Grant: Discovery Grants Program - Individual
McGill University
$34,000.00
Causality and Learning under Selection Bias
EF 2024-25
Résumé
Causality has been a subject of prime interest in statistical inference since the inception of statistics as a branch of science. Starting with Fisher's work on design of experiments, and in particular the notion of randomization, leading to Cochran's seminal papers in 60s and early 70s on observational studies in human populations followed by the work of other statisticians and econometricians over the past five decades, there is now a vast literature on causality. In contrast, the precursor question as to how we may identify causal directions has, however, been considered to a much lesser extent. If we hope to take learning to another level where we can have beyond domain generalization, identification of causal direction and learning the cause and effect relationships are imperative. The recent work of machine learners has brought us useful and important techniques to address this question. An underlying assumption to the methodology in the current literature is the so-called independence of cause and mechanism. This fundamental assumption can fail to hold when there is a possibility of selection bias. Applying the usual methods without proper adjustment for selection bias may lead to mis-identification of causal directions. Other complexities, such as incomplete data or missingness, can compound the issue. Although the literature on selection bias in machine learning is not as extensive, there has been a growing interest in the community, particularly among those who study large language models.
Over the next five years, I aim to study the issue of selection bias and causal discovery where in addition to selection bias the training data may also be incomplete. I am particularly interested in incompleteness caused by censoring, a common phenomenon in follow-up studies often conducted by medical and epidemiological researchers. To the best of my knowledge, there is very little on training Neural Networks (NN) using censored data. When the goal is prediction, such studies can lead to solid and robust methods for prediction using incomplete data without making restrictive model assumptions and fill a big gap in our current literature on lifetime data analysis. This is particularly important when the number of covariates (features, attributes) are large. Statistical properties of Deep Neural Networks (DNN) has only been formally studied recently. The existing literature is limited and only covers complete representative data from the population of interest, i.e. no selection bias or censoring. Developing such results under selection bias for incomplete data is one of my goals. The interplay between the hyper parameters of DNN and the dimension of the input data manifold, the intrinsic dimension, and the error rate of DNN under selection bias and censoring is another project I aim to work on. I also want to invest time on the Missing Mass problem under selection bias and its ramifications in DNN, particularly language models.
Programme
Discovery Grants Program - Individual
Bénéficiaire
Emplacement
Quebec, Canada
Type de bénéficiaire
Academia
Type
Grant
Source
NSERC's Awards Data
Tel que publié
Application id
RGPIN-2024-05640
Name nom
Asgharian, Masoud
Department département
Mathematics and Statistics
Organization id
61
Institution établissement
McGill University
Province en
Québec
Province fr
Québec
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2024
Competition year année de concours
2024
Award amount
34000
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
1508
Committee name en
Mathematics and Statistics
Committee name fr
Mathématiques et statistique
Area of application code
801
Area of application group en
Information and communication services
Area of application group fr
Services d'information et de télécommunications
Area of application en
Computer software
Area of application fr
Logiciel
Research subject code
3010
Research subject group en
Statistics and probability
Research subject group fr
Statistique et probabilité
Research subject en
Nonparametric inference
Research subject fr
Inférence non paramétrique
Field of research list names en
Not available
Field of research list names fr
Non disponible
Socioeconomic objective list names en
Not available
Socioeconomic objective list names fr
Non disponible
Application title
Causality and Learning under Selection Bias
Keywords
lifetime data; selection bias; censoring; causal discovery; independence of cause and mechanism; semi-supervised learning; nonparametric methods; semiparametric methods; universal approximation; large sample theory
Application summary
Causality has been a subject of prime interest in statistical inference since the inception of statistics as a branch of science. Starting with Fisher's work on design of experiments, and in particular the notion of randomization, leading to Cochran's seminal papers in 60s and early 70s on observational studies in human populations followed by the work of other statisticians and econometricians over the past five decades, there is now a vast literature on causality. In contrast, the precursor question as to how we may identify causal directions has, however, been considered to a much lesser extent. If we hope to take learning to another level where we can have beyond domain generalization, identification of causal direction and learning the cause and effect relationships are imperative. The recent work of machine learners has brought us useful and important techniques to address this question. An underlying assumption to the methodology in the current literature is the so-called independence of cause and mechanism. This fundamental assumption can fail to hold when there is a possibility of selection bias. Applying the usual methods without proper adjustment for selection bias may lead to mis-identification of causal directions. Other complexities, such as incomplete data or missingness, can compound the issue. Although the literature on selection bias in machine learning is not as extensive, there has been a growing interest in the community, particularly among those who study large language models.
Over the next five years, I aim to study the issue of selection bias and causal discovery where in addition to selection bias the training data may also be incomplete. I am particularly interested in incompleteness caused by censoring, a common phenomenon in follow-up studies often conducted by medical and epidemiological researchers. To the best of my knowledge, there is very little on training Neural Networks (NN) using censored data. When the goal is prediction, such studies can lead to solid and robust methods for prediction using incomplete data without making restrictive model assumptions and fill a big gap in our current literature on lifetime data analysis. This is particularly important when the number of covariates (features, attributes) are large. Statistical properties of Deep Neural Networks (DNN) has only been formally studied recently. The existing literature is limited and only covers complete representative data from the population of interest, i.e. no selection bias or censoring. Developing such results under selection bias for incomplete data is one of my goals. The interplay between the hyper parameters of DNN and the dimension of the input data manifold, the intrinsic dimension, and the error rate of DNN under selection bias and censoring is another project I aim to work on. I also want to invest time on the Missing Mass problem under selection bias and its ramifications in DNN, particularly language models.