Constructing Structured Decision Sources for Consensus-Based Pseudo-Label Learning

Long Wang

cs.LGOct 8, 2026

Constructing Structured Decision Sources for Consensus-Based Pseudo-Label Learning

Authors: Long Wang

Abstract

Consensus can make pseudo-label learning more reliable, but only when its predictors contribute genuinely different evidence. Multiple models that repeat the same boundary provide additional votes without additional information. We address this problem by con structing decision sources through controlled changes to within-class structure. Starting from a shared graph representation, we vary center granularity and neighborhood mixing, reproduce each resulting source to test its stability, and select a complementary subset using node pair coassignment. Unanimous predictions from the selected sources are then ranked for student training. On the public fixed splits of Cora, CiteSeer, and PubMed, evaluated with five random seeds, the constructed sources improve fixed-budget training pseudo-label precision by 1.19 to 4.39 percentage points over three conventionally initialized GCN sources. Under matched structural filters, three-source consensus is more precise than each constituent source in all 45 dataset slot eed comparisons. The gains are strongest in pseudo-label quality: downstream accuracy remains competitive but does not lead on every dataset. These results identify source construction rather than model count alone as an important design problem for consensus-based pseudo-label learning.

Figures & tables

Figure 1: Overview of structured decision-source construction. Structural interventions $(k,\beta)$ generate four candidates from a shared warm-start representation. Reproduction stability and between-source co-assignment differences select three sources. Their consensus candidates enter either the q90 quality diagnostic or minimum-confidence ranking for student training; the q90 gate is not used in student training.

Figure 2: Structural intervention, co-assignment probing, and reproduction-based selection. Varying $k$ and $\beta$ creates different within-class partitions. Co-assignment matrices remove center-label permutations, and within-source stability $s_{\mathrm{reg}}$ and between-source difference $d_{\mathrm{co}}$ select three stable, nonduplicate sources. The illustrated nodes and matrices do not represent experimental samples.

Dataset	Nodes	Edge entries	Features	Classes	Train/val/test	$\|U\|$
Cora	2708	10556	1433	7	140/500/1000	2568
CiteSeer	3327	9104	3703	6	120/500/1000	3207
PubMed	19717	88648	500	3	60/500/1000	19657

Table 1: Statistics and fixed masks for the three Planetoid public datasets. Edges are directed PyG edge entries.

Dataset	Voting rule	Before $N$	Before prec.	After $N$	After prec.	After cov.	Removed correct/incorrect
Cora	Structured sources 3/3	2473.6	84.27	1452.4	92.53	56.56	741.0/280.2
	DiPat 3/3 + structural gate	2437.4	83.23	1416.4	93.47	55.16	705.0/316.0
	Joint 4/6	2388.4	85.36	1431.8	93.38	55.76	702.2/254.4
	Joint 5/6	2298.8	86.87	1415.4	93.86	55.12	668.8/214.6
	Joint 6/6	2208.8	88.21	1396.8	94.19	54.39	632.8/179.2
CiteSeer	Structured sources 3/3	3081.2	73.40	1322.0	85.86	41.22	1126.8/632.4

Table 2: Pseudo-label quality before and after q90 filtering (five-seed means). Precision and coverage are percentages. Removed correct/incorrect gives the numbers removed by q90. Within each dataset, bold marks the largest post-gate precision and the largest post-gate coverage.

Figure 3: Precision–coverage tradeoff after the common q90 structural gate (five-seed means). Lines connect adjacent voting strictness levels only to show direction and are not continuous threshold curves.

Dataset	Method	Training pseudo-labels	Test Accuracy	Test Macro-F1
Cora	Supervised GCN	0	$81.66\pm 0.82$	$80.51\pm 0.82$
	Graph-only DiPat adaptation	1734	$82.48\pm 0.33$	$81.31\pm 0.47$
	Structured sources	1734	$\bm{84.04\pm 1.12}$	$\bm{83.03\pm 0.89}$
CiteSeer	Supervised GCN	0	$71.46\pm 0.66$	$\bm{68.29\pm 0.62}$
	Graph-only DiPat adaptation	2165	$73.54\pm 0.66$	$67.26\pm 0.61$
	Structured sources	2165	$\bm{73.74\pm 0.63}$	$\bm{68.29\pm 1.16}$

Table 3: Downstream results under the shared local protocol (Accuracy/F1 in percent; five-seed mean

\pm

sample standard deviation). Students do not use the q90 gate. Within each dataset and metric, bold marks the largest mean; ties are included.

