Accurate evaluation of off-ball player contributions in soccer remains a critical challenge, as players spend the majority of match time without possession. While existing performance models focus predominantly on on-ball actions, they often overlook the subtle, context-dependent movements that influence team success. This study presents a machine learning framework for evaluating off-ball player contributions during attacking sequences, grounded in expert annotations. We construct a graph-based dataset from tracking and event data of J. League matches, and train a model combining Graph Convolutional Networks (GCN) and Long Short-Term Memory (LSTM) networks to capture both spatial and temporal aspects of play. To improve model robustness under limited and imbalanced data, we additionally applied data augmentation techniques such as mirroring, time shifting, and controlled noise addition. Evaluation experiments show that the model not only aligns well with expert judgments but also effectively identifies high-quality plays when trained under strategically filtered conditions. These findings suggest that our method offers interpretable, context-aware evaluations of off-ball player behavior, contributing to more comprehensive soccer analytics.