Spatial convolutional models of Graph Neural Networks (GNNs) updates embeddings of nodes by the neighborhood aggregation, it has obvious advantages in reducing time complexity and improving accuracy. Therefore, it is a very important task to change the way of neighborhood aggregation to learn better node embeddings. Attention mechanisms are usually used to assign trainable weights to nodes in neighborhood aggregation, so that the node influence can also participate in the process of neighborhood aggregation. We propose the ATT-MLP model that combines attention mechanism and multi-layer perception(MLP), and applys node attention weights to graph pooling. Experiments on graph prediction show that our algorithm performs better than other baselines on commonly used datasets.