Multi View Clustering Eval

multi-view-clustering-eval

Automatically Identify and Rectify: Robust Deep Contrastive Multi-view Clustering in Noisy Scenarios — Xihong Yang et al. (2025) (arXiv:2505.21387, 2025)

What this evaluates

Evaluates the robustness of deep multi-view clustering models when input data is corrupted by randomly injected noise at varying proportions. It measures how effectively the model can identify noisy samples, rectify them, and produce accurate cluster assignments across multiple feature views.

Datasets

• BBCSport — total 544; splits: test (-1)
• WebKB — total 1051; splits: test (-1)
• Reuters — total 1200; splits: test (-1)
• UCI-digit — total 2000; splits: test (-1)
• Caltech101 — total 9144; splits: test (-1)
• STL10 — total 13000; splits: test (-1)

Metrics

• ACC (primary) — range: percent
  • Accuracy (ACC) measures clustering performance by finding the best one-to-one mapping between predicted clusters and ground-truth labels using the Hungarian algorithm, then computing the ratio of correctly mapped samples.
• NMI — range: percent
  • Normalized Mutual Information (NMI) quantifies the mutual dependence between predicted cluster assignments and true labels, normalized by the entropy of both distributions.
• PUR — range: percent
  • Purity (PUR) calculates the proportion of the most frequent class in each predicted cluster, averaged over all clusters, representing the fraction of correctly assigned samples.

Input / output format

Input: Multi-view feature vectors for each sample, with a dataset-specific number of views. Noisy versions are generated by randomly corrupting a specified percentage of the input data.

Output: Cluster assignment labels for each sample across all views.

Scoring recipe

def compute_clustering_metrics(pred, gold):
    from scipy.optimize import linear_sum_assignment
    from sklearn.metrics import normalized_mutual_info_score
    n_clusters = len(set(pred))
    contingency = np.zeros((n_clusters, len(set(gold))))
    for p, g in zip(pred, gold):
        contingency[p][g] += 1
    row, col = linear_sum_assignment(-contingency)
    acc = contingency[row, col].sum() / len(pred)
    pur = acc
    nmi = normalized_mutual_info_score(gold, pred)
    return acc * 100, nmi * 100, pur * 100

Common pitfalls

• Noise is injected randomly into multi-view inputs at specific rates (10%, 30%, 50%, 70%, 90%), meaning results are not evaluated on a single fixed test set but across multiple corruption levels.
• All reported results are averaged over 10 independent runs to account for randomness in initialization and noise injection, not single-run scores.
• Baseline comparisons require reproducing original source codes and configurations as stated, rather than relying on public checkpoints or default hyperparameters.

Evidence (verbatim from paper)

Evaluation Metrics: To provide a thorough evaluation of the model’s clustering performance, we utilize three widely recognized metrics: Accuracy (ACC), Normalized Mutual Information (NMI), and Purity (PUR). To ensure a fair comparison, all methods are assessed across 10 independent runs, and the average results are reported.

Citation

@misc{yang2025airmvc,
  title={Automatically Identify and Rectify: Robust Deep Contrastive Multi-view Clustering in Noisy Scenarios},
  author={Xihong Yang et al. (2025)},
  year={2025},
  note={arXiv:2505.21387}
}

• arXiv: 2505.21387