Medha Journal Presentation

From Semantic Portal Wiki

Presentation given at CSCI 6966 Advanced Semantic Web (Fall 2008)#Lesson 12

• Speaker: Medha Atre
• Title: Discovering and Explaining Abnormal Nodes in Semantic Graphs
• Authors: Shou-de Lin, Hans Chalupsky
• Conference: IEEE Transactions on Knowledge and Data Engineering, vol. 20(8): 1039-1052
• URL: http://ieeexplore.ieee.org/search/wrapper.jsp?arnumber=4358968
• Date of Presentation: 2008/11/20

Questions

ID	Question	Name	Answer
Abnormal Nodes Ankesh	I wish authors could have written a line describing MI & PMI. It would be useful to extend this approach to consider semantic relations between labels (subPropertyOf). I perceive a disconnect between method of finding outliers and the explanation that is generated. Do you see that as a problem? (Has it got anything to do with getOutliers(profile) algorithm?) Outliers are found using distance based algorithm, where as explanation is generated by selecting the features used by the classifier to classify the nodes into outliers and non-outliers found earlier. I'm not really able to connect the two things. Because o/w we can simply train our classifier to classify based on feature selection- and we can isolate outliers from it.	Ankesh Khandelwal
Lin2008discovering question 1 by lebo	The authors make a very good point regarding the "ill defined" nature of using probabilistic measures for a deterministic graph structure. The two Random Experiments that they propose are unique, intuitive, and probabilistically sound methods for obtaining probabilistic measures. But from the time that they propose the method, they do not discuss the methods' computational expense until the last paragraph of the paper, "an important future direction is to improve the scalability of the system. What is most expensive is the computation of feature values, since it requires the system to count a potentially large number of paths." How long should the Random Experiments be reasonably run to obtain characterize the input while trading off the time required to do so?	Tim Lebo
Medha Journal Presentation GTW 1	In defending a maximum path length metaconstraint, section 2.3 says "Moreover, the longer a path is, the harder it is for humans to make sense of it." Isn't this overly simplistic, in that what makes interpreting a long path difficult is the path length *and* a varied path makeup? Wouldn't a long path of, for example, all 'knows' relations be easy to interpret regardless of path length?	Gregory Todd Williams
Medha Journal Presentation GTW 2	Would this work benefit from considering more complex node features than just simple paths? Would using more complex features (trees, DAGs, or full graphs) to capture more inter-dependence be able to find abnormal nodes that simple paths might not?	Gregory Todd Williams
Medha Journal Presentation Jesse Weaver	Section 2.3 states: "By default, our system uses a maximum path length of five, which has worked well on the various data sets we have analyzed so far." However, the authors do not give any particular reason as to why a path length of five works well. Are you aware of any particular reason? Do you think that perhaps this assumption skews the results presented in Table 3? Perhaps a path length of five is particularly good for the Mafiya dataset (and their other datasets as mentioned above), but is it generally good for all datasets?	Jesse Weaver
Medha Journal Presentation Joshua Shinavier 1	It would be helpful to see a time complexity analysis of the proposed technique. The authors state that a graph of 40,000 nodes and 475,000 edges takes around two minutes to analyze on a typical PC, which is better than one might expect. But now, how about a graph with a million edges? Does the algorithm take four minutes or four years to produce a result?	Joshua Shinavier
Medha Journal Presentation Joshua Shinavier 2	To what extent does this technique depend on the nature of the MRN under analysis? Are there meaningful networks (particularly, networks of interest to the intelligence domain) for which the diversity of simple paths makes it impossible to identify "abnormal" nodes?	Joshua Shinavier

Attendees

Tim Lebo