Mining Time Series Using Rough Sets a Case Study

1
Mining Time Series Using Rough Sets - a Case Study

• Anders Torvill Bjorvand
• Troll Data Inc.
• torvill_at_trolldata.no

2
Objective

• We want to reason about sequences of events
• Including historic attributes increases the
  number of attributes initially
• Is it possible to obtain shorter reducts from
  this approach?

3
Temporal Information Systems

• A Temporal Information System, TIS, is a regular
  Information System, IS, where one of the
  attributes are designated as a sequence variable

4
TIS to IS - I

• We have to transform our TIS to an IS in order to
  effectively reason with sequences
• Based on how many time periods we are interested
  in going back, we create new attributes
  representing the normalized change between each
  subsequent period of time

5
TIS to IS - II

• Lets say that we have some market data with 100
  sequential measurements spanning 30 attributes
• 1 month back ? 99 objects, 30 attributes
• 2 months back ? 98 objects, 60 attributes
• 3 months back ? 97 objects, 90 attributes
• ...

6
TIS to IS - III
7
Data Reduction Potential

• Both increasing the number of attributes and
  decreasing the number of original objects gives
  us a potential for shorter reducts
• Is this true only in theory, or will we observe
  this with real world data ?

8
Experiment

• We computed the shortest reduct of two
  transformations of stock market data (W. Ziarko,
  120 obj. 30 attr.)
• 1 month back 6 attributes
• 2 months back 4 attributes

9
Future Work - I

• Compose further experiments - to verify/evaluate
  the theory both qualitatively and quantitatively
• Prediction of missing values
• Scaling of time intervals for the real time
  approach
• Handling of branching time structure

10
Future Work - II

• Jbin - OODBMS in Java
• Emphasis on the distributed aspects of KDD
• Persistent storage of objects conforming to the
  JavaBean component standard
• Mining integrated in the kernel through the
  engines of Rough Enough and Neuromania

11
Future Work - III

• Synthesis of embedded decision support systems
• Means An algorithm mapping from rule-sets to
  Java-code conforming to the EmbeddedJava API
• Goals Obtaining support for mobile computing
  units comforming to the JavaOS specifications
• Domestic appliances
• Cellular phones, etc.

12
Available Papers

• SCAI97 Rough Enough - A System Supporting the
  Rough Sets Approach
• IMACS97 - Rough Set session Practical
  Applications of Genetic Algorithms for Efficient
  Reduct Computation
• IMACS97 - Rough Set session Rough Enough -
  Software Demonstration
• Masters thesis Time Series and Rough Sets
• Also available at
• http//home.sn.no/torvill