---
title: "Feature selection vs feature extraction"
subject: "Data Mining"
chapter: 2
key: "semester_7/data_mining:feature-selection-vs-extraction"
source: https://www.bctengineer.com/q/semester_7/data_mining/feature-selection-vs-extraction
site: BCT Engineering
---

# Feature selection vs feature extraction

Data Mining · Chapter 2: Data Pre-processing

- Asked in: year not given
- Video: [Machine Learning - Dimensionality Reduction - Feature Extraction & Selection](https://www.youtube.com/watch?v=AU_hBML2H1c) — Cognitive Class
- Answer checked by Claude Opus 5.5 (high effort) on 2026-10-14
- Page: https://www.bctengineer.com/q/semester_7/data_mining/feature-selection-vs-extraction
- Whole chapter: https://www.bctengineer.com/md/semester_7/data_mining/chapter-2

## Question

Differentiate between feature selection and feature extraction.

## Short answer

Feature selection chooses a subset of original attributes, preserving meaning. Feature extraction creates new attributes by transforming originals, often losing interpretability. Selection is used when interpretability matters; extraction when many correlated features exist.

## Answer

**Feature selection** chooses a subset of the existing attributes and discards the rest. **Feature extraction** builds a smaller set of new attributes by transforming or combining all the original ones.

| Basis | Feature selection | Feature extraction |
| --- | --- | --- |
| Idea | Select the best subset of original features | Create new features from the originals |
| Features after | Same features, fewer in number | New, transformed features |
| Meaning | Original meaning kept; easy to interpret | Meaning often lost (e.g. "PC1") |
| Examples | Information gain, chi-square, forward/backward selection | PCA, SVD, LDA, wavelets, autoencoders |
| Information | Dropped features' information is lost | Information from all features is compressed |
| Cost | Subset search can be costly (wrappers) | Matrix computations (eigenvectors) can be costly |
| Need to measure the dropped features? | No, only the selected ones are needed later | Yes, all original features are needed to compute the new ones |
| Typical use | When interpretability and cost of measuring matter | When many correlated features exist; visualisation |

*Example*: from attributes {height, weight, BMI, name}, selection may keep {BMI}; extraction (PCA) would combine height and weight into one component.

## Check yourself

The answer's key points as multiple-choice questions, with the correct option and why.

1. Which is the main difference?
   A. Selection creates new features; extraction keeps originals
   B. Selection keeps a subset of original features; extraction creates new transformed ones
   C. Both create new features
   D. Both keep exactly the same features
   Answer: B. Selection discards some originals; extraction combines all of them. (Key point: Core difference)

2. Which method generally keeps the original meaning?
   A. Feature extraction
   B. Both equally
   C. Feature selection
   D. Neither
   Answer: C. Extracted features like PC1 often lose their meaning. (Key point: Interpretability)

3. Which needs all original features to be measured later?
   A. Feature selection, since it uses all
   B. Neither needs it
   C. Both need only selected ones
   D. Feature extraction, since new features are computed from all originals
   Answer: D. With selection only chosen features are needed afterwards. (Key point: Measurement need)

## Related in Data Mining

- [DM vs RDBMS SQL; project failure](https://www.bctengineer.com/question.md?key=semester_7%2Fdata_mining%3Adm-vs-rdbms-failure) — Data Mining, chapter 1
- [Need of clustering; hierarchical vs partitioning; K-means](https://www.bctengineer.com/question.md?key=semester_7%2Fdata_mining%3Acluster-need-hier-vs-partition-kmeans) — Data Mining, chapter 5
- [Descriptive vs predictive data mining](https://www.bctengineer.com/question.md?key=semester_7%2Fdata_mining%3Adm-descriptive-predictive) — Data Mining, chapter 1
- [Data warehouse vs database](https://www.bctengineer.com/question.md?key=semester_7%2Fdata_mining%3Adw-vs-database) — Data Mining, chapter 1
- [OLAP vs OLTP; data cube, operations](https://www.bctengineer.com/question.md?key=semester_7%2Fdata_mining%3Aolap-oltp-data-cube-fact-dimension) — Data Mining, chapter 2
- [OLAP vs OLTP; schema, fact, dimension](https://www.bctengineer.com/question.md?key=semester_7%2Fdata_mining%3Aolap-oltp-schema-fact-dimension) — Data Mining, chapter 2

## Related in other subjects

- [ASIP vs GPP and processor selection](https://www.bctengineer.com/question.md?key=semester_6%2Fembedded_system%3Aasip-vs-gpp-selection-issues) — Embedded System, chapter 2
- [Single vs double earth wire selection](https://www.bctengineer.com/question.md?key=semester_8%2Ftransmission_and_distribution_system_design%3Asingle-double-earth-wire-criteria) — Transmission and Distribution System Design, chapter 3
- [Triac vs diac; BJT vs MOSFET selection](https://www.bctengineer.com/question.md?key=semester_7%2Fpower_electronics%3Atriac-diac-bjt-mosfet-selection) — Power Electronics, chapter 1
- [Bonded and unbonded gauges, selection, installation](https://www.bctengineer.com/question.md?key=semester_4%2Finstrumentation_and_measurement%3Abonded-unbonded-gauges-selection-installation) — Instrumentation and Measurement, chapter 6