Online learning has attracted a large number of participants and is increasingly becoming very popular. However, the completion rates for online learning are notoriously low. Further, unlike traditional education systems, teachers, if any, are unable to comprehensively evaluate the learning gain of each student through the online learning platform. Hence, we need to have an effective framework for evaluating students' performance in online education systems and to predict their expected outcomes and associated early failures. To this end, we introduce Deep Online Performance Evaluation (DOPE), which first models the student course relations in an online system as a knowledge graph, then utilizes an advanced graph neural network to extract course and student embeddings, harnesses a recurrent neural network to encode the system's temporal student behavioral data, and ultimately predicts a student's performance in a given course. Comprehensive experiments on six online courses verify the effectiveness of DOPE across multiple settings against representative baseline methods. Furthermore, we perform ablation feature analysis on the student behavioral features to better understand the inner workings of DOPE. The code and data are available from https://github.com/hamidkarimi/dope. [For the full proceedings, see ED607784.]