The execution time prediction for query tasks in graph database has become difficult and challenging due to the complexity of query plan and system. It is difficult for Database Administrators (DBA) or Database Management System (DBMS) to catch the accurate execution time during and before the execution of a query task. Before executing a query task, predicting its execution time can help the DBA or DBMS to efficiently management in the fields of load management, task scheduling, permission control, progress monitoring, system scale customization, etc. Therefore, accurately and efficiently predicting the execution time for query tasks is a key technology in these fields. In this paper, motivated by the combination of artificial intelligence technologies and graph database theories, we first propose a novel deep learning method to predict the execution time for query tasks in graph database. First, each query plan tree of tasks is encoded into an operation sequence. Second, top-20 features are selected from 68 candidate system features using random forest (RF), and the selected top-20 features are reduced to five principal components using principal component analysis (PCA). Finally, an accurate and efficient model based on the long short-term memory (LSTM) is designed and implemented to predict the execution time. The model can predict the execution time in advance before executing a query task in graph database. The experimental results from six kinds of benchmarks with the public data set Yelp show that the average accuracy of the proposed model can reach 81.34% with a high prediction efficiency rate, which proves the feasibility of the deep learning method. In particular, the proposed model can achieve the state-of-the-art prediction performance for query task execution time.

由于查询计划和系统的复杂性，图数据库中查询任务的执行时间预测已变得困难和挑战。数据库管理员（DBA）或数据库管理系统（DBMS）很难在执行查询任务期间和之前捕获准确的执行时间。在执行查询任务之前，预测其执行时间可以帮助DBA或DBMS在负载管理，任务调度，权限控制，进度监视，系统规模定制等领域进行有效管理。因此，准确而有效地预测执行时间查询任务是这些领域中的关键技术。本文受人工智能技术和图数据库理论相结合的启发，我们首先提出一种新颖的深度学习方法，以预测图数据库中查询任务的执行时间。首先，将每个任务的查询计划树编码为一个操作序列。其次，使用随机森林（RF）从68个候选系统特征中选择前20个特征，然后使用主成分分析（PCA）将选择的前20个特征缩减为5个主要成分。最后，设计并实现了基于长短期记忆（LSTM）的准确高效的模型，以预测执行时间。该模型可以在执行图数据库中的查询任务之前预先预测执行时间。根据公开数据集Yelp的六种基准测试结果，该模型的平均准确率可以达到81.34％，预测效率很高，证明了深度学习方法的可行性。特别地，提出的模型可以实现查询任务执行时间的最新预测性能。