Canada Spends Logo

Explorez notre analyse du budget 2026

Grant: Discovery Grants Program - Individual

Dalhousie University

$31,000.00

Towards Unbiased, Robust and Highly Scalable User Preference Estimation on Social Media Platforms

EF 2023-24

Résumé
The average Canadian spends 1.5 hours/day browsing content from social media platforms such as Facebook, Tiktok, and Twitter based on Statista report. To alleviate information overloading, the content users see is pruned and ranked by personalized recommendation algorithms, where user preference estimation plays a critical role. However, the personalized recommendation could be seriously biased and even polluted by actors seeking profit or disinformation. Consequently, the content pushed to readers may be one-sided, leading to wrong perceptions of the facts or triggering poor decisions. Recent research revealed that content spread on social networks could significantly impact human behaviour; fashion trends, consumption habits, and even political decisions could be influenced by the fermentation of posts imperceptibly. The recommender systems are not innocent as they can intensify this influence. Thus, preventing the personalized recommender system from exacerbating bias and harmful social impact has become a key research challenge. Recent research often focuses on categorizing the source of biases, evaluating their potential impact, or mitigating them, but overlooks how the biases corrupt a recommender system. While a few causal inference models aim to estimate the relationship between sources of bias and user behaviours (e.g., clicks, shares, and favorites), they often do not model the internal components of recommender systems explicitly. My proposed research will take a fresh look at the problem of bias handling in large-scale personalization systems and will focus on enhancing reliability of user preference estimation (UPE). In general, a recommender system is a software system that 1) calculates similarity score between user and item given user preference and item attribute estimations and then 2) sorts the candidate items based on the similarity and other criteria (e.g., diversity). Considering the simplicity of similarity computation and objectiveness of item attribute extraction, UPE stands out as the most vulnerable component to ignorantly encode bias. The dynamic and subjective nature of user preference may foster the absorption of bias. My research program includes two main themes. Theme 1 focuses on understanding how data bias corrupts the UPE model. We propose two projects that complement each other. Project 1 aims to detect user preference bias through counterfactual explanation and the deep generative model. Project 2 introduces a context-aware conversational system to verify the potential bias encoding. Theme 2 focuses on achieving unbiased user preference tracking. There are also two projects (3 and 4). Project 3 leverages the attention mechanism to reach context-aware training control, so user preference can be disentangled from influence of social events. Project 4 incorporates the Mixture of Experts framework into sequential modelling such that UPE model is equivalently robust to different user groups.
Programme
Discovery Grants Program - Individual
Bénéficiaire
Emplacement
Nova Scotia, Canada
Type de bénéficiaire
Academia
Type
Grant
Source
NSERC's Awards Data

Tel que publié

Application id
RGPIN-2023-05913
Name nom
Wu, Ga
Department département
Faculty of Computer Science
Organization id
79
Institution établissement
Dalhousie University
Province en
Nova Scotia
Province fr
Nouvelle-Écosse
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2023
Competition year année de concours
2023
Award amount
31000
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
1507
Committee name en
Computer Science
Committee name fr
Sciences informatiques
Area of application code
800
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 and communication services
Area of application fr
Services d'information et de télécommunications
Research subject code
2604
Research subject group en
Robotics
Research subject group fr
Intelligent systems applications
Research subject en
Robotique
Research subject fr
Applications de systèmes intelligents
Application title
Towards Unbiased, Robust and Highly Scalable User Preference Estimation on Social Media Platforms
Keyword
Conversational System; Deep Generative Model; Deep Learning; Large-scale Application; Probabilistic Graphical Model; Recommender System; Reliable Machine Learning; Responsible AI; User Preference Estimation
Application summary
The average Canadian spends 1.5 hours/day browsing content from social media platforms such as Facebook, Tiktok, and Twitter based on Statista report. To alleviate information overloading, the content users see is pruned and ranked by personalized recommendation algorithms, where user preference estimation plays a critical role. However, the personalized recommendation could be seriously biased and even polluted by actors seeking profit or disinformation. Consequently, the content pushed to readers may be one-sided, leading to wrong perceptions of the facts or triggering poor decisions. Recent research revealed that content spread on social networks could significantly impact human behaviour; fashion trends, consumption habits, and even political decisions could be influenced by the fermentation of posts imperceptibly. The recommender systems are not innocent as they can intensify this influence. Thus, preventing the personalized recommender system from exacerbating bias and harmful social impact has become a key research challenge. Recent research often focuses on categorizing the source of biases, evaluating their potential impact, or mitigating them, but overlooks how the biases corrupt a recommender system. While a few causal inference models aim to estimate the relationship between sources of bias and user behaviours (e.g., clicks, shares, and favorites), they often do not model the internal components of recommender systems explicitly. My proposed research will take a fresh look at the problem of bias handling in large-scale personalization systems and will focus on enhancing reliability of user preference estimation (UPE). In general, a recommender system is a software system that 1) calculates similarity score between user and item given user preference and item attribute estimations and then 2) sorts the candidate items based on the similarity and other criteria (e.g., diversity). Considering the simplicity of similarity computation and objectiveness of item attribute extraction, UPE stands out as the most vulnerable component to ignorantly encode bias. The dynamic and subjective nature of user preference may foster the absorption of bias. My research program includes two main themes. Theme 1 focuses on understanding how data bias corrupts the UPE model. We propose two projects that complement each other. Project 1 aims to detect user preference bias through counterfactual explanation and the deep generative model. Project 2 introduces a context-aware conversational system to verify the potential bias encoding. Theme 2 focuses on achieving unbiased user preference tracking. There are also two projects (3 and 4). Project 3 leverages the attention mechanism to reach context-aware training control, so user preference can be disentangled from influence of social events. Project 4 incorporates the Mixture of Experts framework into sequential modelling such that UPE model is equivalently robust to different user groups.