Many real-world applications can be modeled as signed directed graphs wherein the links between nodes can have either positive or negative signs. Social networks can be modeled as signed directed graphs where positive/negative links represent trust/distrust relationships between users. In order to predict user behavior in social networks, several studies have addressed the link-sign prediction problem that predicts a link sign as positive or negative. However, the existing approaches do not take into account the time when the links were added which plays an important role in understanding the user relationships. Moreover, most of the existing approaches require the complete network information which is not realistic in modern social networks. Last but not least, these approaches are not adapted for dynamic networks and the link-sign prediction algorithms have to be reapplied each time the network changes. In this paper, we study the problem of link-sign prediction by combining random walks for graph sampling, Doc2Vec for node vectorization and Recurrent Neural Networks for prediction. The approach requires only local information and can be trained incrementally. Our experiments on the same datasets as state-of-the-art approaches show an improved prediction.