Continual Multi-View Clustering with Consistent Anchor Guidance

Chao Zhang, Deng Xu, Xiuyi Jia, Chunlin Chen, Huaxiong Li

Abstract

Multi-view clustering (MVC) has recently attracted much attention. Most existing approaches are designed for fixed multi-view data and cannot deal with the common streaming data in the real world. In this paper, we propose a consistent Anchor guided Continual MVC (ACMVC) method in a two-stage way. In the initial learning stage, a low-rank anchor graph-based model is constructed. In the continual learning stage, to leverage historical knowledge, multi-level anchor information is reused to refine the model via adding consistency regularization. This provides prior knowledge to enhance exploration on current data and captures similarity relationships between previous and current data. The proposed model can be optimized efficiently with linear time and space complexity. Experiments demonstrate the effectiveness and efficiency of our method compared with state-of-the-art approaches.

Introduction

In real-world scenarios, data is often dynamic and multi-modal, arriving in continuous streams from various sources such as social media images, text, and video. Traditional Multi-View Clustering (MVC) methods typically assume that all data is available upfront. When new data arrives, these static methods either ignore it or require expensive retraining on the entire accumulated dataset, which is computationally prohibitive and may violate data privacy policies.

This paper introduces ACMVC (Consistent Anchor guided Continual Multi-View Clustering), a framework designed to learn from streaming multi-view data chunks incrementally. By utilizing "anchors" as a bridge between data chunks and applying consistency regularization, ACMVC effectively transfers historical knowledge to new data while maintaining linear time and space complexity, avoiding the "catastrophic forgetting" common in machine learning.

The ACMVC Framework

The ACMVC method operates in two distinct stages to handle streaming data efficiently.

Initial Learning Stage (I-Stage)

When the first data chunk arrives, the model constructs a low-rank anchor graph. Instead of a full sample-to-sample comparison, it identifies a small set of representative "anchors." It then learns a shared similarity matrix ($Z_1$) that captures the relationships between samples and these anchors across all views.

Continual Learning Stage (C-Stage)

As subsequent data chunks arrive, ACMVC reuses the anchors and latent anchor embeddings from previous rounds rather than starting from scratch.

  • Multi-level Knowledge Reuse: A consistency regularization term is applied to ensure that new anchors and their embeddings remain semantically aligned with historical ones.
  • Knowledge Transfer: This alignment serves as a "prior," enabling the model to understand the structure of new data based on past insights while allowing the anchors to adapt slightly to new data distributions.
Overview of the ACMVC framework showing the Initial Stage (I-Stage) and Continual Stage (C-Stage) with knowledge transfer via shared weights and anchor embeddings.

Experimental Evaluation

The authors evaluated ACMVC on seven diverse datasets, including BBCSport, Mfeat, Wiki, and VGGFace, adapted into chunks to simulate streaming.

Accuracy and Scalability

ACMVC consistently achieved top-tier performance. For instance, on the Mfeat dataset, it reached an accuracy of 86.18%, outperforming other online methods. Its linear complexity ($O(n)$) allows it to scale to large datasets like VGGFace (34,000+ samples) in under 50 seconds, whereas traditional offline methods would require a full, time-consuming re-training on the entire accumulated dataset.

Convergence and Efficiency

The optimization algorithm is highly efficient. As shown in the convergence plots, the objective function typically stabilizes within 10 to 15 iterations per data chunk, proving the algorithm's rapid convergence and suitability for real-time applications.

Convergence curves for BBCSport and VGGFace datasets, showing the objective function value decreasing rapidly and stabilizing within a few iterations for both I-stage and C-stage.

Key Insights and Discussion

  • Anchors as Knowledge Bridges: Anchors represent the underlying global cluster structure. Maintaining their consistency across chunks prevents the model from forgetting previously learned patterns.
  • Stability vs. Plasticity: By using regularization (parameter $\beta$) instead of strictly fixing the anchors, the model balances remembering the past with the "plasticity" needed to adapt to new data shifts.
  • Limitations: The current model assumes a fixed cluster count ($c$). Future work may explore class-incremental learning to handle emerging categories and strategies for incomplete multi-view data where some views might be missing for certain samples.

Conclusion

ACMVC provides an industrial-strength solution for continual multi-view clustering, offering a robust balance between accuracy and computational efficiency. By leveraging consistent anchor guidance, it effectively bridges the gap between static clustering methods and the requirements of real-world streaming data.

Summary updated

Abstract

Multi-view clustering (MVC) has recently attracted much attention. Most existing approaches are designed for fixed multi-view data, and cannot deal with the common streaming data in real world. In this paper, we address this problem by proposing a consistent Anchor guided Continual MVC (ACMVC) method in a two-stage way. In initial learning stage, a low-rank anchor graph based model is constructed. In continual learning stage, to leverage the historical knowledge, the multi-level anchor information is reused to refine the model via adding consistency regularization. It not only provides prior knowledge to enhance the exploration on current data, but also captures the similarity relationship between previous and current data, enabling a comprehensive exploitation on streaming data. The proposed model can be optimized efficiently with linear time and space complexity. Experiments demonstrate the effectiveness and efficiency of our method compared with some state-of-the-art approaches.
BibTeX
@inproceedings{zhang2024continual,
  author = {Chao Zhang and Deng Xu and Xiuyi Jia and Chunlin Chen and Huaxiong Li},
  title = {{Continual Multi-View Clustering with Consistent Anchor Guidance}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=5ZGxF6HvBj},
}
Figure