Dataset	Decision sources	All-vote prec.	q90 prec./cov.	Training PL prec.	Test Acc.	Test Macro-F1
Cora	Structured sources	84.27	92.53 /56.56	93.04	84.04	83.03
	Three random-seed GCNs	84.80	91.94/57.06	91.30	83.30	81.95
CiteSeer	Structured sources	73.40	85.86 /41.22	83.71	73.74	68.29
	Three random-seed GCNs	71.55	84.63/42.03	79.33	73.08	69.21
PubMed	Structured sources	79.36	84.35 /71.75	88.18	79.94	79.53
	Three random-seed GCNs	79.96	84.06/71.10	87.00	79.82	79.35

Table 4: Replacement ablation against three conventionally initialized GCN sources (five-seed means, percent). Both post-gate rows use the same additional q90 structural gate; task training does not use the gate. Within each dataset and metric, bold marks the larger value; q90 coverage is not ranked because it trades off against precision.

Figure 4: Pseudo-label quality for structured sources and three random-seed GCNs (five-seed means). Left: conditional precision after the shared q90 structural gate. Right: fixed-budget training pseudo-label precision without q90 filtering. The panels evaluate different branches and should not be compared numerically across panels.

Dataset	Selected slot 0	Selected slot 1	Selected slot 2
Cora	+0.61 (5/5)	+0.80 (5/5)	+0.42 (5/5)
CiteSeer	+0.66 (5/5)	+0.40 (5/5)	+0.38 (5/5)
PubMed	+0.34 (5/5)	+0.25 (5/5)	+0.32 (5/5)

Table 5: Post-gate precision gain of unanimous 3/3 voting over one source under the same single structural gate (percentage points; five-seed mean). Parentheses show seeds with positive gain.

Dataset	Mean reproduction stability	Mean co-assignment distance	Distance range
Cora	0.833	0.222	0.140–0.303
CiteSeer	0.814	0.263	0.128–0.412
PubMed	0.918	0.234	0.088–0.329

Table 6: Shared-pair structural probes for selected sources (five seeds; each statistic summarizes 15 values). Distances and stability lie in

[0,1]

Explore similar work

Stop Replacing Noise with Noise: Two-Source Reliability Assessment for Label Correction and Sample Reweighting in Label-Noise Learning
Aug 4, 2026Wenxiao Fan, Kan LiNoisy-Label Learning Pseudo-Labeling
School of Computer Science, Beijing Institute of Technology
Fractional Heat Kernel for Semi-Supervised Graph Learning with Small Training Sample Size
Oct 6, 2025Farid Bozorgnia, Vyacheslav Kungurtsev, Shirali Kadyrov +1Graph-Based Semi-Supervised Learning Graph Diffusion
Safe-Subspace Pseudo-Label Refinement for Source-Free Graph Domain Adaptation
May 30, 2026Yingxu Wang, Xinwang Liu, Siyang Gao +1Source-Free Domain Adaptation Pseudo-Labeling
Department of Computer Science and Engineering, Chinese University of Hong Kong · College of Computer, National University of Defense Technology, Changsha, 410073, China · Department of Data Science, City University of Hong Kong +1

Aug 4, 2026, cs.LGJ/K move · Enter open · S save

Stop Replacing Noise with Noise: Two-Source Reliability Assessment for Label Correction and Sample Reweighting in Label-Noise Learning

Wenxiao Fan, Kan Li

School of Computer Science, Beijing Institute of Technology

Noisy-Label Learning Pseudo-Labeling

Open paper AskPDF arXiv

Refurbishment-based noisy-label learning mixes an observed label with a model-derived pseudo target, typically using one sample-wise cleanliness score to control both branches. This creates a hidden coupling: reducing trust in the observed label automatically increases trust in the pseudo target. We show that this complementarity can replace one unreliable signal with another because a pseudo target learned from corrupted supervision may reproduce the noise it is meant to correct. Our representation diagnostics provide a consistent account of this mismatch: noisy supervision redirects deeper layers more strongly, whereas shallower relations remain comparatively stable and provide information beyond the loss posterior. We therefore propose TRACE, a Two-Source Reliability Assessment framework for Label Correction and Sample Reweighting. TRACE assesses the observed label using loss fit, shallow relation stability, and prediction agreement, while separately assessing the pseudo target using model confidence. Its source-specific scores control target correction and supervision strength without assuming complementary reliability. Across synthetic and real-world noisy benchmarks, TRACE improves representative refurbishment baselines and yields more reliable pseudo supervision.

Abstract

Explore similar work

Stop Replacing Noise with Noise: Two-Source Reliability Assessment for Label Correction and Sample Reweighting in Label-Noise Learning

Fractional Heat Kernel for Semi-Supervised Graph Learning with Small Training Sample Size

Safe-Subspace Pseudo-Label Refinement for Source-Free Graph Domain Adaptation