Welcome to the First Workshop on recent advances in behavior prediction and pro-active pervasive computing
Behavior and context prediction breaks the border from reaction on past and present stimuli to proactive anticipation of actions. Researchers have for about one decade now considered the prediction of such stimuli to enable pro-active context computing. Research directions spread from applications for behavior and context prediction over event prediction, machine learning, architectures for context prediction, data formats and algorithms.
Even though a great diversity of applications for behavior and context prediction has been proposed, a common methodology or platform has not yet crystallised. Application developers are forced to start from scratch since previous authors seldom provided usable sources of their applications that could be extended. To foster the integration into applications, support for application developers has to be improved. We require a widely accepted architecture and toolkit – easy to use; with an open API; comprising common algorithms, accepted data sets and benchmarks. It should enable researchers to test prediction algorithms in a common environment on accepted data sets as well as to extend and import it by own algorithms and data sets.
When it comes to algorithms for context prediction, a comprehensive comparison of strengths and weaknesses on benchmark data sets is yet missing. The motivation for choosing an algorithm for a specific application is not seldom driven by the experience and education of the researcher. Therefore, inherent properties such as the structure and requirements of the data as well as the application regarding accuracy and processing load are ignored. To raise the field to a level at which it might be integrated in commercial applications, common, widely accepted data sets need to be established as well as accepted benchmarks.
Analytic studies mainly consider the computational complexity of time-series forecasting methods. They are required to establish a theoretically sound background for applications. Data formats or impacts of the restriction to few symbolical formats might foster comprehension of basic issues for prediction. Additionally, the computational complexity and the ability to distribute computational load among nodes in a network are promising research directions in order to enable prediction in systems of distributed, resource limited nodes.
Promising ideas, broadening the field have been mentioned but are addressed superficially. A prominent example is prediction sharing among nodes in proximity. Related questions regard privacy and trust, service quality, communication and storage cost and accuracy amplification through redundancy. Likewise, the sharing of time series might be utilised for correction of measurement errors or predictions. Also, authors seldom address the prediction of rare events. In particular, for disaster or accident prevention, we would like to prevent extremely unlikely events for which possibly no training data exists.
After about one decade of activities in the field of behavior and context prediction, the workshop will bring together researchers of this field and reveal important open issues. Among these are
Accurate prediction of seldom events
Accurate prediction of seldom events
Important events are frequently also seldom events. For instance, several conditions and errors might lead to a car accident. How can we train a system on events which are not likely covered by training data sets?
Continuous learning
User behaviour and habit changes over time. To guarantee constant quality of prediction, the method must be able to ‘forget’ patterns that grow unimportant. Personalisation of prediction: Different users behave differently. Although a common scheme may exist, prediction algorithms have to tolerate noise.
Public spaces
The utilisation of context data in public spaces will grow more important. Since context prediction inherently relies on personal context patterns, new concepts that utilise shared data in a personalised way are desired.
Data sets and benchmarks
Currently, comprehensive data-sets are created for context-computing. However, these data-sets are hardly sufficient to be applied for context prediction applications. In particular, data has to be sampled over longer time-spans and cover stochastic processes which are inherently predictable.
Sharing of prediction and time series
Since context time-series of nodes in proximity are related, this redundancy might be utilised to correct and detect errors in the data and improve prediction accuracy.
Privacy and trust
Shared time series but also the fact that context time series might cover events and actions of remote entities rises questions of privacy and trust.