Demo for Non-Parametric Classification

1
Demo for Non-Parametric Classification

• Euclidean Metric Classifier with Data Clustering

2
Data Classification by Similarity

• The similarities among the data is the basis of
  this type of classification
• Similar data is classified together
• Similar term in the mathematical sense, it must
  be mathematically defined
• In the metric Space
• Euclidean Distance, Manhattan distance, etc
• Nearest-Neighbor approach

3
Classification Method

• Steps
• A priori information. Classified Observations
• Model Reduction to reduce computation.
• Use of nearest-neighbor approach.
• A cluster point represents a group of neighbor
  data points. Voronoid Tesselation
• Selection of no. of Cluster centers is
  important.
• The unclassified measurements are evaluation
  against the clusters points.
• K-nearest neighbor rule is applied for Euclidean
  distance and k 1

Training data Class a
Training data Class b
Cluster Partitioning
Cluster Centers a
Cluster Centers b
Unclassified data
Neighbor Evaluation Minimum distance
Data classified as Class a
Data classified as Class b
4
K-means and K-medoid algorithms
5
Feature Space for Training Observations
6
Kmedoid Function

• Syntax
• resultKmedoid(data.X,param.c)
• Function
• The objective function Kmedoid algorithm is to
  partition the data set X into c clusters
• Result
• The calculated cluster center vi (i ? 1,
  2,..c) is the nearest data point to the mean of
  the data points in cluster i.

7
Cluster Partition for Case a
8
Cluster Partition for Case b
9
Cluster Partition all Cases
10
New Observations in the Feature Space
11
New Observations Classified
12
Improvements and Suggestions

• To Validate the cluster perfomance classifying
  the a priori training data
• To test the effect on the clusters perfomance the
  no. of cluster protoypes
• To try classification using the complete training
  data, without Cluster partitioning
• To Increase K in the nearest neighbor selection

13
The End
Thank You!