Abstract: To enhance the learning ability of graph neural networks for node representation on heterophily graphs, a simple and effective combinatorial filter learning module from the perspective of node-neighborhood differences is proposed, which allows the filter parameters to be learned adaptively on a node-wise basis. Specifically, this module learns the low-pass and high-pass filter coefficients of the filters via pairwise similarity measurement between a node and its neighbors, and later enables the nodes to learn more robust higher-order representations through dimensional adaptive gating signal fusion intermediate layer learning. The node representation of heterophily graphs is better learned by adaptively adjusting the proportion of nodes acquiring low-frequency and high-frequency information and the aggregation of positive and negative signals. Experiment results show that the method proposed outperforms in node classification task compared with other state-of-the-art methods and effectively alleviates the phenomenon of oversmoothing.