Temporal Concept Learning (or how to predict the future?)

קירה רדינסקי, הרצאה סמינריונית למגיסטר
יום רביעי, 7.4.2010, 13:00
טאוב 337
Prof. S. Markovitch

Man has long desired to predict events that are likely to take place in the future. Such predictions can be beneficial for planning, resource allocation and identification of risks and opportunities. Predicting events in politics, economics, society, etc. is a difficult task that is usually performed by human experts possessing extensive domain-specific and common-sense knowledge. In this research we aim to learn how to predict in temporal environments with dynamically changing features, labels and temporal correlations with other objects in the domain. We develop a general framework for temporal concept learning, as a generalization of the traditional learning architecture, and present a family of algorithms for efficient concept learning in temporal domains. We present experiments on several real-world prediction tasks, such as foreign exchange, weather, oil trading and hurricane prediction, that confirm the superior performance of our method compared to the previous state of the art methods.

בחזרה לאינדקס האירועים