01/28/2026
By Nidhi Piyush Vakil

The Kennedy College of Science, Richard A. Miner School of Computer & Information Sciences, invites you to attend a doctoral dissertation proposal defense by Nidhi Vakil, titled: "Foundations for Controllable and Curriculum-Based Graph Intelligence."

Date: Feb. 13, 2026
Time: 3 - 4 p.m. EST
Location: Dandeneau 309 or via Zoom 

Committee members

  • Hadi Amiri (Advisor), Assistant Professor, Miner School of Computer and Information Sciences, UMass Lowell
  • Hong Yu, Professor, Miner School of Computer and Information Sciences, UMass Lowell
  • Tingjian Ge, Professor, Miner School of Computer and Information Sciences, UMass Lowell
  • Yunyao Li, Director of Machine Learning, Adobe

Abstract:

Graph learning has become a foundational paradigm for modeling relational and structured data. Beyond node and edge attributes, graphs embody rich structural attributes–such as density, centrality, connectivity, and motif distributions–that naturally quantify structural complexity and can serve as steering signals for both learning and generation. However, existing methods either treat all graphs as having a uniform difficulty level or rely on a single, static notion of difficulty (e.g. instantaneous training loss) when designing training curricula. Such approaches do not fully capture the multifaceted nature of complexity in graphs. Moreover, current graph generation methods typically lack explicit, fine-grained control over structural attributes or focus on a limited number of graph properties, which limits their applicability in settings that require precise structural guarantees. This thesis proposes to use graph topological attributes to (1) design principled curriculum learning schemes to improve the efficacy of message-passing in training graph neural networks (GNNs), and (2) enable controllable graph generation that can be steered toward desired structural configurations. This proposal is organized into two parts:  Part I focuses on curriculum learning for graph data. First, we present a curriculum that integrates structurally-informed node and text embeddings with instantaneous loss trajectories to dynamically quantify the evolving difficulty of individual samples and adapt their difficulty with respect to both the model and data [Vakil and Amiri, 2022b]. Second, we develop a multi-view competency-based curriculum learning (MCCL) framework that overcomes the limitation of single view difficulty criteria by operationalizing difficulty through multiple topology-derived attributes [Vakil and Amiri, 2023b]. The curriculum is updated based on the GNN’s competency (learning progress) to determine the next most informative view for effective training, resulting in better performance of link prediction and node classification tasks over non-curriculum and single view curricula. Finally, we introduce text-graph curriculum learning (TGCL) that combines a multi-view complexity formalism with space repetition-based scheduling strategy to guide training [Vakil and Amiri, 2023a]. The approach enables transferring learned curricula across datasets and GNN architectures, while reducing data usage by avoiding exposure to all samples at every iteration, leading to both efficiency and performance gains.

Part II  introduces controlled graph generation using fine-grained graph attributes as conditioning variables. We propose the controlled graph translator (CGT) model that performs controllable, multi-objective translation (morphing) from a given source graph to a target graph that satisfies multiple desired structural attributes at granular level [Vakil and Amiri, 2024]. This framework supports tasks such as taxonomy refinement, discovery of missing edges, and graph augmentation in low resource regimes. The above contributions establish foundations for controllable and curriculum-based graph intelligence that tightly integrates topological structure with learning and generation.

The proposal concludes by outlining the renaming thesis work, which will extend the use of structural attributes from guiding the learning process and generation to also removing and mitigating undesirable structures in graphs. First, we will study how fine-grained graph attributes can be used to steer graph generation toward desired structural attributes. These methods are expected to be particularly important for data augmentation in low-resource and imbalanced regimes, as well as for molecular graph generation where domain-specific structural and functional properties must be satisfied. Second, we will design principled curricula for graph unlearning, where specific undesirable patterns, biases, or subgraphs are removed from trained GNN models while preserving model’s knowledge on the remaining data. In addition, we will investigate unlearning through mode connectivity for graphs, where we formulate unlearning as movement along a low-loss manifold between an initial GNN and an updated unlearned model. This will allow comparing optimized curricula for learning versus unlearning to investigate differences in their dynamics and to provide principled strategies for removing undesirable information from graph data.