零工式排程問題為製造與生產管理中典型之組合最佳化問題。隨著問題規模與決 策複雜度提升，傳統最佳化方法通常需要較高之搜尋成本與參數調整。然而，現有學習式排程研究多著重於單一目標，對於同時考量生產效率與交期風險之多目標排程問題仍較少探討。本研究針對雙目標零工式排程問題，同時最小化最大完工時間與總逾期階段數量。研究中利用NSGA-II 產生參考解作為模型訓練與比較基準。模型以分離圖表示排程結構，並結合圖神經網路與指針式解碼器，逐步建構排程解。 本研究比較僅監督式學習、僅自標籤優化機制與混合策略三種訓練方式。實驗結 果顯示，混合策略在5×5設定下表現較穩定，並在5×5訓練、10×10驗證時展現較佳跨規模泛化能力。相對地，僅自標籤優化機制在10×10同規模設定下具有較強之持續改善能力。特徵消融分析顯示，工作相關資訊對模型表現影響最為明顯，而交期與逾期壓力資訊有助於改善逾期目標。整體而言，學習式模型雖未在所有設定下皆優於NSGA-II，但其在跨規模泛化與多目標排程決策上展現出應用潛力。;The job shop scheduling problem is a typical combinatorial optimization problem in manufacturing and production management. As problem size and decision complexity increase， traditional optimization methods often require high search costs and parameter tuning. However， most learning-based scheduling studies focus on single-objective settings， while multi-objective scheduling with due-date risk remains less explored. This study investigates a bi-objective job shop scheduling problem that minimizes makespan and the total number of tardy stage-outs. NSGA-II is used to generate reference solutions for model training and comparison. The proposed model represents the problem as a disjunctive graph and combines a graph neural network with a pointer-network-based decoder to construct schedules step by step. Three training strategies are compared: supervised learning， self-labeling improve ment， and a hybrid strategy. The results show that the hybrid strategy performs more stably in the 5×5 setting and achieves better cross-scale generalization when trained on 5×5 instances and validated on 10×10 instances. In contrast， self-labeling improvement shows stronger continuous improvement in the 10×10 same-scale setting. Feature ablation analysis indicates that job-related information has the greatest impact on model performance， while due-date and tardiness-pressure features help improve the tardiness objective. Overall， although the learning-based model does not outper form NSGA-II in all settings， it shows potential for cross-scale generalization and multi objective scheduling decisions.
