Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep Contrastive Multi-view Clustering under Semantic Feature Guidance

Siwen Liu, Jinyan Liu, Hanning Yuan, Qi Li, Jing Geng, Ziqiang Yuan, Huaxu Han

TL;DR

A multi-view clustering framework named Deep Contrastive Multi-view Clustering under Semantic feature guidance (DCMCS) to alleviate the influence of false negative pairs by minimizing instance-level contrastive loss weighted by semantic similarity.

Abstract

Contrastive learning has achieved promising performance in the field of multi-view clustering recently. However, the positive and negative sample construction mechanisms ignoring semantic consistency lead to false negative pairs, limiting the performance of existing algorithms from further improvement. To solve this problem, we propose a multi-view clustering framework named Deep Contrastive Multi-view Clustering under Semantic feature guidance (DCMCS) to alleviate the influence of false negative pairs. Specifically, view-specific features are firstly extracted from raw features and fused to obtain fusion view features according to view importance. To mitigate the interference of view-private information, specific view and fusion view semantic features are learned by cluster-level contrastive learning and concatenated to measure the semantic similarity of instances. By minimizing instance-level contrastive loss weighted by semantic similarity, DCMCS adaptively weakens contrastive leaning between false negative pairs. Experimental results on several public datasets demonstrate the proposed framework outperforms the state-of-the-art methods.

Deep Contrastive Multi-view Clustering under Semantic Feature Guidance

TL;DR

A multi-view clustering framework named Deep Contrastive Multi-view Clustering under Semantic feature guidance (DCMCS) to alleviate the influence of false negative pairs by minimizing instance-level contrastive loss weighted by semantic similarity.

Abstract

Contrastive learning has achieved promising performance in the field of multi-view clustering recently. However, the positive and negative sample construction mechanisms ignoring semantic consistency lead to false negative pairs, limiting the performance of existing algorithms from further improvement. To solve this problem, we propose a multi-view clustering framework named Deep Contrastive Multi-view Clustering under Semantic feature guidance (DCMCS) to alleviate the influence of false negative pairs. Specifically, view-specific features are firstly extracted from raw features and fused to obtain fusion view features according to view importance. To mitigate the interference of view-private information, specific view and fusion view semantic features are learned by cluster-level contrastive learning and concatenated to measure the semantic similarity of instances. By minimizing instance-level contrastive loss weighted by semantic similarity, DCMCS adaptively weakens contrastive leaning between false negative pairs. Experimental results on several public datasets demonstrate the proposed framework outperforms the state-of-the-art methods.
Paper Structure (27 sections, 14 equations, 4 figures, 8 tables, 1 algorithm)

This paper contains 27 sections, 14 equations, 4 figures, 8 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Multi-view clustering.
  4. Contrastive learning.
  5. Method
  6. Proposed Statement
  7. View-specific feature fusion
  8. Cluster-level contrastive learning
  9. Semantics-guided instance-level contrastive learning
  10. Instance pair weights
  11. Instance-level contrastive learning under semantic guidance
  12. Optimization
  13. Experiment
  14. Experimental Setup
  15. Datasets.
  16. ...and 12 more sections

Figures (4)

  • Figure 1: The framework of DCMCS. The view-specific features $\mathrm{Z}^{v}$, the fusion view features $\hat{\mathrm{Z}}$, the instance-level features $\mathrm{H}^{v}$,$\hat{\mathrm{H}}$ and the semantic features $\mathrm{C}^{v}$,$\hat{\mathrm{C}}$ are learned from the raw features $\mathrm{X}^{v}$. The reconstruction objective $L_{\mathrm{con}}$ is individually conducted on $\mathrm{Z}^{v}$. In semantics-guided contrastive learning(SGCL) and cluster-level contrastive learning(CLCL) modules, two contrastive losses (i.e., $L_i$ and $L_c$) are conducted on the instance-level features and cluster-level features, respectively. Moreover, $\mathrm{R^C}$ represents the weight matrix generated from semantic features to establish the relationship between negative pairs in $L_i$.
  • Figure 2: Attenion module. $\mathrm{W}_{\mathrm{Q}_{1}}$, $\mathrm{W}_{\mathrm{Q}_{2}}$, and $\mathrm{W_O}$ are utilized to achieve feature space transformation, and the weight matrix $\mathrm{R^C}$ is used to obtain instance pair weights.
  • Figure 3: Visualization of GCFAgg (a-c) and DCMCS (d-f) on Hdigit's two views and fusion view
  • Figure 4: (a) Convergence analysis. (b) The similarities of feature pairs and cluster pairs in contrative learning(CL). (c) and (d) Parameters sensitivity analysis.