Environmental and Occupational Health Sciences Institute
Rutgers University Community RepositoryNot a member yet
58345 research outputs found
Sort by
Essays on assortative mating in marriage
No full textIn Chapter 1, increased assortative mating by educational attainment within the U.S. marriagemarket since 1960 is found. Unlike previous research, I study assortative mating by educational
attainment for different educational groups instead of the whole population. Using the samples
from the census and ACS from 1960 to 2015, I find evidence to support increased assortative
mating in the marriage market among the less educated and decreased assortative mating among
the highly educated. Previous research has tried to tie assortative mating to increased income
inequality in the U.S. since 1960 but has found ambiguous results. My findings prove why
assortative mating is overestimated in explaining increased income inequality. In addition, I find
some other changes in the marriage market should be helpful in explaining the increasing income
inequality since the 1960s. Specifically, increasing married female labor participation and
changing family earning structure contributed more to increased income inequality over the last
several decades than the changes in assortative mating. The marriage rate has declined after WWII. However, in the marriage market, assortativemating has risen; people are more likely to marry someone with a similar culture or economic
features. Previous papers usually study the effect of assortative mating by education attainment.
In Chapter 2, my paper studies assortative mating in a bachelor's degree field. The sample from
the American Community Survey (ACS) for the years 2009 to 2021 was used in this paper. I found
strong assortative mating in the marriage market by degree field. Meanwhile, similar to previous
research like Greenwood et al. (2014), I found that assortative mating by degree field would
increase household income inequality. Additionally, I found that assortative mating by degree field
strongly and significantly impacts endogamy. Mating with a spouse with the same degree field will
increase the probability of endogamy (marrying someone of the same race). In Chapter 3, this study examines the association between spousal working hours and men'sand women's employment when spouses have different education attainments. Working longer has
become increasingly common since the Industrial Revolution. With computers and a globalized
workforce, the competition for jobs has intensified. Using the sample from ACS from 2000 to 2020,
I find that having a husband who works long hours significantly correlated with his wife's odds of
unemployment, especially when he has higher education attainments. Furthermore, having a wife
who works long hours is associated with her husband's odds of unemployment when she holds
advanced degrees but goes in the opposite direction when she doesn't finish college. The lack of
longitudinal data makes it difficult to reveal the causality, and it needs to be revisited.Ph.D.Includes bibliographical reference
Historical evolution and future retreat of a lagoon-terminating glacier in Alaska
No full textBear Glacier Lagoon within Kenai Fjords National Park, Alaska, USA, is a quickly evolving glacier-lagoon system driven by recurrences of glacial lake outburst floods. The lagoon is approximately 70 years old, and very little information exists on the timing of these outburst floods, their effect on the local landscape and glacier, and changes in the properties of water that circulate through the lagoon. Here, we combine limited historical data with recent lagoon salinity measurements to describe the historical evolution and modern dynamics of this glacier-lagoon system, which exhibits both lake-terminating and tidewater attributes. Specifically, we present the evolution of Bear Glacier Lagoon over the last half century using historical maps and satellite imagery in the context of glacial retreat. Further, we describe the recent physical changes to the lagoon following a number of outburst flood events through time series charts that may begin to explain the glacier’s fast retreat and the lagoon’s new circulation patterns. Landsat imagery starting from 1984 allows for the creation of yearly lagoon outlines to better observe the retreat of the glacier and determine how the changes to the structure of the moraine are related to periodic outburst floods identified by National Park Service monitoring. The reconstruction of Bear Glacier Lagoon’s history and modern development may provide a base analysis to the new concept of glacier-lagoon system dynamics as they become more frequent among rapidly retreating Arctic glaciers.M.S.Includes bibliographical reference
Artificial intelligence aided track degradation prediction
No full textTrack geometry defects and broken rails are among primary causes of freight-train derailments in the United States. Track degradation modeling and prediction can be used to support preventive maintenance by identifying and prioritizing track segments that are more prone to track defects, thus reducing accidents and service disruptions.This dissertation develops a machine learning approach to predict foot-by-foot track geometry change accounting for spatial and temporal dependencies of track geometry change. First, a robust position synchronization method is proposed to geospatially align track geometry data from multiple inspections accounting for positional errors, which provides high-quality geometry data for track condition evaluation and prediction. Subsequently, based on the processed data, a spatial-temporal model, CNN-LSTM, is proposed for foot-by-foot track geometry prediction by effectively capturing spatial and temporal information. A case study has been conducted using data from one Class I railroad. It shows that the positional error correction method can reduce the relative positional error to less than 1 foot with a 99% confidence interval. The proposed CNN-LSTM model outperforms four alternative models, including a naïve model (using the last observation), multi-layer perceptron (MLP), plain CNN, and plain LSTM, for both short-term and long-term prediction periods.
