Temporal Machine Learning Research

What's new

Last modified: 18 March 2001.

7 May 2001: Added link to Frank Hoeppner's work.
18 March 2001: Added link to the Machine Learning for Spatial and Temporal Data workshop that is part of ICML 2001.

Introduction

Reliable and robust classification techniques have been developed for static domains for several years now. However, recently, there has been increased interest in classification, clustering, searching and other processing of things that vary over time. These include things like sensor information from robots, signals from biomedical sources like electrocardiographs, financial markets, gesture and more. The goal is to find patterns in data that varies over time.

This page lists the researchers that I know about that work in the area. If you have any suggestions, people I've missed, etc., I would really appreciate it if you would contact me.

Dimensions of research

There are a variety of different sub-problems within this domain. The main dimensions of these variations are:

Supervised vs Unsupervised: The temporal data can be labelled, telling us what type of pattern is, or alternatively the goal can be more difficult, along the lines of finding interesting and recurring patterns.
Univariate vs Multivariate: Is a single "channel" of data being analysed, or is it a domain where there is more than one source of data that is simultaneously being analysed?

Reinventing the wheel?

Time series have also been analysed for a long time before machine learning researchers began taking a closer look at it. In view of this, it is especially important to avoid reinventing the wheel constantly. This section highlights some of the more established areas that correspond to parts of the temporal machine learning problem:

Hidden Markov Models
Recurrent Neural Networks
Allen's Interval Logic
Knowledge-based signal processing

Researchers

Gautam Das is doing work in finding similar time series.
Doug Fisher is looking at techniques for clustering time series. In addition, two of his students have worked or are working on the problem:
- Stefanos Manganaris has built a system for classifying univariate data streams for his PhD. His PhD also has a good literature survey for getting into the area.
- Doug Talbert is looking at clustering temporal structures.
Giorgios Paliouras did a PhD thesis on refining temporal recognition expert systems. He applied his techniques to whale songs. Also has a good literature survey, especially on knowledge-based signal processing.
Eamonn Keogh is looking at ways of picking up patterns from temporal data.
Heikki Mannila is interested in sequence data (like network logs) as well as time series (ECGs and the like). He has a wide range of papers available in the area.
Padhraic Smyth also does some work in the area, as well a great deal of work on a lot of other topics related to time series analysis.
Mohammed Waleed Kadous is developing a multivariate classification algorithm.
Jeffrey D. Scargle is working on techniques for analysing time series from astronomical sources.
Frank Hoeppner on the qualitative aspects of time series used by humans in analysing time series.

Data sets

Heart Data: As used by Keogh, one of the datasets from the Nonlinear Techniques in Physiological Time Series Analysis datasets.
Waleed Kadous's Sign Language Data collected by the author.
The UCI KDD Archive now has several temporal datasets available.

Related Conferences & Workshops

Predicting the future: AI Approaches to Time-series Analysis was held as a workshop to AAAI'98. It covered some of the most recent and interesting work in the area. The chair was Andrea Danyluk.
Machine Learning for Spatial and Temporal Data is being held as a workshop as part of ICML 2001. It is being held in Williamstown, Massachusetts.

Related Journals

Reminder

If you have any suggestions, corrections, etc., please don't hesitate to e-mail me at waleed@cse.unsw.edu.au

Mohammed Waleed Kadous

Last modified: Mon May 7 23:01:29 EST 2001