Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Contrastive Continual Multi-view Clustering with Filtered Structural Fusion

Xinhang Wan, Jiyuan Liu, Hao Yu, Ao Li, Xinwang Liu, Ke Liang, Zhibin Dong, En Zhu

TL;DR

This paper tackles the problem of clustering data when views arrive sequentially and prior raw views cannot be stored, causing catastrophic forgetting in multi-view clustering. It introduces CCMVC-FSF, a method that uses a fixed-size buffer to store filtered structural information and a clustering-then-sample strategy to guide contrastive learning for updating a robust partition matrix, with a two-step alternating optimization and convergence guarantees. Key contributions include the fixed data buffer design, an efficient positive/negative sample generation scheme, a contrastive loss that leverages prior knowledge, and theoretical links to semi-supervised learning and knowledge distillation, plus extensive experiments showing robustness to CFP and scalability. The approach is privacy- and memory-friendly and well-suited for streaming, real-time multi-view data scenarios.

Abstract

Multi-view clustering thrives in applications where views are collected in advance by extracting consistent and complementary information among views. However, it overlooks scenarios where data views are collected sequentially, i.e., real-time data. Due to privacy issues or memory burden, previous views are not available with time in these situations. Some methods are proposed to handle it but are trapped in a stability-plasticity dilemma. In specific, these methods undergo a catastrophic forgetting of prior knowledge when a new view is attained. Such a catastrophic forgetting problem (CFP) would cause the consistent and complementary information hard to get and affect the clustering performance. To tackle this, we propose a novel method termed Contrastive Continual Multi-view Clustering with Filtered Structural Fusion (CCMVC-FSF). Precisely, considering that data correlations play a vital role in clustering and prior knowledge ought to guide the clustering process of a new view, we develop a data buffer with fixed size to store filtered structural information and utilize it to guide the generation of a robust partition matrix via contrastive learning. Furthermore, we theoretically connect CCMVC-FSF with semi-supervised learning and knowledge distillation. Extensive experiments exhibit the excellence of the proposed method.

Contrastive Continual Multi-view Clustering with Filtered Structural Fusion

TL;DR

This paper tackles the problem of clustering data when views arrive sequentially and prior raw views cannot be stored, causing catastrophic forgetting in multi-view clustering. It introduces CCMVC-FSF, a method that uses a fixed-size buffer to store filtered structural information and a clustering-then-sample strategy to guide contrastive learning for updating a robust partition matrix, with a two-step alternating optimization and convergence guarantees. Key contributions include the fixed data buffer design, an efficient positive/negative sample generation scheme, a contrastive loss that leverages prior knowledge, and theoretical links to semi-supervised learning and knowledge distillation, plus extensive experiments showing robustness to CFP and scalability. The approach is privacy- and memory-friendly and well-suited for streaming, real-time multi-view data scenarios.

Abstract

Multi-view clustering thrives in applications where views are collected in advance by extracting consistent and complementary information among views. However, it overlooks scenarios where data views are collected sequentially, i.e., real-time data. Due to privacy issues or memory burden, previous views are not available with time in these situations. Some methods are proposed to handle it but are trapped in a stability-plasticity dilemma. In specific, these methods undergo a catastrophic forgetting of prior knowledge when a new view is attained. Such a catastrophic forgetting problem (CFP) would cause the consistent and complementary information hard to get and affect the clustering performance. To tackle this, we propose a novel method termed Contrastive Continual Multi-view Clustering with Filtered Structural Fusion (CCMVC-FSF). Precisely, considering that data correlations play a vital role in clustering and prior knowledge ought to guide the clustering process of a new view, we develop a data buffer with fixed size to store filtered structural information and utilize it to guide the generation of a robust partition matrix via contrastive learning. Furthermore, we theoretically connect CCMVC-FSF with semi-supervised learning and knowledge distillation. Extensive experiments exhibit the excellence of the proposed method.
Paper Structure (32 sections, 4 theorems, 23 equations, 7 figures, 5 tables, 3 algorithms)

This paper contains 32 sections, 4 theorems, 23 equations, 7 figures, 5 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Multiple Kernel Clustering
  4. Late Fusion Multi-view Clustering
  5. Continual Multi-view Clustering
  6. Contrastive Continual Multi-view Clustering with Filtered Structural Fusion
  7. The Design of A Fixed Data Buffer
  8. Clustering then Sample
  9. Contrastive Loss
  10. Formulation of Our Algorithm
  11. Optimization
  12. --Subproblem
  13. --Subproblem
  14. Convergence Analysis
  15. - Subproblem
  16. ...and 17 more sections

Key Result

Theorem 1

Assume that $S=\left(x_1, \ldots, x_n\right)$ be a finite population with $n$ points, and $\tilde{S}_i=\left\{y_j\right\}_{j=1}^r$ denote a set randomly selected from $S$ without replacement, then we derive that holds with probability at least $1-\delta$. The proof is given in the appendix.

Figures (7)

  • Figure 1: The basic framework of our proposed algorithm. The filtered structural information $\mathbf{C}_{t-1}$ of previous views is stored and updated in a buffer with fixed size. When the $t$-th view is collected, the model first extracts and updates the filtered structural information (FSI) to attain $\mathbf{C}_{t}$, then the consensus information optimization (CIO) is conducted with three parts of information, i.e., the partition matrix $\mathbf{H}_t$, the consensus matrix $\mathbf{H}^*_{t-1}$ of previous views, and the updated filtered structural information $\mathbf{C}_t$.
  • Figure 2: The objective values of CCMVC-FSF vary with iterations on 3Sources, Mfeat, Reuters, and YTB10.
  • Figure 3: Visualization of the similarity matrix computed by $\mathbf{H}^*_{t}{\mathbf{H}^*_{t}}^{\top}$ on Mfeat varies with iterations.
  • Figure 4: The sensitivity of CCMVC-FSF with the variation of $\lambda$ on 3Sources, Mfeat, Reuters, and YTB10.
  • Figure 5: The effect on CFP with views collected and fused in order on 3Sources, Mfeat, Reuters, and YTB10 in terms of ACC, respectively.
  • ...and 2 more figures

Theorems & Definitions (8)

  • Remark 1
  • Theorem 1
  • Theorem 2
  • Remark 2
  • Lemma 1
  • Remark 3
  • Remark 4
  • Theorem 3