The ability to learn from user interactions provides an effective means to understand user intent through their behaviors that is instrumental to improve user engagement, incorporate user feedback, and gauge user satisfaction. By leveraging important cues embedded in such interactions, a learning system can collect key evidence to uncover users’ cognitive, affective, and behavioral factors, all of which are critical to maintain and/or increase its user base. However, in practical settings, interactive systems are still challenged by sparse user interactions that are also dynamic, noisy, and highly heterogeneous. As a result, both traditional statistical learning methods and contemporary deep neural networks (DNNs) may not be properly trained to learn meaningful patterns from sparse, noisy, and constantly changing learning signals. In recent years, the learning to learn (L2L) (or meta-learning) paradigm has been increasingly leveraged in diverse application domains, such as computer vision, gaming, and healthcare to improve the learning capacity from sparse data. L2L tries to mimic the way how humans learn from many tasks that allows them to generalize often from extremely few examples. Inspired by such an attractive learning paradigm, this dissertation aims to contribute a novel L2L framework that is able to effectively learn and generalize well from sparse user interactions to collectively address the challenges of learning from sparse user interactions. The L2L framework is comprised of four interconnected components. The first component focuses on learning from sparse interactions that constantly change over time. For this, we introduce a Dynamic Meta Learning (DML) model that integrates a meta learning module with a sequential learning module, where data sparsity is tackled by the first module and the later module captures constantly changing user interaction behaviors. The second component focuses on dealing with noisy interactions to detect true user intent through uncertainty quantification. This component integrates evidential learning with meta learning, in which the former quantifies the uncertainty by leveraging evidence of each interaction and the later tackles sparse interactions. Furthermore, the concept of evidence is utilized to guide the predictive model to find more important and informative interactions in sparse data to enhance model training. The third component emphasizes dealing with dynamic and heterogeneous user behavior in sparse interactions. This component aims to ensure long-term user satisfaction by combining reinforcement learning, which handles constantly evolving user behaviors and evidential learning, which leverages evidence-based exploration to tackle data heterogeneity. The last component aims to advance the contemporary dynamic models which require to partition the time into arbitrary intervals to support model training and inference. We develop a novel Neural Stochastic Differential Equation (NSDE) model in the L2L setting that captures continuously changing user behavior and integrates with evidential theory to achieve evidence-guided learning. This method leverages the power of a NSDE solver to capture user’s continuously evolving preferences over time which results in richer user representation than previous discrete dynamic methods. Furthermore, we derive a mathematical relationship between the interaction time gap and model uncertainty to provide effective recommendations.

Library of Congress Subject Headings

Learning models (Stochastic processes); Deep learning (Machine learning); Neural networks (Computer science)

Publication Date


Document Type


Student Type


Degree Name

Computing and Information Sciences (Ph.D.)

Department, Program, or Center

Computing and Information Sciences Ph.D, Department of


Golisano College of Computing and Information Sciences


Qi Yu

Advisor/Committee Member

Xumin Liu

Advisor/Committee Member

Rui Li


RIT – Main Campus

Plan Codes