Anchor Learning Framework Boosts Multi-View Clustering Performance Across Thirteen Datasets

A new machine learning framework termed Cross-View Consistent Representation via Anchor Learning for Multi-View Clustering, or CCR-MVC, demonstrates significant improvements in both clustering performance and computational efficiency across 13 benchmark datasets. Published in the journal Applied Intelligence, the study establishes that CCR-MVC ranks first overall across four primary evaluation metrics, namely clustering accuracy, Normalized Mutual Information, Purity, and Fscore, according to comprehensive statistical rank tests comparing the technique against existing multi-view clustering alternatives.

In terms of clustering accuracy measured by the ACC metric, the newly proposed method achieved the best or second-best results across nine of the 13 evaluated datasets. The performance advantages were particularly pronounced on challenging benchmark collections, such as the bbcsport dataset, where CCR-MVC outperformed the second-best comparative methods by an accuracy margin of 26.89 percent. These quantitative comparisons highlight the model’s capacity to resolve complex grouping patterns in datasets where competing clustering techniques encounter substantial difficulty.

The research specifically addresses key technical challenges that arise within anchor-based multi-view clustering algorithms, a family of methods that has attracted substantial attention primarily due to their operational efficiency when processing complex data. Despite their processing advantages, existing anchor-based approaches struggle because individual views inherently possess distinct feature dimensions, varying data distributions, and disparate levels of discriminative ability. Consequently, the specific anchors selected within each separate view can yield conflicting and inconsistent expressions of what should be the same underlying latent consensus representation.

Because specific anchors across disparate views produce inconsistent representations of the shared latent space, traditional clustering pipelines struggle to accurately depict the common clustering structure shared among the multiple views. The local deviations and imbalances present in an individual view can introduce substantial interference, misaligning the learned consensus structure and degrading the overall quality of multi-view data integration.

To overcome this limitation, the authors introduced CCR-MVC to jointly learn the anchor matrix of each individual view and the shared latent consensus representation within a unified optimization framework. By structuring the learning process in a single unified architecture, the method directly coordinates the anchor representations across all views, enabling the model to comprehensively integrate both the complementary information unique to particular views and the common information shared across the entire collection of views.

In addition to joint anchor matrix learning, the methodology introduces a novel cross-view consistency constraint designed to enforce harmony across differing perspectives. This mathematical constraint ensures that different views consistently express the identical latent consensus representation, thereby suppressing the disruptive interference of local deviations stemming from any single view. As a direct result of this regularization, the framework can more accurately extract and isolate the underlying clustering structure shared across multiple views.

Beyond clustering accuracy and statistical ranking on standard datasets, the experimental evaluation examined the computational scalability of CCR-MVC when applied to large-scale data environments. In these large-scale tests, the method demonstrated strong computational scalability by achieving the lowest running time among all compared clustering methods while simultaneously maintaining high clustering accuracy, confirming that its consistent anchor formulation provides significant speed advantages alongside superior grouping accuracy.

Subject of Research: Multi-view clustering via anchor learning and cross-view consistency constraints

Article Title: Cross-view consistent representation via anchor learning method for multi-view clustering

Article References: Bai, L., Xue, Q., Liu, Y., Jiang, X., & Pan, X. (2026). Cross-view consistent representation via anchor learning method for multi-view clustering. Applied Intelligence, 56(16), Article 534. https://doi.org/10.1007/s10489-026-07566-7

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07566-7

Keywords: multi-view clustering, anchor-based clustering, CCR-MVC, anchor learning, cross-view consistency constraint, latent consensus representation, anchor matrix, data distribution, discriminative ability, feature dimensions, bbcsport dataset, ACC metric, Normalized Mutual Information, Purity metric, Fscore metric, statistical rank tests, computational scalability, large-scale datasets