Additionally, this dissertation adopts an ensemble learning framework to predict fatigue-related broken rail by extracting features from the multi-source databases related to infrastructure, operation, inspection, and weather conditions. Three data treatment methods (i.e., no resampling, undersampling, and oversampling) are incorporated into the proposed models, which include logistic regression, random forests, and gradient boosting. The findings show that gradient boosting (specifically XGBoost) with undersampling approach outperforms other machine learning methods for the studied dataset. It also identifies features that have a significant influence on broken rail prediction, such as the ratio of annual traffic density to rail weight, days since last broken rail, and days since last rail defect. Using this model, when 10% of the railroad network is scanned, approximately 33% of broken rails can be captured using the proposed methods.
Finally, a spatial-temporal neural network model based on ResNet-Transformer architecture is developed for broken rail prediction to consider both spatial and temporal dependencies. Utilizing the ResNet architecture, the proposed model captures detailed spatial correlations among rail characteristics. Subsequently, the Transformer architecture is utilized to effectively handle long-term temporal data patterns and dependencies that reflect dynamic changes over time. Experiment results demonstrate that ResNet-Transformer model significantly outperforms the traditional machine learning approach, especially in identifying segments that might be more prone to broken rails. When screening 10% of the railroad network, the model can capture approximately 42% of broken rails. It indicates the effectiveness of the proposed ResNet-Transformer framework for broken rail prediction. The proposed artificial intelligence aided framework can aid in proactive maintenance and strategic capital planning in the railroad industry.Ph.D.Includes bibliographical reference
Advancing efficient deep learning: an exploration of model compression from classification to generative models
No full textDeep Neural Networks (DNNs) have gained prominence across a range of applications, consisting of computer vision, speaker recognition, and natural language processing. Nowadays, there has been a consistent trend towards designing large neural networks due to the perception that bigger models yield better performance and even exhibit characteristics of “real intelligence”. The rapidly increasing size of these models, however, brings challenges for deployment on resource-constrained platforms, such as mobile devices, and embedded devices. Direct deployment on such platforms is difficult due to memory constraints and the time-intensive nature of the current large-scale model. To address the deployment challenges with these constraints, this thesis explores three prevalent model compression techniques to enhance efficiency in the AI era: pruning, low-rank decomposition, and quantization. These techniques are applied to deep neural networks, including image classification models and text-to-image generative models. Specifically, we first propose systematically investigating the quantification metric, measuring scheme, and sensitiveness and reliability of the channel independence in the context of filter pruning. Second, we start by exploring structured sparsification and low-rank decomposition toward compact DNN models, then propose a unified DNN compression framework that can co-explore the benefits of structured sparsification and tensor decomposition efficiently. Third, we compress large-scale text-to-image models to extremely low bit-widths by conducting a comprehensive analysis that examines the impact of each layer during quantization. We establish a mixed-precision strategy and propose a series of effective techniques for quantizing such generative models. This study on model compression aims to provide insightful approaches for efficient deep learning in deploying AI systems, significantly enhancing energy efficiency and reducing resource consumption.Ph.D.Includes bibliographical reference
Sources of Per- and Poly-Fluoroalkyl Substances to the influent of municipal wastewater treatment plants
No full textThe objective of this study was to investigate the sources of and processes affecting per- and polyfluoroalkyl substances (PFAS) in Wastewater Treatment Plants (WWTPs). A data set of paired samples of the influent and effluent collected from 27 WWTPs across California and analyzed for a suite of PFAS was obtained from the California Geotracker database and was analyzed using Positive Matrix Factorization (PMF). The data was compiled into two ways, first with each sample (influent and effluent) entered into data (set A) independently, giving rise to 16 analytes measured in 154 samples, and the second (set B) with each pair of influent and effluent treated as a single sample (30 analytes in 52 samples). Sixteen PFAS compounds were above detection often enough to be included in the analysis. All but two of the 16 PFAS (PFPeS and NMeFOSE) displayed higher concentrations in the effluent than in the paired influent, suggesting minimal partitioning into the biosolids and generation of measured PFAS compounds from precursors during treatment. The 16x154 data set generated five factors that represent PFAS sources, including conversion of precursors to perfluorocarboxylic acids (PFCAs) during transport. PMF analysis generated three fingerprints that described PFAS sources from the 30x52 data set. The first fingerprint showed high PFOA, PFPA, and PFHA contributions in both influent and effluent, suggesting that it represents PFAS that are already stable upon entering the WWTP. The second fingerprint showed high contributions of PFOS, PFHxS, and 6.2 FTS in the influent and increased amounts of PFOA, PFPA, PFHA, and other PFCAs in the effluent. This source therefore appears to represent precursors converted to PFCAs during treatment. The third fingerprint was characterized by very low contributions of all measured PFAS in the influent coupled with high levels of PFAS (PFOA, PFPA, and PFHA) in the effluent, suggesting that it represents PFAS from unmeasured precursors converted to PFCAs during treatment. The partitioning of PFAS to biosolids and the generation of PFAS from unmeasured precursors complicates the calculation of how much of the PFCAs in the effluent are generated during treatment. The approach of combining one pair of influent and effluent samples as one sample in the PMF input allows a reasonable estimate of this fraction. The results indicate that 40% to 80% of PFCAs in the effluent are products of the conversion of precursors during treatment.M.S.Includes bibliographical reference
