The ever-expanding volume of data presents considerable challenges in storing and processing semi-supervised models, hindering their practical implementation. Researchers have explored reducing network versions as a potential solution. Real-world networks often comprise diverse vertex and edge types, leading to the adoption of k-partite network representation. However, existing methods have mainly focused on reducing uni-partite networks with a single vertex type and edges. This study introduces a novel coarsening method designed explicitly for k-partite networks, aiming to preserve classification performance while addressing storage and processing issues. We conducted empirical analyses on synthetically generated networks to evaluate their effectiveness. The results demonstrate the potential of coarsening techniques in overcoming storage and processing challenges posed by large networks. The proposed coarsening algorithm significantly improved storage efficiency and classification runtime, even with moderate reductions in the number of vertices. This led to over one-third savings in storage space and a twofold increase in classification speed. Moreover, the classification performance metrics exhibited low variation on average, indicating the algorithm’s robustness and reliability in various scenarios.