 |
|
|
|
|
|
|
|
|
|
|
|
|
 |
|
|||||||||||
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
|
|
 |
 |
|
||||||||||
 |
|
|||||||||||
 |
|
|||||||||||
 |
 |
|
||||||||||
| Seminar Schedule (Seminar Organizer: Prof. Ziv Bar-Joseph) | |
Machine Learning Lunchtime Chats
|
Spring 2006 Seminar
Date: April 27, 2006 Time: 12:30 PM - 1:30 PM (Pizza while it lasts!) Location: 3305 Newell-Simon Hall Speaker: Ping Chen KDD Masters Candidate Title: Data Mining in Macroeconomic Data Sets Abstract: National Economic Input-Output (EIO) data describes the monetary transactions among economic sectors. The monetary transactions among these sectors form a weighted bi-directional network from a supply sector to a demand sector and the weight is equivalent to the transaction value between them. In this research, we study the web structure and properties of this network and identify patterns of inter-sector dependence evolution by investigating the historical EIO data over the years 1947-1982. Here we make the following contributions: The first is the discovery that economic transactions (the distribution of the weight) are highly skewed, but follow the double Pareto-lognormal distribution (dPlN). The second contribution is the design of Multiple Steps of Pattern REcognition in skewed DAta (M-SPREAD) which identifies patterns and clusters despite the skewness of the data set. We applied our methods on the EIO data and we found interesting and explainable patterns, such as correlations among sectors, various evolution patterns within different transaction scales, outlier sectors and outlier time-stamps.