Leman Akoglu et al KDD'10

From Cohen Courses
Revision as of 22:17, 15 February 2011 by Diliu (talk | contribs)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigationJump to search

Citation

Leman Akoglu, Bhavana Dalvi Structure, Tie Persistence and Event Detection in Large Phone and SMS Networks In KDD'10

Summary

This paper focuses on a large phone call data set from an Indian phone company. The paper tries to answer three questions: First, what is the structural property of the dataset? What are the relationships between different quantities taken from the egonet? Second, if a link exists between two people, will the link still exist in the future? Third, how to detect change-point anomalies?

Brief Description

The authors considered tie attributes and node attributes in the analysis. For tie attributes, they considered reciprocity and topology overlap. For node attributes, they considered node degrees, Cluster coefficients and user reciprocity defined as the proportion of reciprocity ties.

We first talk about general network property of the data. The authors found that the weights on mutual node pairs in Mobile Call Graph or Mobile Text graph are both small and even. They also found that nodes with high degree tend to connect to other high-degree nodes; node strength grows in a power law manner comparing to the node degree. Moreover, node strength increases as the number of overlapping neighbors between two nodes increases.

Moreover, the authors attacked the tie persistence problem using logistic regression. They compared their method with the traditional LR approach as well.

The authors also looked at the change-point behavior. Specifically, they are interested in detecting the time stamp where the behavior of each node changes. They first extracted features including degree, weight, reciprocity edges, triangles in egonet etc. They next considered an eigen decomposition approach using tensor. The third dimension is the time. After this step each node can be represented as a matrix changing over time. By taking eigenvectors over time, they are able to identify different behaviors using dot product between these vectors.