Researchers propose DiffLift to boost topological neural networks

Researchers from the University of São Paulo, Federal University of Ceará, and Simula Metropolitan Center for Digital Engineering have introduced DiffLift (∂lift), a differentiable framework designed to automate the identification of higher-order structures in graphs for topological neural networks (TNNs). Existing TNNs rely on unsupervised graph lifting to extract structures like cycles and cliques, but these methods are static and can limit performance. DiffLift addresses this by learning vertex-level latent representations to parameterize and include higher-order cells dynamically, enabling end-to-end training alongside TNN architectures.

In experiments across multiple benchmarks for graph and node classification, DiffLift outperformed traditional static lifting methods by up to 45%. The framework is scalable and compatible with any TNN, offering a plug-and-play solution for leveraging high-order graph structures. The work was presented at ICLR 2026 and published as a conference paper, signaling its significance in advancing TNNs and graph representation learning.