Go to ICLR 2025 Conference homepageTowards Accurate Validation in Deep Clustering through Unified Embedding Learning
Deep clustering integrates deep neural networks into the clustering process, simultaneously learning embedding spaces and cluster assignments. However, significant challenges remain in evaluating and comparing the performance of different deep clustering algorithms—or even different training runs of the same algorithm. First, evaluating the clustering results from different models in the same high-dimensional input space is impractical due to the curse of dimensionality. Second, comparing the clustering results of different models in their respective learned embedding spaces introduces discrepancies, as existing validation measures are designed for comparisons within the same feature space. To address these issues, we propose a novel evaluation framework that learns a unified embedding space. This approach aligns different embedding spaces into a common space, enabling accurate comparison of clustering results across different models and training runs. Extensive experiments demonstrate the effectiveness of our framework, showing improved consistency and reliability in evaluating deep clustering performance.