Predicting signed edges
This notebook contains results from the work Predicting Positive and Negative Links: Theory & Applications, authored by:
- Charalampos E. Tsourakakis (Boston University, Harvard University)
- Michael Mitzenmacher (Harvard University)
- Jaroslaw Blasiok (Harvard University)
- Ben Lawson (Boston University)
- Preetum Nakkira (Harvard University)
- Vasileios Nakos (Harvard University)
The file Constructing features.ipynb demonstrates the construction of features used to predict the sign of an edge. The features we use include (a) the features proposed by Leskovec et al, (b) and the number of (positive, negative)x paths of length 3 and 4 respectively. Therefore, we add 4 new features. This notebook shows a quick implementation. We use the Highland tribes dataset from KONECT.
In Prediction.ipynb we analyze the signedd Wikipedia dataset from SNAP. We train a logistic regression classifer using stratified 10-fold cross-validation. We observe that paths of short length an informative feature.