Nearest Neighbor Editing and Condensing Techniques

1
Nearest Neighbor Editing and Condensing
Techniques
Organization

• Nearest Neighbor Revisited
• Condensing Techniques
• Proximity Graphs and Decision Boundaries
• Editing Techniques

Last updated Oct. 7, 2005
2
Nearest Neighbour Rule
Non-parametric pattern classification. Consider a
two class problem where each sample consists of
two measurements (x,y).
k 1
For a given query point q, assign the class of
the nearest neighbour.
k 3
Compute the k nearest neighbours and assign the
class by majority vote.
3
Example Digit Recognition

• Yann LeCunn MNIST Digit Recognition
• Handwritten digits
• 28x28 pixel images d 784
• 60,000 training samples
• 10,000 test samples
• Nearest neighbour is competitive

Test Error Rate () Test Error Rate ()
Linear classifier (1-layer NN) 12.0
K-nearest-neighbors, Euclidean 5.0
K-nearest-neighbors, Euclidean, deskewed 2.4
K-NN, Tangent Distance, 16x16 1.1
K-NN, shape context matching 0.67
1000 RBF linear classifier 3.6
SVM deg 4 polynomial 1.1
2-layer NN, 300 hidden units 4.7
2-layer NN, 300 HU, deskewing 1.6
LeNet-5, distortions 0.8
Boosted LeNet-4, distortions 0.7
4
Nearest Neighbour Issues

• Expensive
• To determine the nearest neighbour of a query
  point q, must compute the distance to all N
  training examples
• Pre-sort training examples into fast data
  structures (kd-trees)
• Compute only an approximate distance (LSH)
• Remove redundant data (condensing)
• Storage Requirements
• Must store all training data P
• Remove redundant data (condensing)
• Pre-sorting often increases the storage
  requirements
• High Dimensional Data
• Curse of Dimensionality
• Required amount of training data increases
  exponentially with dimension
• Computational cost also increases dramatically
• Partitioning techniques degrade to linear search
  in high dimension

5
Exact Nearest Neighbour

• Asymptotic error (infinite sample size) is less
  than twice the Bayes classification error
• Requires a lot of training data
• Expensive for high dimensional data (dgt20?)
• O(Nd) complexity for both storage and query time
• N is the number of training examples, d is the
  dimension of each sample
• This can be reduced through dataset
  editing/condensing

6
Decision Regions
Each cell contains one sample, and every location
within the cell is closer to that sample than to
any other sample. A Voronoi diagram divides the
space into such cells.
Every query point will be assigned the
classification of the sample within that cell.
The decision boundary separates the class regions
based on the 1-NN decision rule. Knowledge of
this boundary is sufficient to classify new
points. The boundary itself is rarely computed
many algorithms seek to retain only those points
necessary to generate an identical boundary.
7
Condensing

• Aim is to reduce the number of training samples
• Retain only the samples that are needed to define
  the decision boundary
• This is reminiscent of a Support Vector Machine
• Decision Boundary Consistent a subset whose
  nearest neighbour decision boundary is identical
  to the boundary of the entire training set
• Consistent Set --- the smallest subset of the
  training data that correctly classifies all of
  the original training data
• Minimum Consistent Set smallest consistent set

Original data
Condensed data
Minimum Consistent Set
8
Condensing

• Condensed Nearest Neighbour (CNN) Hart 1968
• Incremental
• Order dependent
• Neither minimal nor decision boundary consistent
• O(n3) for brute-force method
• Can follow up with reduced NN Gates72
• Remove a sample if doing so does not cause any
  incorrect classifications

1. Initialize subset with a single training example
2. Classify all remaining samples using the subset,
   and transfer any incorrectly classified samples
   to the subset
3. Return to 2 until no transfers occurred or the
   subset is full

Produces consistent set
9
Condensing

• Condensed Nearest Neighbour (CNN) Hart 1968
• Incremental
• Order dependent
• Neither minimal nor decision boundary consistent
• O(n3) for brute-force method
• Can follow up with reduced NN Gates72
• Remove a sample if doing so does not cause any
  incorrect classifications

1. Initialize subset with a single training example
2. Classify all remaining samples using the subset,
   and transfer any incorrectly classified samples
   to the subset
3. Return to 2 until no transfers occurred or the
   subset is full

10
Condensing

• Condensed Nearest Neighbour (CNN) Hart 1968
• Incremental
• Order dependent
• Neither minimal nor decision boundary consistent
• O(n3) for brute-force method
• Can follow up with reduced NN Gates72
• Remove a sample if doing so does not cause any
  incorrect classifications

1. Initialize subset with a single training example
2. Classify all remaining samples using the subset,
   and transfer any incorrectly classified samples
   to the subset
3. Return to 2 until no transfers occurred or the
   subset is full

11
Condensing

• Condensed Nearest Neighbour (CNN) Hart 1968
• Incremental
• Order dependent
• Neither minimal nor decision boundary consistent
• O(n3) for brute-force method
• Can follow up with reduced NN Gates72
• Remove a sample if doing so does not cause any
  incorrect classifications

