Multi-View Discrete Clustering: A Concise Model.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Existing graph-based multi-view clustering often uses a two-stage approach involving eigen-decomposition and post-processing.
• This two-stage process can lead to deviations from directly solving the primal clustering problem.
• Effective integration of information from multiple views is crucial for enhancing multi-view clustering performance.

Purpose of the Study:

• To propose a concise model, Multi-view Discrete Clustering (MDC), that directly addresses the primal problem of multi-view graph clustering.
• To enhance multi-view clustering by automatically weighing view-specific similarity matrices and directly obtaining a discrete indicator matrix.
• To develop a hyper-parameter free model for simplified and effective multi-view graph clustering.

Main Methods:

• Developed the Multi-view Discrete Clustering (MDC) model to directly solve the primal multi-view graph clustering problem.
• Implemented automatic weighting of view-specific similarity matrices.
• Obtained the discrete indicator matrix directly from the aggregated similarity matrix without post-processing.
• Designed an efficient optimization algorithm to solve the objective problem.

Main Results:

• The proposed MDC model directly obtains discrete cluster indicators without post-processing.
• MDC effectively integrates information from multiple views through automatic similarity matrix weighting.
• The model operates without additive components or tunable hyper-parameters.
• Extensive experiments demonstrated the superiority of MDC on synthetic and real benchmark datasets.

Conclusions:

• The Multi-view Discrete Clustering (MDC) model offers a superior approach to graph-based multi-view clustering.
• Directly solving the primal problem and avoiding post-processing leads to improved clustering accuracy.
• The model's hyper-parameter free nature and efficient optimization enhance its practical applicability.