Transfer learning using graph-based label transfer (GLT)

Transfer learning aims to leverage the learned knowledge from the labeled source domain and transfer it to an unlabeled target domain. Unsupervised domain adaptation (UDA), a core challenge in transfer learning, typically involves aligning distributions across domains to reduce the domain shift. Although most existing UDA methods rely on deep neural networks, they require extensive end-to-end training, consume substantial computational resources, lack interpretability, and are prone to overfitting. In this work, we propose a graph-based label transfer (GLT) method built upon GraphHop, a non-parametric and label-efficient framework developed initially for label propagation. GLT constructs label-wise adaptive k-NN graphs based on aligned features and performs iterative self-training through entropy-aware pseudo-label refinement. We introduce a multifold validation-driven entropy filtering mechanism that adaptively selects high-confidence target samples across training rounds. This mechanism allows us to progressively expand the labeled set with reliable pseudo-labeled data, improving robustness and generalization. We evaluate GLT on standard domain adaptation benchmarks. Experiments show that it achieves competitive classification accuracy with significantly reduced computational complexity, enabling an intuitive and transparent transfer process that reflects intrinsic geometric relationships among samples.