Sina Weibo has evolved into a daily social tool for people, yet effectively leveraging its data for sentiment analysis remains a challenging task due to the presence of information beyond text, such as emojis or images. In this paper, we propose an attention graph convolutional network (AGCN) for fine-grained sentiment classification of Weibo posts. Utilizing an attention network based on cosine similarity, the rich emotional information embedded in emoji features interacts with the textual content, effectively enhancing the capability to represent emotions in the text. Leveraging the characteristics of attention networks to construct a graph structure effectively enables graph convolutional networks to capture higher-order relationships between words in textual features. This approach addresses the challenge of extracting sentiment tendencies from Weibo comments. Experimental results on public data sets demonstrate the effectiveness of AGCNs.