Geographical data mining: key design issues

1
Geographical data mining key design issues

• Stan Openshaw, Centre for Computational
  Geography, School of Geography, University of
  Leeds, United Kingdom GeoComputation99
• ??????

2
Introduction

• Explosion in Geographically referenced data
• Occasioned by developments in IT, digital
  mapping, remote sensing, and the global diffusion
  of GIS
• A GIS context DM tool are to be made in order to
  exploit the flood of geoinformation

3
Introduction

• Data mining tools analyze very large commercial
  databases in order to model and predict
  customer-buying behaviour.
• As Geographical Data Mining should do more in
  spatial pattern recognition.
• The focus here is on developing an explicitly
  Geographical Data Mining(GDM) technology.

4
Why geographical data mining?

• data uncertainty and errors are often spatial
  structured
• whole map statistics are seldom helpful
• relationships are often geographical localised -
  rather than global
• non-linearity is the norm
• time often interacts with space
• most GIS data layers are categorical
• the locational element is important
• there can be a fair proportion of junk data

5

• many conventional statistical tools continue to
  be useful and that they have already well served
  quantitative geography for more than 3 decades
• however, the view here is that many of these
  conventional, general purpose, statistical
  techniques are not sufficiently focused on, and
  tailored to, the special needs of geographical
  analysis

6
Typical generic data mining functions

• exploratory data analysis tools
• linked map and graph displays
• other visual data mining tools
• most multivariate statistical methods
• linear and logistic regression
• classification
• decision trees and regression trees
• association rules detection
• neural networks
• memory based reasoning

7
example

• If you input some X, Y referenced data into a
  data mining package and expect it to identify
  localised clusters of excess incidence of a
  disease, then you would probably be disappointed.
  
• These packages could only treat the X,Y
  co-ordinates as if they were merely two ordinary
  variables (such as age or income) and it is very
  likely that nothing useful would be achieved.

8
Geographical analysis machines

• GIS-relevant GDM ?????1.a human being based
  exploratory graphical methods approach, and2.the
  development of automated analysis machines.
• The first prototype of an automated GAM was
  developed in mid 1980s.(the original GDM)

9
The justification for automation

• there are many possible locations of potential
  localised clusters (many millions),
• the search should be locationally unbiased and
  geographically comprehensive, and
• it should handle spatial data uncertainty.

10
the explosion of GIS databases

• the need to speedily analyse in hours,
• an increasing need to perform routine analysis
  for monitoring purposes
• few skilled spatial analysts,
• high imperative for geographical analysis
• as computer hardware has become faster,
• automation makes it easier to develop
  user-friendly interfaces

11

• the current GAMK version is a very powerful
  cluster detector that in blind testing
• however, it is only a descriptive tool
• GAM uses a brute force search
• A space-time version has recently been developed
  (GAMK-T)

12
Geographical Explanations Machines

• a traditional map-based GIS analysis tool that
  of map overlay
• It has long been found useful to overlay two
  coverages
• The hope is that some of the polygons created by
  the map overlay process will define "interesting"
  results

13

• Instead of overlaying complete maps, the data
  inside each of the GAM search circles are
  examined using permutations of coverages
• This new method is termed the Geographical
  Exploration Machine (GEM).

14
GIS data characteristics

• Openshaw (1994,5) pointed out that GIS databases
  consist of three broad classes of data types that
  form a complex multivariate tri-space and that it
  is these spaces, and the interactions between
  them,
• 1. Geographical coordinates,
• 2. Temporal coordinate(s), and
• 3. Multiple attributes relating to the
  geographical entities.

15

• note that all three types of space are measured
  in different units that cannot be directly
  related to each other

16
GDM Basic functionality

• Serve basic spatial data exploratory needs,
• Have the potential to create new insights, ideas,
  and hypotheses from the analysis,
• Offer artistic impressions of pattern structure
  to stimulate the imagination,
• Spot major unusual localised database patterns
  and detect empirical location-based regularities,
  and
• Be easy to use and meet basic GISability criteria
  (Openshaw and Fischer, 1995).

17
Trispace characteristics of a GIS database

• Datatype Nature of Data
• 1. geography data
• 2. time data
• 3. multiple attribute data
• 4. geography and time data
• 5. time and multiple attribute
  data
• 6. ?geography and multiple attribute data
• 7. geography, time, and multiple attribute
  data

18

• most current geographical analysis tools only
  work with data type 1
• Time Data (data type 2) can be handled via
  well-developed time series methods, but these are
  implicitly space free
• Attribute only data (type 3) can be handled via
  conventional multivariate statistical (and
  current data mining) methods

19

• space-time clustering measures exist (i.e., Knox
  and Mantel statistics)
• space-time models have so far been developed
  (i.e., STARIMAs).
• Time and multiple attribute data (data type 5)
  are even harder to analyse. The best example
  would be multiple time series modelling for a
  fixed area.
• Methods that can handle data with this level of
  complexity should have little difficulty with all
  the rest.

20
Visiting hyperspace

• 1. Progressively re-map the hyperspace greatly
  reducing its dimensionality until it becomes
  manageable
• 2. Develop some kind of virtual reality
  hyperspace explorer and data visualiser so that
  you can go into the database's hyperspace
• 3. Create GDM agents that are able to visit these
  spaces for you and then report back

21

• Human beings can only easily handle a small
  number of dimensions of multivariate space.
• impossible problems in visualising four, five, or
  fifty more dimensions
• create intelligent agents or artificial creatures
  or robots able to search these higher dimensional
  spaces in

22
Handling time-varying population at risk data

• disease incidence data might be monthly, but the
  best census-based population at risk data are
  updated once every 10 years

23

• Ignore the problem on the grounds that any
  patterns created by this neglect may still be of
  considerable interest
• Use an expected value based on historic data or
  long-term average values or estimates of a
  maximum value so as to minimise the risk of false
  patterns being uncovered and
• Estimate population-at-risk values for each time
  period for which there are data to be analysed,
  perhaps using linear interpolation.

24
Understand the findings

• A summary view in geography space of any and all
  clustering that has been found
• List of best search regions together with
  coordinates in hyperspace, and
• Hypertext linkages or databases that connected
  (2) to (1) and show the contribution (2) makes to
  1.

25
conclusion

• The paper has focused on the design issues in
  building a GIS-appropriate Geographical Data
  Mining tool.
• Preliminary results will be presented at
  GeoComputation '99 using synthetic data with
  varying levels of pattern complexity.
• The belief is that GDM actually works
  surprisingly well, at least on synthetic data.