Aiming at the top-bottom aggregation-type multigrid preconditioner from top to bottom based on graph partitioning, we investigate the multigrid construction method using METIS software package. Given that software package METIS can only process integer weights but not float-type weights, we propose an effective scheme to convert float-type edge weights to integer edge weights. Then it is used to select edge weights in the METIS graph partitioning software, and we propose an improved algorithm for aggregation-type multigrid preconditioner from top to bottom. Numerical experiments on the solution to the sparse linear system derived from two-dimensional and three-dimensional model partial differential equations show that the improved method with edge weights can greatly improve the iterative efficiency of the multigrid preconditioned conjugate-gradient method, and the improvement for anisotropic problems are especially significant.