Selected Talks

Efficiency and Effectiveness in Large-Scale Learning

May 28, 2019

Proposal Talk, Khoury College of Computer Sciences PhD Proposal, 208, WVH (TBD)

Thesis Proposal

Efficiency and Effectiveness in Large-Scale Learning

Authors

Bingyu Wang

Abstract

In the past decade, the data have grown faster than ever across many domains. For example, in survival analysis, there are more than 60 million Medicare beneficiaries across 40 thousand ZIP Code areas in United States from 2000 to 2012, which is up to 5.7 billion person-months of follow-up. In multi-label classification, Wikipedia data contains more than 500 thousand labels, millions of features and instances. For many of such datasets, machine learning models are facing unprecedented challenges associated with effectiveness and efficiency in both time and memory. This thesis aims to develop learning models that scale well on large data while being able to maintain or even increase its level of performance based on the inherent structures of the dataset and learning algorithms. By working on the following key questions for each model: 1) how to adapt to the intrinsic structure of the dataset and 2) how to take into account the special design of the learning formula and algorithm, we develop state-of-the-art algorithms for both regression and classification problems, and scale such algorithms well on multiple real-world datasets, such as millions of Medicare enrollees in survival analysis, Wikipedia articles and Amazon products categorization in multi-label classification.

Effect Modification of the Association of Long-Term PM_2.5 Exposure and Cause-Specific Mortality: An Analysis of 64 Million U.S. Medicare Beneficiaries

August 28, 2018

Conference Talk, ISES-ISEE 2018 Joint Annual Meeting, Ottawa, Canada

Authors

Bingyu Wang, Ki-Do Eum, Justtin Manjourides, Fatemeh Kazemiparkouhi, Helen Suh, Virgil Pavlu

Abstract

Introduction. Long-term exposure to fine particulate matter (PM_2.5) has been consistently associated with mortality; however, our understanding of how these associations are modified by certain characteristics, such as urbanicity and race, is more limited.
Methods. We considered demographic and mortality data for over 64 million Medicare enrollees across all 40 thousand ZIP codes in the conterminous United States from 2000 to 2008. We linked these data to ZIP code and month-specific PM_2.5 exposures estimated using GIS-based spatio-temporal models and fit the data using Cox proportional hazard model that were modified using data aggregation and limited-memory BFGS optimization. This modification allowed our Cox PH model to analyze data simultaneously for all 64 million Medicare beneficiaries within 10 minutes (Intel Xeon E5-2680, 56 cores). We used these models to estimate the association of PM_2.5 on all cause, cardiovascular (CVD), respiratory disease and cancers (including specifics: ischemic heart disease, cerebrovascular disease, congestive heart failure, COPD, pneumonia, upper respiratory infection (URI), lung cancer), with strata for age, race, gender, and ZIP code of residence. We examined effect modification by age, race, gender, land use, and region and assessed the linearity of the dose-response relationship for each cause and by modifier.
Results. We found significant positive associations of PM_2.5 and all causes of death, except URI, with mortality risk ratios (RR) for non-accidental, CVD, respiratory, and cancer mortality equaling 1.244 (95% CI: 1.238, 1.251), 1.683 (95% CI: 1.669, 1.696), 1.241 (95% CI: 1.223, 1.260) and 1.160 (95% CI: 1.147, 1.172) per 10 $\mu$g/m³ increase in PM_2.5, respectively. We showed risk of death to be higher for beneficiaries living in urban as compared to non-urban areas, for men as compared to women, for younger ages, and individuals living in the northeastern US. PM_2.5 associated RRs were similar for whites and non-whites.