A promotive structural balance model based on reinforcement learning for signed social networks

To solve the structural balance problem in signed social networks, a number of structural balance models have been developed. However, these models neglect the effect of the number of nodes are connected to the changed edges, which is not consistent with the practical requirement of social network systems. For this issue, we propose a novel structural balance model, which jointly takes the minimization of the number of changed edges and the number of nodes connected to the changed edges into account. Then, to optimize the proposed model, we design a novel algorithm based on reinforcement learning, which is a first attempt to use reinforcement learning for structural balance problem. Since nodes in a network don't need to be identified by specific states when solving structural balance problem, a stateless Q-learning is adopted. Furthermore, a policy improvement operator is incorporated into the stateless Q-learning to enhance its ability in exploring solutions in a complex search space. Experimental results on the six networks show that the proposed algorithm has dominance in terms of optimal solutions, stability, and convergence against the other comparison algorithms.