Towards multi-order homophily-aware propagation in multi-view learning

Multi-view learning aims to enhance performance by integrating information from different data perspectives. Existing multi-view learning methods based on graph neural networks often treat all views equally, regardless of their varying quality, and overlook the heterophily of the constructed similarity matrix. This can lead to over-propagation of noise during the aggregation process, which negatively impacts the final multi-view integration and results in sub-optimal performance. This paper proposes a multi-order homophily-aware multi-view learning to mitigate the propagation of heterophily information on the graph. The proposed method consists of two main components: the multi-order decoupling module and the homophily-aware propagation module. The former encodes the multi-order neighborhood and self-embedding independently to capture the neighborhood information at a fine-grained level. The latter constructs a homophily-aware matrix and calculates the homophily contribution of different order neighborhoods as the propagation confidence to guide the aggregation of node information. Experimental results demonstrate that the proposed method outperforms existing state-of-the-art models. The code is released at https://github.com/dynammor/MHML.