Welcome to the repository for our recent study on the differences in data organization between contrastive and supervised learning methods. Our research sheds light on the unique characteristics of clusters formed by these two types of learning methods, specifically focusing on the concept of locally dense clusters.
In our work, we introduce a novel metric known as Relative Local Density (RLD) to quantitatively measure local density within clusters. We provide illustrative examples to differentiate between locally dense clusters and globally dense ones.
Our comparative analysis between contrastive and supervised learning reveals that while the former fosters locally dense clusters, the latter tends to create clusters with both local and global density. We further explore the utilization of a Graph Convolutional Network (GCN) classifier as an alternative to linear classifiers for better handling locally dense clusters.
Our findings are further validated using t-SNE visualizations, which provide a visual representation of the distinctions between the features generated by the two learning methods.
The code here now is the raw code, thus could be very difficult to understand and use. However, we are currently in the process of cleaning and organizing the code to ensure that it is user-friendly and easily understandable. This process could take time so if you want to use it now, please contact us and we are more than willing to give you full instructions as best we can.
For any inquiries related to our research or this repository, please feel free to reach out to us.
Thank you for your interest in our work.