<p>The action quality assessment (AQA) is used to score the actions in the motion video. According to the input method, the AQA method can be divided into RGB video-based and skeleton data-based. Existing algorithms based on skeleton sequence are difficult to make full use of skeleton information, and may ignore the importance of a specific part of the human body for scoring. In view of the above problems, this paper proposes a Body Part-Based Dual-Task Spatio-Temporal Graph Convolutional Network (BPDT-STGCN). Specifically, the human skeleton is partitioned into three anatomical parts, the head, hand, and leg, which are individually fed into Spatio-Temporal Graph Convolutional Networks (ST-GCN) for feature extraction. Considering the varying importance of different body parts, a Multi-Scale Dynamic Feature Aggregation Network (MS-DFAN) is designed to dynamically assign weights to each part, thereby adapting to diverse action types. Furthermore, to enhance the model’s capability to distinguish between samples of varying quality and improve the accuracy of overall ranking assessment, a novel optimization strategy is introduced. This strategy utilizes K-means clustering to categorize samples into three quality levels: good, medium, and poor. Consequently, the model is trained to perform an auxiliary fuzzy prediction of these quality levels. Based on the fuzzy prediction results, the Mean Squared Error (MSE) loss for the good and poor samples is dynamically weighted. The effectiveness of the proposed method is extensively validated on the public Rhythmic Gymnastics dataset and the MIT-Skate dataset, where comprehensive experimental results demonstrate the superiority and validity of our approach. Our code is available at:<a href="https://github.com/kkioll/BPDT-STGCN">https://github.com/kkioll/BPDT-STGCN</a></p>