The analysis and discrimination of time series data has important practical significance. Currently, transforming the time series data into networks through visibility graph (VG) methods is an effective approach for classifying the series data through GNNs. However, there are two main obstacles to the VG method: (1) the tension between efficiency and complexity during weighted graph construction; (2) difficulty in assigning the different importance of nodes. To tackle these difficulties, we propose an improved weighted visibility graph algorithm (WLVG) in this paper. The proposed algorithm can first intelligently assign weights to the network according to the Euclidean distance among nodes, and then resample the network by the weight coefficients resulting in the removal of the unimportant edges. Finally, in order to effectively aggregate the information among neighbors, the graph isomorphism network (GIN) is utilized for identifying the objects. Experimental results show WLVG outperforms other baseline methods on several practical datasets and demonstrate its effectiveness. •We propose a new weighting method, which can reduce the interference of remote noisy nodes.•We propose an improved visibility graph algorithm to actively drop the trivial edges in the network.•We combine the proposed weighting method with VG/HVG, which is then integrated with GIN.