KNN

Feature Reduction For K-Nearest Neighbours

More features reduce accuracy. More info != more discriminative power!

This is because, some features are irrelevant, and introduce noise, and fool the algorithm (especially a lazy algo like Knn) Moreover they may have redundant features, as we have limited computational resources.

So we use 2 methods to reduce this:

  1. Feature Selection: F is fiven, find a subset F’ which has elements less than F, such that it optimises cetrain aspects
  2. Feature Extraction: Transforms or projects the existing features to a dimention m < n (original no. of dimentions)

Both cases we try to improve or maintain classification accuracy while simplifying it.

For 1, we have \( 2^n \) possible subsets so finding F’ can use

For evaluation we have 2 methods: Unsupervised (Filter methods) that looks only at the input and outputs the one with most information and Supervised methods (Wrapper methods) that actually uses it on the learning algo and estimates error on validation set.

Strategies

Feature Selection can be univariate (looking at each feature one at a time) or multivariate

Univariate

We then rank the features

Multivariate

Here -ve or +ve val doesnt matter. Only the abs val determine the rank.