Santiago Eibe, M' Angel Hidalgo, Ernestina Menasalvas

1

• Santiago Eibe, M. Angel Hidalgo, Ernestina
  Menasalvas
• Facultad de Informatica.
• Universidad Politecnica de Madrid
• UKDU- Berlin 2006

2
Agenda

• Motivation and introduction
• Previous work
• The approach

3
Motivation Data Mining Scenarios

• Mobile Health Monitoring mobile phones are being
  used to help manage certain health conditions
• Web Mining Applications
• Automotive continuous on-board monitoring and
  mining vehicle data streams ( H. Kargupta, et al
  04). VEDAS A Mobile and Distributed Data Stream
  Mining System for Real-Time Vehicle Monitoring.
  SIAM04 )
• Network Intrusion use data mining techniques to
  discover consistent and useful patterns to detect
  security anomalies
• Autonomous Space Applications
• remote control stations
• on-board platforms
• mining of data collected from satellite sensors
• And more ...

4
Evaluation and deployment in the new environment
Who is evaluating? Is the miner available?
What results should be deployed
5
Data Mining evaluation in Ubiq scenaria
CRISP-DM (ASSESS MODEL) The data mining
engineer interprets the models according to his
domain knowledge, the data mining success
criteria and the desired test design

• Is the data miner going to be available?
• Multiple viewpoints integrated context

Visualization as a catalyst Models
design Requirements extraction Usability and
reusability Autonomy
6
Related work

• J. Branch, B. Szymanski, R. Wolff, C. Gianella,
  H. Kargupta. (2006). In-Network Outlier Detection
  in Wireless Sensor Networks. (ICDCS)
• S. Datta, K. Bhaduri, C. Giannella, R. Wolff, H.
  Kargupta. (2006). Distributed Data Mining in
  Peer-to-Peer Networks. (Invited submission to the
  IEEE Internet Computing special issue on
  Distributed Data Mining),
• K. Liu, K Bhaduri, K. Das, P. Nguyen, H. Kargupta
  (2006). Client-side Web Mining for Community
  Formation in Peer-to-Peer Environments. SIGKDD
  workshop on web usage and analysis (WebKDD).
• Accessing and analyzing data from a ubiquitous
  computing environment offer many challenges. One
  of this is related to human-computer interaction.
  From a Knowledge Discovery point of view,
  important human-computer interaction issues are
  collaborative problem .

7
Related work

• Charu C. Aggarwal, Jiawei Han, Jianyong Wang,
  Philip S. Yu A Framework for On-Demand
  Classification of Evolving Data Streams. IEEE
  Trans. Knowl. Data Eng. 18(5) 577-589 (2006)
• When the data streams are fast and continuous, it
  becomes important to analyze and predict the
  trends quickly in online fashion
• Dietrich Wettschereck, Alípio Jorge, Steve Moyle
  Visualization and Evaluation Support of Knowledge
  Discovery through the Predictive Model Markup
  Language. KES 2003 493-501
• Towards Effective and Interpretable Data Mining
  by Visual Interaction . Charu Aggarwal ACM02
• Therefore, a natural strategy would be to devise
  a system which is centered around a
  human-computer interactive process. In such a
  system, the particular data mining task can be
  divided between the human and the computer in
  such a way that each entity performs the task
  that it is most well suited to. The active
  participation of the user has the additional
  advantage that he has a better understanding of
  the final results

8
Previous results SolEuNet

• Steve Moyle Collaborative Data Mining. The Data
  Mining and Knowledge Discovery Handbook 2005
  1043-1056
• RAMSYS is a web-based infrastructure for
  collaborative data mining. It is being developed
  in the SolEuNet European Project for virtual
  enterprise services in data mining and decision
  support. Central to RAMSYS is the data of sharing
  the current best understanding to foster
  efficient collaboration.

9
WHAT we proposeInfoVis, the catalyst for the
UDM process

• Simplify evaluation in ubiquitous environments
  through visualization
• Collaborative evaluation awareness model
• Fill the gap between data miner and the domain
  expert
• Domain expert can participate in the
  collaborative process
• Capture Semantics of the underlying process
  (mining and business)
• Tasks Descriptions to minimize the number and
  importance of user (analyst of business expert)
  errors in the process
• User descriptions to cover the diversity of
  interpretation
• Device descriptions
• Systematic approach to forming and reproducing
  visual data mining model evaluation in ubiquitous
  environments

10
Visualization and ubiquity the challenges

• Outputs to different devices/users (diversity
  accessibility)
• Inputs from different devices/users
• Data and knowledge presentation must respect the
  limits imposed by the combination of ubiquitous
  devices and general human perceptual and
  cognitive limitations (e.g., display resolution),
  and the specific requirements on accessibility
  posed by peoples diversity
• Location transparency
• Context sensitive
• Track changes
• Support feedback and iteration
• Support data streams mining
• Visualization of information related to patterns
  extraction together with pattern is a must to
  understand and track changes

11
Visual Evaluation FrameworkDesigning Tasks

• Visualization Requirements Extraction
  Requirement engeneering
• Prototyped based (user cases)
• Output Requirement Extraction
• Format
• Location
• User expertice
• Scope
• context
• Visual Metaphoras Design
• From a visual evaluation components library of
  previous designs
• reusability
• Defining new (ad-hoc) metaphoras

12
Visual Evaluation Framework

• Scenes
• Actors
• Channels

• Standards

13
Scenes

• Data Store
• Centralized or distributed
• Continuos load and management of the data
• Dealing with stream, windows, data aging
• Modeling Scene
• Mining Scene
• At least on data mining model
• Optionally
• User model
• Recommedation model
• Obedience model
• Ubiquitous Scene
• Descriptions related to ubiquitous environment
• Awareness model

14
Scenes(II)

• Transformation Scene
• Two main types
• Visualization-related
• from rules to graphs, filters, plug-ins, ...
• Ubiquity-related
• integration, adaptability, awareness,
• Visualization Scene
• Visual controls
• Visualization algorithms
• Controls and mechanisms to aid user interaction
• etc

15
Actors and channels

• Actors Different user profiles
• Human
• Machine (servicies)
• Channels
• Constraints to limit interactions in the model
• Support the underlying semantics for
  visualization
• Example
• If a user wants to transform a tree into a graph
  and that is not possible then there will not be a
  channel
• We have templates to allow definitions of new
  interaction

16
Advantages of the approach so far

• Models user behaviour in the evaluation
• Models semantics of the context
• Models interactions
• Usability
• Reusability
• Interaction of the user will be minimized
• A step towards collaborative mining
• Some degree of autonomy

17
Thanks!