A decision-making based feature for link prediction in signed social networks

We are interested in signed link prediction where relationships should be predicted as positive (friendship, fan, like, etc) or negative (opposition, anti-fan, dislike, etc). In this problem, feature extraction is an essential step to encode the information needed for prediction. While most current features are based on balance or status theory, we consider the problem of link prediction in the different view of decision-making theory. Our main contribution is a novel feature called Positive-Negative Ratio feature (PNR) which is the ratio between positive and negative links. Our PNR feature, which is based on the strong theory of decision-making, reveals many advantages compared to existing features. It uses a two-dimensional feature but can defeat existing methods at least 3% in classification accuracy and AUC in all three standard databases (Epinions, Slashdot and Wikipedia), even training and testing databases are different. Furthermore, PNR is 5, 1.3 and 1.5 times faster than the other methods in extraction, training and prediction steps, respectively.