[Re]Reproducibility Study: Cluster-guided Contrastive Graph Clustering Network

TMLR Paper2182 Authors

Abstract:

Contrastive learning has shown promise in deep graph clustering, but current methods have notable limitations. Positive sample effectiveness is heavily influenced by data augmentation, risking semantic drift and improper pair selection. Moreover, negative pairs lack
reliability, disregarding crucial clustering information. To address these issues, the authors proposed Cluster-guided Contrastive Graph Clustering Network (CCGCN). It employs a unique Siamese encoder architecture, creating two distinct graph views without complex
augmentations, enhancing positive sample discriminative quality. Negative samples are selected from cluster centers for improved reliability. An objective function encourages intra cluster cohesion and inter-cluster separation. Initially evaluated on six datasets, we expand
to 12 graph and 2 non-graph datasets in our study, aiming to validate and generalize the method’s effectiveness through reproducibility and additional experiments

Submission Length: Regular submission (no more than 12 pages of main content)

Changes Since Last Submission:

EDIT 1:

1. We incorporated a cross-validation study into the paper, revealing that the model exhibits low variance in evaluation metrics. Specifically, the evaluation metrics for the test set demonstrate minimal deviations compared to those of the training set. This consistency suggests that the model performs consistently well on unseen data, indicating low variance.
Moreover, our analysis unveiled an unusual behavior in smaller datasets. In such cases, we observed a peculiar trend wherein evaluation metrics displayed higher standard deviations from the mean values.

2. We have mentioned and discussed in detail the computational challenges, model stagnation which we face while applying the model to non graph datasets and our solutions to these problems.

3. Hyperparameter sensitivity analysis has been extended to additional dataset to confirm that the model shows negligible variation in performance with changes in models hyperparameter, in some cases showing 0 changes.

EDIT 2:

Minor corrections to the formatting of the document

Assigned Action Editor: ~Ran_He1

Submission Number: 2182