CLARE: Cluster-Number-Independent Latent Representation Learning with Graph Autoencoders for Deep Clustering
Abstract
Deep clustering uncovers latent structure in the data, but the number of clusters K is rarely known in advance. Because the true K is unknown, data analysts cluster the data under many candidate values of K and compare the results to find the most reasonable clusters. Many recent deep clustering methods improve accuracy using self-supervised methods, such as pseudo-labeling, but K is defined in the training objective. Evaluating each candidate K requires retraining the model from scratch, making a complete sweep across all K values computationally expensive. We propose CLARE (Cluster-number-independent LAtent REpresentation learning), a representation learning framework that does not use K during training. CLARE builds an adaptive kNN graph that reflects the clustering objective and keeps an edge only when it is consistent with a node's local neighborhood. A graph autoencoder then learns a single embedding from this graph, and we can group the data into any number of clusters without retraining. We further introduce a cluster-number-independent clustering loss that sharpens the cluster structure by extending the global clustering coefficient to two-hop connections. Across eight text and three image benchmarks, CLARE reaches state-of-the-art accuracy on long-text and on two of the four short-text datasets, remains competitive on medium-text, and sets a new state of the art on fine-grained image clustering.