This paper introduces a novel Hierarchical Graph Neural Network (HGNN) model for stock type prediction, addressing the challenge of capturing multi-level information and relational structures in the stock market. The model constructs a stock industry relationship graph to represent inter-stock dependencies and extracts temporal information from historical price sequences using a temporal attention aggregator. A graph convolution operation is designed to model macro market states, integrating these features into a comprehensive prediction framework. The approach effectively utilizes stock relationship data and hierarchical attributes, which are often overlooked in traditional models. The results demonstrate that the HGNN model outperforms baseline methods, highlighting its ability to capture both micro-level stock characteristics and macro-level market dynamics. This work contributes to the field of quantitative finance by providing a robust method for stock type classification that leverages graph neural networks and temporal attention mechanisms.