With the next-generation Web aiming to further facilitate data/information sharing and aggregation, providing data privacy protection support in an open networked environments becomes increasingly important. Learning-fromabstraction is a recently proposed distributed data mining approach which first abstracts data at local sources using the agglomerative hierarchical clustering (AGH) algorithm and then aggregates the abstractions (instead of the data) for global analysis. In this paper, we explain the limitation of the use of AGH for local manifold preserving data abstraction and propose the use of the graph-based clustering approach (e.g., the minimum cut) for local data abstraction. The effectiveness of the proposed abstraction approach was evaluated using benchmarking datasets with promising results. The global analysis results obtained based on the minimum cut abstraction was found to outperform those based on the AGH abstraction, especially when the underlying manifold was complex.