New Graph Neural Network Excels at Untangling Complex Multimedia Data

Graph neural networks (GNNs) have emerged as powerful tools for processing graph-structured data, finding wide application in fields like multimedia information retrieval. However, a significant limitation of many traditional GNN models is their effectiveness primarily with homogeneous graphs, which possess simpler structures and uniform semantics. When confronted with the rich, varied interactions and diverse semantics characteristic of heterogeneous multimedia data, the performance of these conventional models tends to decline significantly.

To overcome this challenge, a new heterogeneous graph neural network algorithm, dubbed MPRW-HGNN, has been developed, leveraging a meta-path random walk strategy. The core of this innovation lies in its multi-stage processing. Initially, a dedicated module generates meta-path instances and their structural representations through random walks, enabling the model to capture the underlying relationships more effectively. Subsequently, a soft-attention mechanism is employed to meticulously fuse information derived from multiple meta-paths, allowing for a finer-grained understanding of the semantic structures around individual nodes.

Further enhancing its capabilities, MPRW-HGNN incorporates a self-attention mechanism. This component is crucial for exploring the semantic correlations and differences that exist between multiple paths, facilitating an adaptive weighted fusion of this multi-path information. The result is the generation of robust node features specifically tailored for heterogeneous graph data. Rigorous experiments conducted on widely recognized datasets, including the IMDB movie dataset and the DBLP academic paper dataset, have demonstrated its effectiveness. As lead author Pei notes in the paper, "Through extensive experiments on information retrieval tasks using the IMDB movie dataset and DBLP academic paper dataset, we demonstrate the significant advantages of our proposed algorithm in handling multimodal data and improving retrieval accuracy."