Abstract: In order to deal with non-linear data structures, a robust representative data sampling strategy based on multi-objective sketch description is proposed. Each data point was coded by solving quadratic programming in manifold bottom structure, and a parallel algorithm was introduced. A multi-objective manifold sketch description method was proposed, which ensured the representativeness, conciseness, and robustness of sampling selection through sparsity measurement of encoding. A highly scalable random algorithm was further introduced to effectively improve scalability and acceleration. The effectiveness of the proposed method was verified through dataset experiments.