Canada Spends Logo

Explore our Budget 2026 analysis

Grant: Discovery Grants Program - Individual

McGill University

$50,000.00

Bayesian Learning for Graph Data

FY 2023-24

Summary
Although there has been great progress in learning methodologies for data on graphs and networks, numerous challenges remain. Many settings give rise to graph-structured data. In some cases, a graph is provided explicitly, with edges depicting relationships between data entities. In other cases, a graph is used to model underlying dependencies. Example application settings include communication network configuration, recommender system development, and drug and protein design. Learning systems have been deployed successfully in real-world settings, but often in controlled settings where failure does not have significant consequences. To encourage further integration, it is vital to develop models that can provide reliable decisions. A fundamental component of reliability is consistent performance when presented with a variety of tasks and unusual or changing data. The deployment environment often differs considerably from the training environment. The assumption that encountered data are drawn from the same distribution as the training data frequently fails to hold. We observe distribution drift and outliers. Models are usually trained using proxy metrics such as accuracy or log-likelihood, but often these do not correspond to the metrics of interest for deployed models, which are more task-centric and application-specific. Moreover, a deployed model can be used for multiple tasks with differing objectives, and tasks can evolve over time with different performance aspects becoming more important. The proposed research program targets the development of reliable learning models and algorithms for graph-structured data. I will design graph learning methods and models that can: (i) quantify their uncertainty to facilitate improved decision making; (ii) detect and adapt to changes in the data distribution; and (iii) facilitate active, continual and sequential learning. We will develop flexible models that can successfully address multiple tasks and scale to extremely large graphs without overly sacrificing expressivity. A key aim is to support continual learning, where a model can be trained sequentially over time with multiple tasks. The cornerstone of our scientific approach is the Bayesian framework. This involves specifying prior distributions for model parameters and updating them based on observed data to construct posterior distributions. The Bayesian paradigm is well-suited to the goal of characterizing uncertainty. Since it targets posterior distributions rather than just point estimates, there is a natural mechanism for deriving confidence intervals. The Bayesian approach also provides the advantages of principled mechanisms for handling missing data, detecting outliers, and incorporating prior information. The approach is also often more effective for settings such as active learning where there is a limited amount of data, and is naturally aligned with a continual learning setting with streaming data and sequential updates.
Program
Discovery Grants Program - Individual
Location
Quebec, Canada
Recipient Type
Academia
Type
Grant
Source
NSERC's Awards Data

As published

Application id
RGPIN-2023-03899
Name nom
Coates, Mark
Department département
Electrical and Computer Engineering
Organization id
61
Institution établissement
McGill University
Province en
Québec
Province fr
Québec
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2023
Competition year année de concours
2023
Award amount
50000
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1510
Committee name en
Electrical and Computer Engineering
Committee name fr
Génie électrique et informatique
Area of application code
802
Area of application group en
Information and communication services
Area of application group fr
Services d'information et de télécommunications
Area of application en
Information systems and technology
Area of application fr
Systèmes d'information et technologie de l'information
Research subject code
3000
Research subject group en
Statistics and probability
Research subject group fr
Statistics and probability
Research subject en
Statistique et probabilité
Research subject fr
Statistique et probabilité
Application title
Bayesian Learning for Graph Data
Keyword
Bayesian inference; Continual learning; Graph neural networks; Graphs; Sequential inference; Time series analysis; Uncertainty
Application summary
Although there has been great progress in learning methodologies for data on graphs and networks, numerous challenges remain. Many settings give rise to graph-structured data. In some cases, a graph is provided explicitly, with edges depicting relationships between data entities. In other cases, a graph is used to model underlying dependencies. Example application settings include communication network configuration, recommender system development, and drug and protein design. Learning systems have been deployed successfully in real-world settings, but often in controlled settings where failure does not have significant consequences. To encourage further integration, it is vital to develop models that can provide reliable decisions. A fundamental component of reliability is consistent performance when presented with a variety of tasks and unusual or changing data. The deployment environment often differs considerably from the training environment. The assumption that encountered data are drawn from the same distribution as the training data frequently fails to hold. We observe distribution drift and outliers. Models are usually trained using proxy metrics such as accuracy or log-likelihood, but often these do not correspond to the metrics of interest for deployed models, which are more task-centric and application-specific. Moreover, a deployed model can be used for multiple tasks with differing objectives, and tasks can evolve over time with different performance aspects becoming more important. The proposed research program targets the development of reliable learning models and algorithms for graph-structured data. I will design graph learning methods and models that can: (i) quantify their uncertainty to facilitate improved decision making; (ii) detect and adapt to changes in the data distribution; and (iii) facilitate active, continual and sequential learning. We will develop flexible models that can successfully address multiple tasks and scale to extremely large graphs without overly sacrificing expressivity. A key aim is to support continual learning, where a model can be trained sequentially over time with multiple tasks. The cornerstone of our scientific approach is the Bayesian framework. This involves specifying prior distributions for model parameters and updating them based on observed data to construct posterior distributions. The Bayesian paradigm is well-suited to the goal of characterizing uncertainty. Since it targets posterior distributions rather than just point estimates, there is a natural mechanism for deriving confidence intervals. The Bayesian approach also provides the advantages of principled mechanisms for handling missing data, detecting outliers, and incorporating prior information. The approach is also often more effective for settings such as active learning where there is a limited amount of data, and is naturally aligned with a continual learning setting with streaming data and sequential updates.