1. Initialize subset with a single training example
2. Classify all remaining samples using the subset,
   and transfer any incorrectly classified samples
   to the subset
3. Return to 2 until no transfers occurred or the
   subset is full

12
Condensing

• Condensed Nearest Neighbour (CNN) Hart 1968
• Incremental
• Order dependent
• Neither minimal nor decision boundary consistent
• O(n3) for brute-force method
• Can follow up with reduced NN Gates72
• Remove a sample if doing so does not cause any
  incorrect classifications

1. Initialize subset with a single training example
2. Classify all remaining samples using the subset,
   and transfer any incorrectly classified samples
   to the subset
3. Return to 2 until no transfers occurred or the
   subset is full

13
Condensing

• Condensed Nearest Neighbour (CNN) Hart 1968
• Incremental
• Order dependent
• Neither minimal nor decision boundary consistent
• O(n3) for brute-force method
• Can follow up with reduced NN Gates72
• Remove a sample if doing so does not cause any
  incorrect classifications

1. Initialize subset with a single training example
2. Classify all remaining samples using the subset,
   and transfer any incorrectly classified samples
   to the subset
3. Return to 2 until no transfers occurred or the
   subset is full

14
Condensing

• Condensed Nearest Neighbour (CNN) Hart 1968
• Incremental
• Order dependent
• Neither minimal nor decision boundary consistent
• O(n3) for brute-force method
• Can follow up with reduced NN Gates72
• Remove a sample if doing so does not cause any
  incorrect classifications

1. Initialize subset with a single training example
2. Classify all remaining samples using the subset,
   and transfer any incorrectly classified samples
   to the subset
3. Return to 2 until no transfers occurred or the
   subset is full

15
Proximity Graphs

• Condensing aims to retain points along the
  decision boundary
• How to identify such points?
• Neighbouring points of different classes
• Proximity graphs provide various definitions of
  neighbour

NNG Nearest Neighbour Graph MST Minimum
Spanning Tree RNG Relative Neighbourhood
Graph GG Gabriel Graph DT Delaunay
Triangulation (neighbours of a 1NN-classifier)
16
Proximity Graphs Delaunay

• The Delaunay Triangulation is the dual of the
  Voronoi diagram
• Three points are each others neighbours if their
  tangent sphere contains no other points
• Voronoi condensing retain those points whose
  neighbours (as defined by the Delaunay
  Triangulation) are of the opposite class
• The decision boundary is identical

• Conservative subset
• Retains extra points
• Expensive to compute in high dimensions

17
Proximity Graphs Gabriel

• The Gabriel graph is a subset of the Delaunay
  Triangulation (some decision boundary might be
  missed)
• Points are neighbours only if their (diametral)
  sphere of influence is empty
• Does not preserve the identical decision
  boundary, but most changes occur outside the
  convex hull of the data points
• Can be computed more efficiently

Green lines denote Tomek links
18
(No Transcript)
19
Not a Gabriel Edge
20
Proximity Graphs RNG

• The Relative Neighbourhood Graph (RNG) is a
  subset of the Gabriel graph
• Two points are neighbours if the lune defined
  by the intersection of their radial spheres is
  empty
• Further reduces the number of neighbours
• Decision boundary changes are often drastic, and
  not guaranteed to be training set consistent

Gabriel edited
RNG edited not consistent
21
Dataset Reduction Editing

• Training data may contain noise, overlapping
  classes
• starting to make assumptions about the underlying
  distributions
• Editing seeks to remove noisy points and produce
  smooth decision boundaries often by retaining
  points far from the decision boundaries
• Results in homogenous clusters of points

22
Wilson Editing

• Wilson 1972
• Remove points that do not agree with the majority
  of their k nearest neighbours

Earlier example
Overlapping classes
Original data
Original data
Wilson editing with k7
Wilson editing with k7
23
Multi-edit

1. Diffusion divide data into N 3 random subsets
2. Classification Classify Si using 1-NN with
   S(i1)Mod N as the training set (i 1..N)
3. Editing Discard all samples incorrectly
   classified in (2)
4. Confusion Pool all remaining samples into a new
   set
5. Termination If the last I iterations produced no
   editing then end otherwise go to (1)

• Multi-edit Devijer Kittler 79
• Repeatedly apply Wilson editing to random
  partitions
• Classify with the 1-NN rule
• Approximates the error rate of the Bayes decision
  rule

Multi-edit, 8 iterations last 3 same
24
Combined Editing/Condensing

• First edit the data to remove noise and smooth
  the boundary
• Then condense to obtain a smaller subset

25
Where are we with respect to NN?

• Simple method, pretty powerful rule
• Very popular in text mining (?seems to work well
  for this task)
• Can be made to run fast
• Requires a lot of training data
• Edit to reduce noise, class overlap, overfitting
• Condense to remove data that are not needed to
  enhance speed

26
Problems when using k-NN in Practice

• What distance measure to use?
• Often Euclidean distance is used
• Locally adaptive metrics
• More complicated with non-numeric data, or when
  different dimensions have different scales
• Choice of k?
• Cross-validation
• 1-NN often performs well in practice
• k-NN needed for overlapping classes
• Reduce k-NN problem to 1-NN through dataset
  editing