Progressive Graph Learning for Open-Set Domain Adaptation

1. Introduction

• Open-set adaptation allows target data to contain an unknown category absent from the source, addressing a setting that closed-set methods do not assume.Domain shift can arise from changes in lighting, background, environment, or data modality.Source: §1 · p. 1
• Existing open-set approaches lack theoretical analysis of partial risks and the target-risk upper bound, motivating a framework aimed at minimizing those risks.The paper proposes combining four strategies in an end-to-end progressive learning framework.Source: §1 · p. 1
• PGL restricts the shared-class hypothesis space to reduce source misclassification as unknown and progressively moves low-confidence target samples into a pseudo-labeled unknown set.The progressive step is intended to suppress negative transfer from falsely aligning private domain representations.Source: §1 · pp. 1–2
• Episodic graph training aggregates source and target features to align conditional distributions, while an adversarial discriminator aligns known-category marginal distributions.The framework addresses conditional shift at sample and manifold levels in a transductive setting.Source: §1 · p. 2
• PGL is evaluated on Office-Home, VisDA-17, and Syn2Real-O open-set object-recognition benchmarks.The paper reports that it outperforms existing state-of-the-art open-set adaptation approaches.Source: §1 · p. 2

2. Preliminaries

• Closed-set UDA assumes equal source and target label spaces and learns a target classifier from labeled source data and unlabeled target data.The classifier maps target inputs to the shared class set.Source: §2 · p. 2
• Open-set UDA extends the target label space with an unknown class that is absent from the source label space.The target classifier is learned from labeled source and unlabeled target samples.Source: §2 · p. 2
• The target-error bound motivates four improvement directions: reduce source risk, domain discrepancy, shared conditional-shift error, and open-set risk.The open-set risk corresponds to unknown-sample misclassification and weighs most when the unknown fraction is large.Source: §2 · p. 3

3. Progressive Open-Set UDA

• POUDA separates the target classifier for shared classes from a pseudo-labeling function for the unknown class.The framework reformulates open-set UDA progressively to minimize the four partial risks.Source: §3 · p. 3
• At each of M steps, the pseudo-labeling function uses ranked shared-class predictions and an openness-dependent threshold to assign target samples known or unknown labels.The step enlargement factor is α = 1/M, and β measures the target set's unknown-sample ratio.Source: §3 · p. 3
• The authors state that progressive learning can achieve a tighter upper bound than conventional open-set domain adaptation.The POUDA error bound is formulated for a fixed target-set openness β.Source: §3 · p. 3

4. Methodology

• PGL targets conditional shift through episodic support from source and pseudo-labeled target samples, plus GNN aggregation over same-class nodes.The approach also iteratively adjusts thresholds to control progressive open-set risk.Source: §4 · p. 4

4.1. Initial Episodic Training with GNNs

• Each initial episode combines per-class labeled source samples with unlabeled target samples, and the resulting batch graph supports transductive inference.Graph vertices represent source or target features, and edges measure node affinity.Source: §4.1 · p. 4

4.1.1. EDGE UPDATE

• Edge-update layers propagate normalized edge features using a learned nonlinear network with sigmoid and degree-matrix terms.The degree matrix is defined from the layer’s adjacency matrix plus the identity matrix.Source: §4.1.1 · p. 4

4.1.2. NODE UPDATE

• Node updates use neighboring-node information and concatenation in a network built from two convolutional layers, LeakyReLU activations, and dropout.Initial node embeddings come from the backbone representations.Source: §4.1.2 · p. 5

4.1.3. JOINT OPTIMIZATION

• Joint optimization combines adversarial domain alignment with node classification and edge classification losses in the final objective.The node classifier uses focal loss; edge labels indicate whether source nodes share a class.Source: §4.1.3 · p. 5

4.2. Pseudo-Labeling in Progressive Paradigm

• At each step, PGL ranks target samples by maximum shared-class likelihood and expands known and unknown pseudo-labeled sets using a threshold schedule.A large enlarging factor can introduce noisy candidates and reduce the number of steps; a small factor can increase computation while allowing steadier performance gains.Source: §4 · p. 4, §4.2 · p. 5

4.3. Episodic Update with Mix-up Strategy

• Each episodic update probabilistically replaces source samples with pseudo-labeled known target samples, then repeats pseudo-labeling with the new episodes until convergence.The replacement probability varies with the progressive step.Source: §4.3 · pp. 5–6

5.1. Datasets

• The benchmarks define known and unknown classes differently: Office-Home uses 25 known classes, VisDA-17 uses 6 of 12, and Syn2Real-O adds 50k unknown target images.Syn2Real-O has openness 0.9; its known source and target data follow the VisDA-17 protocol.Source: §5.1 · p. 6

5.2. Results and Analysis

• PGL improves mean OS∗ accuracy by 8.1%, 28.0% and 29.6% on Office-Home, Syn2Real-O and VisDA-17, respectively, over state-of-the-art results.The comparisons use ResNet-50 for Office-Home and Syn2Real-O and VGGNet for VisDA-17.Source: §5.2 · pp. 6–7, Table 2 · p. 7, Table 3 · p. 7
• On Syn2Real-O, removing progressive learning drops OS by 16.9%, replacing focal loss with NLL lowers OS from 65.5% to 56.9%, and removing GNNs drops OS by 12.5%.Removing mix-up lowers OS by 4.6%; the authors associate progressive pseudo-labeling with reduced conditional-shift effects and GNNs with class-wise manifold learning.Source: Table 5 · p. 7, §5.2 · pp. 7–8
• Smaller enlarging factor α performs consistently better on Syn2Real-O and Office-Home, while requiring more training time.The authors say slower enlargement yields more reliable pseudo-labeled candidates for shared-classifier learning and can prevent error accumulation.Source: Table 6 · p. 7, §5.2 · p. 8
• t-SNE analyses report clearer class-wise boundaries and alignment of same-class source and target representations with PGL, while learned edge values resemble the ground-truth edge labels.The visualizations cover Office-Home and Syn2Real-O; the edge-map comparison is on Syn2Real-O.Source: Figure 3 · p. 8, §5.2 · pp. 8–9, Figure 5 · p. 9
• PGL consistently outperforms OSBP and STA as Syn2Real-O openness varies across 0.25, 0.5, 0.75 and 0.9.The paper notes that OSBP and PGL set a hyperparameter to control openness, whereas STA generates soft weights adversarially and shows performance fluctuation.Source: Figure 4 · p. 8, §5.2 · p. 8
• With loss coefficients µ and γ varied over (0, 2], OS is stable within [0.7, 2]; taking µ toward zero drops performance by 4.6%, and taking γ toward zero drops it by 10.2%.The authors interpret these drops as evidence for edge supervision and adversarial learning.Source: §5.2 · p. 8
• Across training steps, ALL continues rising while OS and OS∗ later decline; the authors stop at step 7 on Office-Home and step 5 on Syn2Real-O.The reported trade-off reflects increasing correct unknown classifications alongside noise from an enlarging pseudo-labeled set.Source: Figure 7 · p. 9, §5.2 · p. 9

• PGL addresses open-set domain shift at both sample and manifold levels by controlling open-set risk.Source: §6 · p. 9
• PGL performs consistently well on challenging open-set object-recognition benchmarks with significant domain discrepancy and conditional shifts.Source: §6 · p. 9