and a simple logistic regression pipeline; one‑hot yields ~57% above random, count encoding reduces feature bloat, target and weight‑of‑evidence encodings produce higher training AUC but show strong overfit (training ~0.87 test ~0.81), and cross‑validated target/James‑Stein style smoothing raises test AUC slightly (e.g. 0.810→0.816) while entity embeddings and ordinal/contrast options offer practical alternatives.
AI-written summary · about
chapters (12)
- 0:00Introduction and speaker
- 0:42Importance of data encoding
- 2:57Data types and variables
- 5:36Rules for encoding methods
- 6:44One-hot encoding analysis
- 10:14Vanilla count encoding
- 12:25Target encoding techniques
- 14:33Weight of evidence
- 15:48Advanced encoding methods
- 18:14Best practices in practice
- 20:15Final recommendations
- 21:22Q&A and session closing
related talks
42mHow am I going to deal with all of these cats
29mDirty Data Science Machine Learning On Non Curated Data
28mClassification Through Regression:Unlock the True Potential of Your Labels
32mPreparing Messy Data for Supervised
38mLeverage Reviews Data for Multi Label Topics Classification
35mUsing Embedding Layers to Manage High Cardinality Categorical Data
37mOptimal on Paper, Broken in Reality
32mSemantic Label Representation and Multimodal Categorization