AGLP: A Graph Learning Perspective for Semi-supervised Domain Adaptation
Houcheng Su, Mengzhu Wang, Jiao Li, Nan Yin, Liang Yang, Li Shen
TL;DR
This paper introduces AGLP, a novel graph learning approach for semi-supervised domain adaptation that leverages structural data information to improve domain-invariant feature learning, outperforming existing methods.
Contribution
It is the first to incorporate structural information via graph convolutional networks into SSDA, enhancing adaptation performance.
Findings
AGLP outperforms state-of-the-art SSDA methods on benchmarks.
Structural information propagation improves domain-invariant features.
The approach effectively reduces domain discrepancies.
Abstract
In semi-supervised domain adaptation (SSDA), the model aims to leverage partially labeled target domain data along with a large amount of labeled source domain data to enhance its generalization capability for the target domain. A key advantage of SSDA is its ability to significantly reduce reliance on labeled data, thereby lowering the costs and time associated with data preparation. Most existing SSDA methods utilize information from domain labels and class labels but overlook the structural information of the data. To address this issue, this paper proposes a graph learning perspective (AGLP) for semi-supervised domain adaptation. We apply the graph convolutional network to the instance graph which allows structural information to propagate along the weighted graph edges. The proposed AGLP model has several advantages. First, to the best of our knowledge, this is the first work to…
Peer Reviews
Decision·ICLR 2025 Conference Withdrawn Submission
(1) The paper introduces a graph learning perspective for SSDA, which is the first work to model structural information in SSDA. (2) The proposed AGLP model effectively learns domain-invariant and semantic representations, reducing domain discrepancies in SSDA. (3) The experimental results show that AGLP outperforms state-of-the-art SSDA methods on multiple benchmarks.
(1) The motivation of this manuscript is not clear. The authors should clearly claim the challenging issues in previous methods. (2) While the GNN adopted in the manuscript seems plausible, it is not exciting. Although the authors emphasize that this is the first semi-supervised domain adaptation method using graph networks, graph networks have been widely used in more difficult unsupervised domain adaptation tasks. (3) While the graph module adopted in the manuscript seems plausible, it is not
The manuscript introduces a graph learning perspective to semi-supervised domain adaptation (SSDA) and leverages structural information in instance graphs, which is a relatively novel approach. By using class centroid alignment to reduce inter-class discrepancies, the overall design is quite reasonable. The method demonstrates performance improvements compared to other approaches, which, to some extent, validates its effectiveness.
1. The manuscript proposes using a graph convolutional network (GCN) for structural information propagation; however, it does not clearly explain why this form of structural propagation effectively reduces domain discrepancies. The overall description lacks clarity, and the innovation points are not accurately conveyed. 2. The manuscript proposes constructing a densely connected instance graph using CNN features of samples, connecting them based on the similarity of their structural characterist
1.AGLP introduces a graph-based perspective in SSDA, effectively modeling structural information to address gaps in current SSDA techniques that primarily rely on domain and class labels without structural insights. 2. he model consistently achieves high accuracy across benchmarks, with ablation and comparative analyses that underline the impact of structure-aware and class centroid alignment mechanisms on performance. 3. This work addresses a vital aspect of SSDA—structural alignment—and introd
1. The integration of GCNs and alignment mechanisms increases model complexity and computational overhead, which might present challenges in low-resource environments. Future work could explore ways to optimize computational efficiency. 2. While the graph structure construction is a strength of the approach, further clarification on hyperparameter tuning would enhance understanding, especially concerning batch size variability and data structure adaptation.
1. The manuscript is well-orgnised. It is easy to follow this work. 2. The ablation studies of each module and hyperparameters are sufficient.
1. The novelty of this paper is limited, it seems to be a combination of existing methods. The structure-aware alignment method and class centroid alignment mentioned in the paper are similar to the existing work GCAN [1]. The method of learning data structure information using a graph convolutional network in the structure-aware alignment module, as well as the centroid alignment objective function in the class centroid alignment module are the same as in GCAN [1]. 2. Most of the theories and e
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsDomain Adaptation and Few-Shot Learning