Useful Patterns

The goal of the Useful Patterns workshop (UP) is to address the problem of making the results of pattern mining useful. Pattern mining is an important aspect of data mining, concerned with finding local structure in the data. Traditionally, the focus of research in pattern mining has been on completeness - and the efficiency in achieving it. This focus, important as it is, has led our attention away from the most important aspect of the exercise: leading to useful results. How to make the results of pattern mining useful; that's exactly what we want to address in UP.

Proceedings

The UP 2010 proceedings are available through the ACM Digital Library here.
The program for UP 2010 consisted of the following talks.

Opening (slides)
by Jilles Vreeken

Keynote Presentation
'Mining Useful Patterns: My Evolutionary View' (abstract, slides)
by Jiawei Han

Keynote Presentation
'Association Discovery' (abstract, slides)
by Geoff Webb

'Multi-Resolution Patterns from Binary Data' (link, slides)
by Prem Raj Adhikari & Jaakko Hollmén

'CloseViz: Visualising Useful Patterns' (link, slides)
by Chris Carmichael & Carson K. Leung

'A Framework for Mining Interesting Pattern Sets' (link, slides)
by Tijl De Bie & Kleantis-Nikolaos Kontonasios & Eirini Spyropoulou

'Point-Distribution Algorithm for Mining Vector-Item Patterns' (link, slides)
by Anne Denton & Jianfei Wu & Dietmar Dorr

'Margin-Closed Frequent Sequential Pattern Mining' (link, slides)
by Dmitriy Fradkin & Fabian Moerchen

'Block Interaction: A Generative Summarization Scheme for Frequent Patterns' (link, slides)
by Ruoming Jin & Yang Xiang & Hui Hong & Kun Huang

'Authorship Classification: A Syntatic Tree Mining Approach' (link, slides)
by Sangkyum Kim & Hyungsul Kim & Tim Weninger & Jiawei Han

'Pattern Selection Problems in Multivariate Time-Series using Equation Discovery' (link, slides)
by Arne Koopman & Arno Knobbe & Marvin Meeng

Workshop Goals

We expect the outcome of the workshop to be new insights with regard to useful pattern mining. UP provides a venue for results on presentation, visualisation and exploration of patterns by users. Moreover, by a great selection of regular papers and invited speakers, we bring together different approaches and views on the field of pattern mining. Great discussion is ensured.

We divide the scope of UP into three areas. Our first area of discussion is how to reduce the amount of returned patterns to useful amounts, whilst retaining the most important information. Our second area is to discuss how to present patterns to a user in a useful manner, e.g., allowing for visual exploration of the pattern set. Our third goal is to discuss how (small) sets of patterns can be used as surrogates for the original data. The main idea here is to challenge the belief that patterns are the end goal but instead use them as an intermediate result.

Invited Speakers

We are proud to have

as the keynote speakers at our workshop.

Jiawei Han will give a keynote presentation titled 'Mining Useful Patterns: My Evolutionary View'. It will detail his evolutionary view on the usefulness of patterns, and present a set of examples on what patterns are considered to be useful in certain practice. This will give some insight on pattern analysis, based on his research, and point out a few open research problems and possible exploration of broad applications of pattern mining.

Geoff Webb will give a keynote presentation titled 'Association Discovery'. In the talk Geoff will introduce what association discovery is, and how it differs from what statisticians do by traditional correlation analysis. A major topic of the talk will be top-most interesting patterns, and how to find these - discussing the strengths and limitations of using statistical testing to do so. The point the talk will drive to, is the question what we should look for: associations, association rules or itemsets.

Organizers

You can contact us at:
contact (at) usefulpatterns.org

Program Committee

  • Michael Berthold, University of Konstanz
  • Björn Bringmann, K.U. Leuven
  • Johannes Fürnkranz, T.U. Darmstadt
  • Vivekanand Gopalkrishnan, Nanyang Technological University
  • Ruoming Jin, Kent State University
  • Eamonn Keogh, University of California - Riverside
  • Arno Knobbe, Universiteit Leiden
  • Arne Koopman, Universiteit Leiden
  • Carson K. Leung, University of Manitoba
  • Srinivasan Parthasarathy, Ohio State University
  • Jian Pei, Simon Fraser University
  • Kai Puolamäki, Aalto University
  • Geoff Webb, Monash University

What's UP?

Pattern mining is an important aspect of data mining, concerned with finding local structure in data. Traditionally, the focus of research in pattern mining has been on completeness and efficiency. That is, trying to find all potentially interesting patterns as fast as possible. This focus, important as it is, has led our attention away from the most important aspect of the exercise: leading to useful results. Let's consider the following example.

Say a domain expert wants to extract novel knowledge from some data, and specifically wants to know what patterns are present in the data. To do so, the expert involves a data analyst. The analyst is provided with the data, and runs his favorite pattern mining algorithm. Due to the pattern explosion, the number of discovered patterns the analyst will find will be enormous; the result of the mining exercise often being much larger than the original data. Nevertheless, let us assume our expert patiently considers the result. Although he might stumble upon some interesting patterns, he will mostly encounter very many patterns that convey roughly the same information. Perhaps worse, however, is that he will find that many of the patterns represent information that is already known.

All things considered, even when convinced of the potential, in the above case the expert would not be very impressed by the usefulness of pattern mining. Unlike in other fields of data mining, such as clustering, in pattern mining presentation and visualization has not been a priority. However, even when we forget about presentation to a user, patterns are not yet as useful as they could be. While they provide highly detailed descriptions of phenomena in data, it remains difficult to make good use of them in, say, e.g., classification or clustering. While this is mostly due to the huge number of discovered patterns, making the result unwieldy at best, it does pose interesting research questions like 'how to select patterns such that they are useful?'. Techniques that summarize the result exist, but focus primarily on being able to reconstruct the full set, instead of targeting the usability of the summarized set. As such, research into techniques that mine small sets of high-quality patterns is required, where high-quality is directly related their intended use.

It is exactly this research, experiences and practices that we want to discuss with UP.