07/17/2026
By Qu Liu

The Miner School of Computer and Information Sciences is proud to announce a Dissertation Proposal by Qu Liu titled "Efficient and Adaptive Online Learning Methods over Graph Streams."

Defense Date: Tuesday, July 28, 2026
Time: 11 a.m. - noon EDT
Location: Please email advisor or student for location.

Committee Members:

  • Tingjian Ge (Advisor), Professor, Graduate Coordinator for MS Programs, Miner School of Computer & Information Sciences, CHORDS
  • Cindy Chen, Associate Professor, Department Chair, Miner School of Computer & Information Sciences, CHORDS
  • Yinghui Wu, Theodore L. and Dana J. Schroeder Associate Professor, Department of Computer and Data Sciences, Case Western Reserve University

Abstract:

Graphs/networks are a general model to represent data and information, especially when they are heterogeneous with various types, as data can be represented by entities (nodes) and relations (edges). Graph streams (aka dynamic graphs) characterize the fact that such data constantly change over time in real-world applications. Machine learning and analytics under the rich topological and temporal dimensions have been a challenging problem that has been a major focus of research efforts. In this thesis, we tackle this problem from the following aspects.

Firstly, we made an effort towards the co-existence of connectionist networks and symbolic rules in the same system, as envisioned by Marvin Minsky 35 years ago. We devise a system called RL2 and show that it is feasible to simultaneously and efficiently perform representation learning and rule learning out of the same online training process for graph streams. We show that it is highly efficient and responsive, and produces good results for both representation learning and rule learning in terms of prediction accuracy and returning top-quality rules for interpretation and building dynamic Bayesian networks.

Secondly, we study the problem of continuously predicting a number of user-subscribed continuous analytics targets (CATs) in dynamic networks. We design an architecture that can include any dynamic graph neural network model as the back end applied over the network data, and per-CAT front end models returning results with their confidence to users. We also devise a data filtering algorithm that feeds a provably optimal subset of data in the embedding space from back end model to front end models. To ensure fairness in terms of query result accuracy for different CATs and users, we propose a fairness metric and a fairness-aware training scheduling algorithm, along with accuracy guarantees on fairness estimation.

Thirdly, we observe the need for continuous, online training of dynamic graph neural network (DGNN) models while at the same time using them to answer continuous predictive queries as data updates stream in, but this implies significant costs in the training time and memory consumption. Along with the DGNN model learning, we simultaneously learn a weight/priority distribution over the nodes via a randomized online algorithm. In turn, the DGNN is continuously trained/learned by sampling nodes from the learned distribution and performing the chosen nodes' partitions of training work. We also devise a novel graph Kernel Density Estimation technique to smooth the distribution and improve the learning quality.

Fourthly, we study the continuous monitoring of multiple correlated predictive events in dynamic graphs. We formulate this problem as continuous multi-label classification and propose SCode, a deep metric-learning framework based on spherical codes. SCode jointly trains a dynamic graph neural network and a code generator so that each possible vector of predictive-event outcomes is represented by a codeword on a unit hypersphere. Predictions for multiple events can then be obtained together by efficiently matching the current graph-state embedding to its nearest codeword. We establish error-correcting and capacity properties of the learned spherical codes, providing robustness to noise and distribution drift, and analyze the generalizability of incremental online learning. Experimental results demonstrate that SCode achieves significantly higher predictive accuracy than dynamic graph neural network baselines while being approximately twice as fast.

Fifthly, we investigate the estimation of potential relationship values in dynamic interaction networks, where many important relationships are unobserved and the available interactions are subject to non-random selection bias. We propose a propensity-regularized representation-learning framework that incorporates the probability of observing an interaction as an endogenous structural regularizer, allowing the model to disentangle latent relationship values from the mechanisms determining which interactions are observed. The framework supports both individual and aggregate potential relationship-value queries through complementary neural prediction and embedding-based matching methods. We also introduce smoothness regularization and derive theoretical bounds that relate estimation error to distances in the learned representation space. In addition, we develop a temporal factor-attribution method to identify historical nodes that influence estimated relationship values. Experiments on four real-world dynamic-network datasets show that the framework improves estimation in both cold-start and warm-run settings and provides robust and interpretable value-based inference under selection bias.

Finally, we study robust and interpretable prediction for heterogeneous and continually changing tasks over temporal graphs. Existing temporal graph neural networks generally combine graph encoding, evidence aggregation, and prediction within a single opaque pipeline, making it difficult to adapt to different tasks or provide faithful explanations. We propose Pilgrim, a task-conditioned interface layer that decouples graph representation learning from task-specific decision-making. Pilgrim uses a differentiable, adaptive top-k mechanism to retrieve a compact set of task-relevant evidence nodes, and predictions are constrained to depend on this selected evidence, providing intrinsic attribution by construction. To improve robustness to class imbalance, noise, and continual task shifts, we combine annealed metric learning with representation-level mixup to impose a stable geometric structure on the evidence space. We further provide theoretical analyses showing that the evidence representation preserves task-relevant information while suppressing nuisance variation. Across diverse temporal-graph tasks and multiple backbone models, Pilgrim consistently improves predictive performance and robustness while producing more faithful attribution than post-hoc explanation methods.