Aggregation-aware MLP: An Unsupervised Approach for Graph Message-passing
This addresses the issue of rigid aggregation in GNNs for graph learning practitioners, offering an unsupervised solution that is incremental by building on existing MLP and graph reconstruction methods.
The paper tackles the problem of fixed aggregation functions in Graph Neural Networks (GNNs) leading to poor performance, especially with heterophily, by proposing an unsupervised framework called Aggregation-aware Multilayer Perceptron (AMLP) that adapts MLPs to aggregation, resulting in superior performance in node clustering and classification tasks.
Graph Neural Networks (GNNs) have become a dominant approach to learning graph representations, primarily because of their message-passing mechanisms. However, GNNs typically adopt a fixed aggregator function such as Mean, Max, or Sum without principled reasoning behind the selection. This rigidity, especially in the presence of heterophily, often leads to poor, problem dependent performance. Although some attempts address this by designing more sophisticated aggregation functions, these methods tend to rely heavily on labeled data, which is often scarce in real-world tasks. In this work, we propose a novel unsupervised framework, "Aggregation-aware Multilayer Perceptron" (AMLP), which shifts the paradigm from directly crafting aggregation functions to making MLP adaptive to aggregation. Our lightweight approach consists of two key steps: First, we utilize a graph reconstruction method that facilitates high-order grouping effects, and second, we employ a single-layer network to encode varying degrees of heterophily, thereby improving the capacity and applicability of the model. Extensive experiments on node clustering and classification demonstrate the superior performance of AMLP, highlighting its potential for diverse graph learning scenarios.