Zubia Mansoor

LASSO

The Least Absolute Shrinkage and Selection Operator (LASSO) is a variable selection and regularization method in regression analysis. As a part of my graduate course on Linear Models and Applications, I aim to provide a brief overview of this technique.

Publication

A hybrid regression model for water quality prediction. Chakraborty, T., Chakraborty, A.K. & Mansoor, Z. Springer, OPSEARCH 56, 1167–1178 (2019). https://doi.org/10.1007/s12597-019-00386-z

As a data science intern at the Indian Statistical Institute, Kolkata, I worked on a project in collaboration with senior statisticians and chemical engineers. Our work proposes a hybrid regression machine learning model to solve a specific problem faced by a modern paper manufacturing company.

Image Classification using WasteNet

1st Place at Vancouver Datajam Hackathon 2020

As the team lead at my first ever hackathon, my team and I built an image classification algorithm to correctly classify images of waste items for proper garbage disposal in Greater Vancouver area. Our nine-member team were tasked to predict the category for different waste items using image classification techniques in machine learning and computer vision. For our project, we made use of secondary trash data maintained by Github users and supplemented it with pictures of trash that the team took themselves during the intense one-day hackathon. Our final solution resulted in using pre-trained ResNet-19 features and fine-tuning them to train a classification network. This is very relevant in that classifiers like ours can be used to correctly identify and sort recyclable items to reduce environmental contamination.

Team Members: Seyed Mahdi Hosseini Miangoleh, Saghar Irandoust, Salina, Warren Ho Kin, Nasreen Mohsin, Hans Sy, Kevin Kang, Amir Parizi (Mentor), Zubia Mansoor (Team Lead)

Predicting Mortgage Default Risk

4th Place at SFU Business Analytics Hackathon 2020

Tasked with solving a real‑world business problem in under 3 hours, my team and I built an ensemble classifier to predict mortgage default for a fintech startup. Our final model achieved an AUC of 73% on a hold‑out test set that averaged the predicted class probabilities by a neural net and random forest. In order to build strategies to target in the upcoming advertising campaigns, we used k-means clustering to help identify key customer demographics. We presented our findings in five rounds to batches of judges sequentially, incorporating each group's feedback before pitching to the next round.

Team Members: Nikola Surjanovic, Claudine Tu, Zubia Mansoor

SARS-CoV-2 Host Genetics

As a part of the Canadian Statistical Sciences Institute (CANSSI) rapid response grant, my lab and I conducted host genetic association studies for SARS-CoV-2 (COVID-19). We aimed to better understand how human genetic variations affect COVID-19 susceptibility informing vaccine and drug development. This manuscript will serve as stepping stone for futher work at CGEn, Canada's national platform for genome sequencing and analysis

Abstract of the Paper

Identifying genetic risk factors explaining why some people are more susceptible or severely affected by COVID-19 than others may give clues to underlying vulnerabilities, and is of great interest to the host genetics community. Due to the way pandemics spread, genetic variation is confounded with several factors like ethnicity, socioeconomic indicators and occupation. This work aims to explore the common biological links between neurodegenerative diseases and COVID-19 after controlling for potential confounders. We employ methods from genome-wide association studies and categorical data analysis, and provide results from a power analysis study we undertook to detect genetic associations. Researchers may use results from our host genetics work to generate hypothesis about antigen binding or etiology, to explore protein networks, or to create links to well-studied genes allowing established work to inform COVID-19 directions.

Online Inference

Sequential Monte Carlo methods are a class of Monte Carlo algorithms used for estimating sequences of distributions and expectations, especially in signal processing and Bayesian statistical inference. In this project, we introduce Sequential Monte Carlo (SMC) as a more efficient alternative to Sequential Importance Sampling (SIS). In particular, SMC builds on the framework of SIS by incorporating a simple adaptation to the SIS algorithm which drastically improves the performance of the method.

Team Members: Renny Doig, Zubia Mansoor

Bayesian Model Averaging

Model Selection

Bayesian model averaging (BMA) provides a systematic and coherent approach for representing model uncertainty by constructing a probability distribution over all possible models, where each probability provides a measure of how likely the different models are. The fundamental idea is to make inferences based on weighted averages of the quantities of interest where the weights correspond to the model probabilities. In this project, we provide a brief overview of the BMA framework and demonstrate its use as a tool for model selection on kid’s cognitive dataset.

Team Members: Louis Arsenault-Mahjoubi, Sonny Min, Zubia Mansoor