Nonvisual Analytical Methods

1
Nonvisual Analytical Methods

• Presented by
• Terry Bobo
• Christina Laurentia
• Siew Lim

2
Introduction

• Although visualization is a very powerful
  technique, users sometimes use other analytical
  methods
• Statistical testing
• Decision trees
• Association rules
• Neural networks
• Genetic algorithms
• Unlike visualization, other approaches are
  employed with the user largely out of the
  analyses loop.

3
Statistical Methods

• The use of descriptive and inferential statistics
  is by far the most standard approach to data
  analysis
• Preferred choice in Science, Medicine, Business
• Requires the use of quantitive data

4
Statistical Methods (Cont.)

• Descriptive statistics include measures such as
• Mean (avg.)
• Median (middle value)
• Mode (most common value)
• Standard deviation (measure of variance)
• Range (high low values)
• Data distribution

5
Assessing Group Differences

• Statistics can be used for hypothesis testing
• Values are predicted before analysis begins
• Ex. Fictious data describing sales pattern during
  the months of November and December.

6
Assessing Group Differences (Cont.)

• The experimental hypothesis, H1, is tested
  against the standard null, H0
• H0 assumes there are no differences between the
  groups H1 assumes groups are different.

7
Assesing Group Differences(Cont.)

• H1 December figures are in fact higher than
  the Nov. Figures.
• H0 There is no real difference any difference
  is due to chance.

8
Size of Statistical Effects

• Important factors to consider before an outcome
  is considered reliable
• Probability level
• Ex. 0.05 or 0.10 (significance level)
• Number of observations included in sample
• Number of groups being compared
• The more degrees of freedom, the smaller the
  ratio needed in order to reject HO.

9
Predictive Regression Analysis

• Linear regression Analysis
• Used to make predictions about the numeric values
  of a variable within the problem space
  represented in the data set.
• Yields a best line fit to the data points

10
Predictive Regression Analysis(Cont.)
11
When to Use Statistical Analysis

• Conditions to be met before statistically
  reliable effects can be observed.
• The data must be in numeric form
• Divided into groups for analysis
• User must have some sort of hypothesis about what
  to expect to find in an analysis

12
Keeping the Big Picture in Mind

• Determine what the overall goal of the data
  mining activity is.
• Whether a statistical test is appropriate depends
  on the question being asked.
• Statistics is not always the best approach.

13
Keeping the Big Picture in Mind (Cont.)

• Ex. A few distributers were discovered to be
  engaged in fraud, which claims constituted a
  small portion of total claims

14
Segregating the data
Figure 6.4 decision tree describing car purchase
profiles of males and female
15
Using decision trees to build rules

• The frame variable( 2door or 4-door) represents
  the root node.
• Engine type (V4 or higher) is a child node of the
  root node, and so on.
• When a record enters the tree, it moves down
  until it reaches the point beyond which it can no
  longer move.

16
Assessing rules

• For each rule you can measure how often the
  submitted data records are properly classified.
• You can compute the error rate of the entire tree
  as weighted sum of the error rates for all of the
  individual leave.

17
When to use decision trees

• Decision trees are useful for problems in which
  the goal is to make broad categorical
  classifications or predictions.
• They are not useful in applications requiring
  quantitative variables.

18
Association rules specific predictions about the
values of

• Association rules are derived from a type of
  analysis that extracts information from
  coincidence.
• Sometimes called market basket analysis.
• This methodology allows you to discover
  correlations, or co-occurrences of transactional
  events.

19
The cross-correlation matrix

• Association rules are derived from analyses based
  on cross-correlation matrices in which the
  likelihood of each event occurring in conjunction
  with every other event is computed.

20
The cross-correlation matrix(cont.)
Figure 6.5 example of cross-correlation matrix
used to infer association rules about the
purchase of grocery store items
21
When to use association rules

• Association rule analysis will be most useful
  when you are doing exploratory analyses, looking
  for interesting relationships that might exist
  within a data set.

22
Neural networks

• A type of computational methodology commonly used
  for pattern identification and classification
• Comprised of nodes that are interconnected by
  excitatory and inhibitory connections
• Pattern of activity are used to represented
  information in the network in a distributed
  fashion.
• Supervised Unsupervised Learning

23
Supervised Learning

• Learning occurs in a supervised mode in which
  system are trained on a known set of target so
  that those targets are readily identified when
  presented as inputs to the system.
• On each trial during supervised learning, an
  input is presented to the system.
• The input activates certain nodes, and the system
  provides an output response based on the pattern
  of activation.

24
Supervised Learning ( contd )

• 4. If the output does not match the desired
  response during learning, the system is provided
  with feedback designed to modify the incorrect
  response.
• 5. Once the system has learned the correct
  responses to the set of training inputs, the
  learning mode is finished.
• 6. When the learning mode is finished, the
  system can be used to automate pattern detection
  and alert the user when incoming patterns match a
  previously learned output response.

25
Unsupervised Learning

• Do not require that the set of permissible output
  responses and their mapping to input be defined a
  priori.
• In unsupervised learning, the network forms its
  own set of outputs during training based on
  features extracted by the network.
• The most popular unsupervised learning is
  Kohonens feature map.

26
Unsupervised Learning ( contd))
Figure 6.6 topographical map produced by an
unsupervised learning network
27
When to use unsupervised neural networks

• The neural network approach to data mining is
  most useful when you are searching for novel ways
  of segmenting the data set.
• This method can be used to discover subgroups of
  data that defined in terms of some common
  feature(s) that separate them from other portions
  of the complete population.

28
Genetic Algorithms

• The utility of genetic algorithms lies more
  within the realm of optimization.
• Genetic algorithms start with a population of
  items and seek to alter and eventually optimize
  their composition for the solution of a
  particular problem.
• The genetic material or information represented
  by each individual can be passed on to subsequent
  generations in a variety of ways with
  optimization occurring in the process.
• Three basic mechanisms in which information is
  chosen, altered, and passed on in order to
  achieve optimization selection, crossover, and
  mutation.

29
Selection

• The process of selection is analogous to the
  process of natural selection that occurs in
  evolution
• Selection is based on the principle of survival
  of the fittest in which the individuals that are
  best suited for the environment are the ones that
  survive to pass their genetic material on to the
  next generation.

30
Crossover

• Crossover occurs when two individuals chosen
  randomly from the population are joined or
  mated such that the resulting offspring contain
  partial replications of the information contained
  in each of the parents.
• The offspring then become full-fledged members of
  the population, competing for survival along with
  the rest.

31
Mutation

• Mutations can occur naturally when there is an
  error in the transmission of genetic information
  from parent to child.
• Mutations can have either good or bad effects.

32
When to use Genetic Algorithms

• Genetic algorithms are most useful in cases in
  which the goal is to find an optimal solution
  given a definable problem space.
• Genetic algorithms are most useful in situations
  in which you are combining data from several
  disparate information sources and types.
• There needs to be a fair amount of uniformity in
  terms of data to be analyzed since all data must
  be coded into vectors of the same dimensionality.

33
Christinas Questions

• What are two types of Neural Networks?
• What are basic mechanisms in Genetic Algorithms?
  When is the best time to use Genetic Algorithms?
  Answer

34
Terrys Question

• Question Distinguish between H1, and HO.

35
Siews question

• When is a good time to use decision trees ?