In the traditional multi label research, they can only roughly locate the semantic regions of the image, and can not fully excavate the label correlation between the semantic regions. To solve the above problems, the author proposes a Semantic Attention Graph Representation (SAGR) algorithm that composed of two key modules for multi-label classification : 1) Semantic Location(SL) module that integrated the semantic information of all labels categories in the image for learning to obtain the feature representation of each label category; 2) Semantic Correlation(SC) module that used graph structure to interact with the obtained semantic feature representation, and captured the dynamic label dependency in image by graph attention network. The experimental results of Pascal VOC2007 and MirFlickr25k datasets show that SAGR algorithm is better than traditional methods, and the mAP of SAGR can be improved to 93.5% and 84.2%.