Abstract <p>Weakly supervised learning algorithms have become increasingly important for modeling complex systems where precise labels are scarce or expensive to obtain. There are specialized algorithms for solving the weakly supervised regression problem, such as the Weakly Supervised Regression algorithm [1], which is based on manifold regularization and cluster ensemble. In this article, we introduce novel improvements to original algorithm, that significantly increase the stability and quality of the algorithm and reduce its dependence on properly selected hyperparameters. This result is achieved through the use of fuzzy clustering and consistency weights when constructing a cluster ensemble.</p>