CDIMC-net: Cognitive Deep Incomplete Multi-view Clustering Network

1 Introduction

• Earlier IMC work includes PMVC-family, weighted-matrix-factorization, graph-learning, and multiple-kernel approaches, but some methods do not cover arbitrary missing-view cases [Wen et al., 2020].The cited methods include KCCA, PMVC, graph-regularized PMVC, and GAN-based approaches.Source: §1 · pp. 1–2
• Prior methods often use shallow models that miss high-level features, and treating marginal samples equally with other samples can make training sensitive to outliers.CDIMC-net is proposed to address both issues.Source: §1 · p. 2
• CDIMC-net combines view-specific deep encoders, graph embedding, and a self-paced k-means layer in a network designed for arbitrary incomplete-view cases.The authors describe graph embedding as promoting network training.Source: §1 · p. 2
• The self-paced strategy adaptively reduces the negative influence of marginal samples and is presented as making CDIMC-net more robust to outliers.The paper identifies this as its first introduction of human-cognition-based learning to IMC.Source: §1 · p. 2

2 Kmeans Clustering

• For n samples with m features, k-means seeks k cluster centers and a cluster-indicator matrix.The data are represented as X = [x1, . . . , xn] ∈ R^m×n and the centers as U = [u1, . . . , uk] ∈ R^m×k.Source: §2 · p. 2
• The cluster-indicator matrix S has dimensions k×n and records assignments of samples to clusters.The section presents S as part of the k-means problem formulation.Source: §2 · p. 2
• An entry S_j,i = 1 means sample x_i is assigned to cluster j; the vector 1 has all elements equal to 1.Source: §2 · p. 2

3.1 Problem Statement

• Incomplete multi-view clustering groups samples into k clusters despite unavailable views, whose presence is recorded per view and instance.The method represents missing instances as NaN and records view availability in a diagonal matrix.Source: §3.1 · p. 2

3.2 CDIMC-net

• CDIMC-net first pre-trains an autoencoder to initialize parameters, then fine-tunes cluster-friendly representations and produces cluster indicators.These are the two phases of the framework.Source: §3.2 · p. 2

Pre-training Network

• View-specific encoders, weighted fusion, and graph embedding capture view features, form a common representation, reduce missing-view influence, and preserve local neighbor structure.Pre-training combines autoencoder and graph-embedding losses.Source: Pre-training Network · pp. 2–3

Fine-tuning and Clustering

• Self-paced k-means fine-tuning selects samples with clustering losses no greater than λ, progressing from confident samples toward more samples while keeping cluster centers fixed.The method combines k-means and graph-embedding losses and produces clustering indicators.Source: Fine-tuning and Clustering · pp. 3–4

3.3 Optimization

• Fine-tuning alternates encoder, cluster-indicator, sample-weight, and age-parameter updates, using SGD and back-propagation for encoder optimization, and converges to a local optimum.The optimization description assigns SGD and back-propagation to the encoder update.Source: §3.3 · p. 4

1 Initialization: Feed

• Initialization feeds the pre-trained network to obtain a consensus representation, then applies k-means to initialize cluster centers and indicators and sets all sample weights to 1.These steps are stated at the start of the fine-tuning algorithm.Source: §1 · p. 4

3.4 Implementation for IMC

• Implementation fills missing instances with view averages for initial clustering, reorders same-cluster samples together, and constructs nearest-neighbor graphs from the rearranged data.The resulting rearrangement is used to organize graph information for network training.Source: §3.4 · pp. 4–5
• The pre-training network uses the rearranged data, graphs, and indicator matrices, setting missing-view features to zero; fine-tuning then updates network parameters, indicators, and weights.The algorithm lists updates to network parameters, cluster indicators, and the weight vector.Source: §1 · p. 4, §3.4 · p. 5

4.1 Experimental Settings

• Experiments use Handwritten, BDGP, and MNIST, with missing-view or paired-view rates p ∈ {10, 30, 50, 70}; clustering is evaluated with ACC and NMI.Comparisons include multiple incomplete-clustering methods, best single view, and Concat.Source: §4.1 · p. 5

4.2 Experimental Results and Analysis

• CDIMC-net significantly outperforms other methods on all three databases; on Handwritten at 50% missing views, it reaches about 91% ACC and 84% NMI, roughly 10% and 16% above the second-best method.It also performs better than PMVC CGAN on MNIST.Source: §4.2 · p. 5, Table 2 · p. 6

4.3 Parameter Analysis

• On Handwritten and BDGP at a 10% missing-view rate, relatively better ACC occurs with smaller α and learning rates; large learning rates hurt local-optimal clustering, while large α lets graph embedding dominate training.The authors suggest choosing both parameters from [1e-5,1e-3].Source: §4.3 · p. 5, Figure 2 · p. 6

4.4 Component Analysis

• On Handwritten and BDGP, CDIMC-net outperforms variants without graph embedding, self-paced learning, or pre-training, supporting the effectiveness of all three components.Source: §4.4 · p. 5, Figure 3 · p. 6

4.5 Convergence Analysis

• On Handwritten and BDGP at a 10% missing-view rate, fine-tuning loss trends downward overall and falls quickly in the first few iterations, which the authors take as evidence of convergence.Source: Figure 4 · p. 6, §4.5 · p. 6

• CDIMC-net handles all kinds of incomplete data, uses self-paced k-means to improve robustness to outliers, and incorporates graph embedding to preserve local structure.Its superior performance is validated on several incomplete cases against state-of-the-art IMC methods.Source: §5 · p. 6