A Hybrid Semi-Supervised Deep Clustering Framework with Pairwise Constraints and Graph-Regularized Representation Learning

This paper proposes a new semi-supervised deep clustering framework. The method improves both representation learning and clustering performance. It integrates several loss components and uses pairwise constraints. The approach is based on an autoencoder structure. A clustering phase is added, guided by a joint objective function. This function includes reconstruction loss, Kullback–Leibler divergence, semi-supervised constraint loss, label supervision, and a graph-based regularization term. The proposed method overcomes the limitations of previous models such as IDEC. It allows the use of user-defined pairwise constraints and ensures local label consistency among neighboring samples. A specific clustering layer and a hybrid loss design help the model transfer label information effectively and preserve structural consistency in the latent space. Experiments show that the method can learn compact embeddings and produce accurate cluster assignments even with limited labeled data. It is therefore a promising solution for semi-supervised clustering problems.