A two-stage representation-learning method called ACGRL recorded the highest reported clustering accuracy on all four benchmark datasets tested. It beat the strongest competing method by 5.40 percentage points on NGs, 0.15 on Hdigit, 12.88 on BBCSport and 9.34 on Cora. The result appears in an arXiv preprint, arXiv:2609.35212v1, dated 28 September 2026.
The study asks whether a model can improve clustering across multiple views of the same data while separating information shared across views from information specific to one view. It also targets a reduction in view-identifying information in the shared representation.
ACGRL uses two stages. Stage I learns view-invariant representations. Stage II freezes those representations and uses them as consistency supervision while learning view-specific information. The resulting representations are then integrated for reconstruction and clustering.
A lead across four tests
Tests used four benchmark datasets. NGs contained 500 samples across three views and five classes; BBCSport had 544 samples, two views and five classes; Cora had 2,708 samples, two views and seven classes; and Hdigit had 10,000 samples, two views and 10 classes.
ACGRL was compared with nine representative multi-view clustering methods, using baseline settings recommended in the respective papers. The study assessed clustering with three reported measures: ACC, ARI and NMI.
ACGRL's lead was uneven across the benchmarks. It was narrowest on Hdigit at 0.15 percentage points and widest on BBCSport at 12.88, with margins of 5.40 on NGs and 9.34 on Cora.
Looking across all three measures, the reported ACC/ARI/NMI triplets were 97.00/92.26/93.31 for NGs, 99.95/99.89/99.86 for Hdigit, 93.58/87.34/85.10 for BBCSport and 64.58/39.58/43.61 for Cora.
Testing the parts of the method
To examine the contribution of each part, the researchers removed each of four objectives in turn. On both Cora and BBCSport, every such removal reduced ACC, ARI and NMI. Separate component tests on NGs and BBCSport also found the full ACGRL model ahead of MSE-based, MI-based and w/o Consistency Supervision variants.
Another analysis used CCSI to compare representation similarity within a view and across views. ACGRL had higher CCSI values than DMVCS and DFL-NET in both settings across all four datasets. The paper interprets higher CCSI as lower average similarity under its selected measures, but the score is not direct proof of semantic disentanglement.
t-SNE visualizations offered a qualitative view of the learned features. On Cora, they showed clearer cluster structure than raw features, including improved separation between clusters. On Hdigit, the learned features formed more compact and distinguishable clusters.
What the result does not settle
The findings remain a benchmark report. The comparison covers four datasets and nine representative methods, and the supplied analysis reports no confidence intervals, repeated-run variability or inferential tests for the performance differences. That leaves the stability and uncertainty of the reported gaps unresolved.
The study used publicly available benchmark datasets, involved no human or animal subjects and required no ethical approval. It reports support from the Natural Science Foundation of Fujian Province, China, under Grants 2026J001408 and 2024J01067, and the National Natural Science Foundation of China under Grant U1805263. The authors reported no relevant financial or non-financial interests.
Paper data and sources
Original title: Adversarial Consistency-Guided Representation Learning for Multi-view Clustering
Authors: Yuchen Lin, Kunpeng Xu, Ying Fang, Lifei Chen
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-28
DOI: Not available
Original paper · Full text