A systematic overview of Data Mining Algorithms


1
A systematic overview of Data Mining Algorithms

• Data Mining David Hand Ch 5

2

• Input data
• Output models or patterns

3
Data Mining algorithm specification

• Task (visualization classification clustering
  regression)
  
• Structure (functional form) of the model that we
  are fitting to the data (linear regression
  hierarchical clustering )
  
• Score function
• Goodness of fit
• Generalization
• Search/optimization method(hill climbing
  simulated annealing convergence specification )

4
CART classification

• Mapping an input vector x to a categorical class
  y
  
• Task prediction (classification)
• Model structure tree
• Score function cross-validation loss function
• Search method greedy local search

5
CART Tree

• Internal node binary test
• Leaf node class
• tree picture p.146
• Data vector x descends a unique path from root to
  leaf.
  
• Structure of the tree is derived from the data
• Choose the best variable to split the data
• Score function is misclassification loss
• S C ( y(i) y(i) )
• C is m x m matrix (m is number of classes)

6
Tree classification vs. Linear classification

• Tree can deal with mixed data types (combination
  of categorical and real-valued data)
  
• Easier to deal with large number of variables
  (because process one at a time)
  

7
CART tree classification

• Goal of CART is to find a tree closest to optimal
  tree
  
• Complex enough to capture the structure
• Not too complex to avoid overfit
• Uses cross-validation technique

8
CART

• Search space is all possible trees (combinatorial
  large).
  
• Approximation algorithm Uses greedy local
  search to identify good candidate tree
  structures
  
• 2 phase
• Recursively expands the tree from root
• Prunes branches

9
Reductionist view on data mining algorithms

• Algorithms are tuples of
• model structure score function search method
  data management
  
• When deciding on an algorithm for an application
  think of which components are suitable

10
Multilayer Preceptrons (MLP) for Regression and
Classification

• Non-linear mapping from a real-valued input
  vector x to a real-valued output y
  
• Can be used as
• Nonlinear model for regression
• classification

11
MLP Basic idea
12
MLP algorithm tuple

• Task prediction
• Structure layers of nonlinear transformations
  of weighted sums of the inputs
  
• Score function sum of squared errors
• Search method steepest-descent from randomly
  chosen initial parameter values

13
MLP

• No simple summary (as with a tree)
• Nonlinear nature of its model structure
• Y is a nonlinear function of the inputs
• Parameter ( the weights) appear non-linearly in
  the score function ()

14
Score function

• Commonly used Score function
• S S( y(i) y(i))2
• y(i) is a true target value and y(i) is a output
  of the network for the ith data point
  
• y(i) is a function of the input vector x(i) and
  the weights
  
• Training network minimizing S

15
Training methods

• Back-propagation
• Steepest descent on the score function descends
  to local minimum given a randomly chosen point in
  the parameter space.
  
• Non-linear optimization techniques

16
Tree vs. MLP

• Tree algorithm search through models of different
  complexities in a relatively automated manner.
  
• No accepted procedure for network structure of
  MLP (number of hidden nodes layers)
  
• In practice trial-and-error

17
A Priori Alg. For Association Rule Learning

• Association Rule
• IF A 1 AND B1 THEN C1 with prob p
• ABC are binary variables.
• accuracy pa is conditional probability
  p(C1A1B1)
  
• support ps p(A1B1C1)
• Specify pa ps as thresholds
• Goal is to find all rules that satisfy threshold
  values

18

• Summaries of co-occurrence patterns in the
  observed data
  
• Correlation relation not causation.

19
Association Rules

• Task description associations between
  variables
  
• Structure probabilistic association rules
• Score function threshold on accuracy and
  support
  
• Search method systematic (BFS) with pruning

20
Example
21
Association Rules Summary

• Systematic search explicitely tries to minimize
  number of linear scans through database
  
• Designed to operate on very large data sets.
• Focus on computational efficiency

22
Vector Space Algorithms for Text Retrieval

• Given Query object Q and a large database
• Find k objects that are most similar to Q
• How is similarity defined
• Reduce documents too a vector representation
• Angle measures similarity

23

• Task retrieval of k most similar documents in a
  database relative to a given query
  
• Representation vector of term occurrences
• Score function angle between two vectors
• Search method
• Model representation is the key idea (which terms
  to use in a vector)

24

• Statistical approach emphasize theoretical
  aspect of inference procedures (parameter
  estimation model selection)
  
• Computer science approach focus on more
  efficient search and data mangement.