TDM Workshop at ICDM 2005

Many real-world applications deal with huge amounts of temporal data. Examples include alarms/events and performance measurements generated by distributed computer systems and by telecommunication networks, the web server logs, online transaction logs, financial data, workflow process logs, and sensor data collected from sensor networks. Conventionally, temporal data is classified to either categorical event streams or numerical time series and both types have been intensively studied in data mining and statistics. However, several previously less emphasized aspects of temporal data have proven their importance in emerging applications and posed several challenges calling for more research. In addition, the applications of temporal data analysis, such as web services, information navigation, system management, adaptive workflow management, program behavior analysis, security management and bioinformatics, are enjoying a growing amount of attention. This workshop aims to gather researchers and practitioners to tackle these challenges and attempts to study the common tasks that need to be addressed in practical applications.

The setting of traditional temporal data analysis is to apply one algorithm on a static, regular and relatively small temporal data set. Many practitioners found existing analysis methods inadequate for their real-world data. Many struggle to transform the data in order to apply existing methods or even to reduce the original problems to better studied ones; either ways induce in more preprocessing effort, more artificial parameters and less interpretable results. We believe these new aspects of temporal data deserve theories and algorithms of their own. Some of these new aspects are:

• Irregularity: Many types of numerical temporal data are not equally paced.
  
  
• Asynchronism: In distributed computing environments like sensor networks, data from different sources tend to be not aligned and hence can not apply synchronous methods.
  
  
• Distributed analysis: A trend in temporal data analysis is to perform data filtering, transformation and analysis as close as possible to the data sources to avoid the prohibitive amount of data being transmitted and analyzed. This new computing paradigm calls for a new theoretical foundation.
  
  
• Streaming Data: Some temporal data is stored only temporally and requires near real-time analysis.
  
  
• Heterogeneous data types: It is very common that temporal data is partly categorical events and partly numerical time series. It remains to be an interesting challenging to best analyze all possible data in a uniform way.
  
  
• Huge Volume: The stream of data can be huge for a long, continuous observation period. Many types of measurements can be obtained from a large number of data sources. This requires designing scalable solutions in analyzing a large volume of temporal data, in terms of both the large number of data points and the large number of types of measurements.

Driven by the new aspects of temporal data, several fundamental problems need to be revisited. Just to name a few of them:

Prediction Correlation Regression Benchmarking Periodic Pattern Mining Temporal Association Finding

Causality Analysis Sequential Event Patterns Threshold selection Frequency Analysis Anomaly Detection Clustering and Classification