Abstract:

Kunqu， hailed as the ancestor of all Chinese dramas and a gem of traditional Chinese opera， holds a significant place in the protection list of the world’s cultural heritage. Kunqu related data have characteristics of large amount and complex relationships. Therefore， by analyzing the temporal and spatial correlation in the development process of Kunqu， a Spatio-Temporal Graph Neural Network （STGNN） model was proposed to utilize historical performance information of Kunqu to predict Kunqu performance popularity. Besides， aiming at the locality of performance venues， a density clustering method based on geographic location and performance popularity was designed to further highlight the role of Kunqu’s performance popularity information. Firstly， a clustering analysis of the performance venues was performed， considering not only the correlation between geographic locations， but also the Kunqu performance popularity information. Then， for the clustered performance information， the spatial correlation was analyzed using Graph Convolutional Network （GCN）， and temporal correlation was analyzed using Long Short-Term Memory （LSTM） network. Experimental results indicate that STGNN surpasses both GCN and LSTM methods in predicting the popularity of Kunqu performances， and demonstrates a faster convergence during the training process.

Key words: Graph Convolution Network (GCN), intangible cultural heritage, drama heritage, clustering algorithm, Long Short-Term Memory (LSTM) network

摘要：

昆曲被誉为中国的百戏之祖，作为中国传统戏曲的瑰宝之一，在世界文化遗产的保护名录中占据重要地位。昆曲相关数据具有数量庞大、关系复杂等特点。通过分析昆曲发展过程中空间与时间的相关性，提出一种时空融合的图神经网络（STGNN）模型，以利用昆曲历史演出信息预测昆曲演出的热度。此外，针对演出地点具有局部性的特点，为进一步突出昆曲演出热度信息的作用，设计一种基于地理位置与演出热度的密度聚类方法。首先对演出地点进行聚类分析，不仅考虑地理位置的相关性，还考虑昆曲演出的热度信息；其次，对于聚类后的演出信息，利用图卷积网络（GCN）进行空间相关性分析，并利用长短时记忆（LSTM）网络进行时间相关性分析。实验结果表明，STGNN对昆曲演出热度的预测指标均优于GCN和LSTM方法，且训练过程能较快收敛。

关键词: 图卷积网络, 非物质文化遗产, 戏剧遗产, 聚类算法, 长短时记忆网络