Concept - Dictionary Reduction

Dictionary Reduction (feature reduction)

Also called feature reduction techniques; needed due to the Concept - Curse of Dimensionality (02 Pre-processing s.73). For BoW models:

1. Local dictionary
2. Removing stopwords (Concept - Stopword Removal)
3. Frequent words
4. Feature selection (Concept - Feature Selection; highlighted on slide)
5. Token reduction: stemming and synonyms (Concept - Stemming)
6. Feature transformation: PCA, or topic models

Feature selection by attribute ranking (s.74): form a local dictionary by ranking features by predictive ability for the category (Sports: soccer, football; Travel: airport, cruise) and keeping the top-ranking ones. Scoring approaches: document frequency, Concept - Information Gain (TWM), mutual information, CHI. Flow: score each feature -> sort -> first k features -> learn model.

Related

Concept - Bag of Words | Concept - Sparse Matrix (CSR)

Appears in

• 02 Pre-processing - Chunk 03 (s.61-90)