Link prediction based on temporal similarity metrics using continuous action set learning automata

Link prediction is a social network research area that tries to predict future links using network structure. The main approaches in this area are based on predicting future links using network structure at a specific period, without considering the links behavior through different periods. For example, a common traditional approach in link prediction calculates a chosen similarity metric for each non-connected link and outputs the links with higher similarity scores as the prediction result. In this paper, we propose a new link prediction method based on temporal similarity metrics and Continuous Action set Learning Automata (CALA). The proposed method takes advantage of using different similarity metrics as well as different time periods. In the proposed algorithm, we try to model the link prediction problem as a noisy optimization problem and use a team of CALAs to solve the noisy optimization problem. CALA is a reinforcement based optimization tool which tries to learn the optimal behavior from the environment feedbacks. To determine the importance of different periods and similarity metrics on the prediction result, we define a coefficient for each of different periods and similarity metrics and use a CALA for each coefficient. Each CALA tries to learn the true value of the corresponding coefficient. Final link prediction is obtained from a combination of different similarity metrics in different times based on the obtained coefficients. The link prediction results reported here show satisfactory of the proposed method for some social network data sets.