Generalizing neural networks to graph data is one of the recent important challenges in machine learning. However, most existing works is on graphs that simply encode whether relationships exist or not, and they can not be applied to signed graphs, which relationships between nodes can be positive (“like”, “trust”) or negative (“dislike”, “distrust”). Proven usefulness of the analysis of signed graphs in many real-world applications prompted us to propose a new method to deal with this type of graphs. To be more specific, we propose a neural network approach for semi-supervised node classification in the signed network. Our approach incorporates the information in negative links with the positive ones without making structural balance theory assumptions. In our method, usage of several hidden layer neural network facilitates node classification by allowing each node to borrow information from not only its immediate neighbors but also other nodes in the graph. By conducting experiments on some real-world interesting graphs, we demonstrate the effectiveness of our algorithm compared to the current state of the art methods.