ataki / guacamole Goto Github PK

Goal Implement trust metric as outlined here and in Trustlet. Reference code is here on Pymmetry

Output A script tm_epinions.py, and a visualization in figs/tm_epinions.png

tagging @Hongxia for trust metric reference

Background

We have successfully generated a Dandekar graph for the milestone with ~7k nodes. We now need to evaluate trust metrics. A part of this process is to find seed parameters which will allow us to obtain a non-trivial trust result, if such exist (non-trivial means that a positive trust score is assigned to nodes that are further than 3 hops away from the source node). If this result is not possible, explain why.

Goal

The goal is to:

• find seed parameters such that we obtain a non-trivial result when running the algorithm on a Dandekar-generated graph, and
• if none exist, explain why

Background

We want to devise and implement a procedure which allows us to evaluate the relative trust metric of two graphs. The approach we have settled on at the moment is:

• given the two graphs, pick source nodes at random
• evaluate the trust metric for both graphs, and generate list of "untrusted" nodes. Designate this the "bad" set.
• choose nodes to delete subject to a function inversely proportional to distance from the source node
  (i.e. much more likely to be deleted when farther away)
• evaluate the trust metric for both graphs again see how much the newly obtained "bad" set compares to the old "bad" set from the first run.

We will generate a scoring function based on the severity of the bad nodes chosen. Our scoring function needs to decide whether to punish additional / missing bad nodes the same way, or to score additions more heavily than bad nodes. This will dictate what type of attack we want to mitigate most.

Since this is a large and important task, there will be three big steps:

1. @ataki to research / brainstorm scoring functions and discuss significance with TA's
2. @Hongxia and @LCanceled to weigh in and ultimately decide on scoring function for their graphs

We want to generate a graph based on the Dandekar algorithm here (page 4)

http://snap.stanford.edu/class/cs224w-2010/proj2009/final_report_Dandekar.pdf

Place a script in the project dir called generate_dandekar.py, as well as an exported SNAP graph in data/dandekar/dandekar.edges.

Goal

Implement trust metric as outlined here in Trustlet for the Advogato dataset.

Dataset (Keep in mind these are in the .dot format, and so may need to be parsed first)

Summary here. Reference dataset here. Try to run on the daily output data here. Summary of dataset

Output

A script tm_epinions.py, and a visualization in figs/tm_epinions.png