Low separation rank in tensor-variate generalized linear models
No full textSince many real-world measurements are inherently multidimensional – involving multiple dimensions or modes – tensors, which extend vectors and matrices to multiple dimensions, are crucial in handling such data. Recent research in machine learning and statistics has shown that the performance of traditional vector- based approaches in solving regression problems often falls short in high-dimensional and small sample regimes. Many tensor-structured machine learning problems are high-dimensional, exemplifying the need for flexible tensor decompositions to tune model complexity to match the available data. Specifically, when the data is tensor structured, imposing tensor structures on the coefficient tensor in regression problems can lead to a more reliable parameter estimation and lower sample complexity compared to vector-based methods. The choice of tensor structure to be imposed affects the representation power and generalizability of the regression model parameter. Compact tensor structures greatly reduce sample complexity yet may not capture more complex interactions in tensor data. Conversely, more flexible tensor structures may suffer from a larger variance and sample complexity. This dissertation investigates a new low-rank tensor model, called Low Separation Rank (LSR), in Generalized Linear Model (GLM) problems. The LSR model – which generalizes the well-known Tucker and CANDECOMP/PARAFAC (CP) models, and is a special case of the Block Tensor Decomposition (BTD) model – is imposed onto the coefficient tensor in the GLM problem as a means to enhance model representation power whilst maintaining an acceptable level of compactness. This dissertation explores both the theoretical and algorithmic aspects of parameter estimation in LSR tensor- structured GLMs (LSR-TGLMs). It is comprised of three parts. The first part of this dissertation establishes the learning theoretic framework of Tensor Generalized Linear Models (TGLMs), through the lens of LSR tensor-structured regularizers. To estimate the LSR- structured model parameter in such problems, a numerical method based on alternating minimization / block coordinate descent is developed. The computational advantage of such methods is highlighted by the fact that a larger non-convex problem can be divided into smaller convex subproblems. Numerical experiments on synthetic datasets demonstrate the efficacy of the proposed LSR tensor model for three regression types (linear, logistic and Poisson). Experiments on a collection of medical imaging datasets demonstrate the usefulness of the LSR model over other tensor models (Tucker and CP) on real, imbalanced data with limited available samples. The second part of the dissertation delves into the theoretical aspects of the LSR-TGLM problem. It ex- amines the fundamental limits on estimation error through minimax and excess risk bounds. These bounds are critical for understanding the sample complexity of the parameter estimation problem in the LSR-TGLM framework, and provide a benchmark for algorithm efficiency. The minimax lower bound on the estima- tion error is proportional to the intrinsic degrees of freedom in the LSR-TGLM problem, suggesting that its sample complexity may be significantly lower than that of vectorized GLMs. These bounds can also be specialized to CP and Tucker-structured GLMs, and their tightness is assessed numerically and through the derivation of sample complexity results (excess risk bound). Additionally, sufficient conditions for local identifiability of the LSR-structured tensor in the LSR-TGLM problem is discussed and results on asymp- totic consistency and normality are established. This involves proving new properties of LSR-structured tensors, particularly compactness of the set of LSR-structured tensors. The asymptotic consistency result shows that maximum likelihood estimates of the parameter estimation problem converge to the true values as sample size increases. The asymptotic normality result shows that the maximum likelihood estimates converge to a normal distribution. The third part of the dissertation extends the algorithmic framework of LSR-TGLMs to the federated learning setting. Federated learning enables collaborative analysis without centralizing data, which is par- ticularly relevant for biomedical imaging where data is often distributed across multiple sites. By balancing local update steps with communication rounds, the federated algorithm achieves performance comparable to centralized methods, thus enabling effective collaborative analysis without the need for data centralization.Ph.D.Includes bibliographical reference
Weight Loss Promotion in the Arabic-Speaking Population
No full textPurpose of the Project: The primary aim of the project was to achieve weight loss over the course of 8 weeks through individual coaching on the use of My Fitness Pal, available on a SMART phone with a cadre of Arabic-speaking participants from a local Mosque in northern NJ.Methods: Overweight and obese members of the local mosque were invited to participate if they had any interest in losing weight. Participants tracked their nutrition and exercise for 8-weeks. Weekly coaching check-ins were held with each individual. During this time their latest weight and their compliance to the diet and exercise recommendations were collected. Data on weight and blood sugars were compiled into an excel file and then analyzed using SPSS
Results: A statistically significant decrease in weight with a mean difference of 5.61 pounds (SD= 3.07, 95% CI [4.08, 7.14]), t(17) = 7.75, p <.001) was seen. A statistically significant decrease in mean blood glucose with a mean difference of 13.61 (SD= 10.29, 95% CI [8.50, 18.73]), t(17)= 5.62, p <.001 was experienced. Five out of the 18 participants reverted to a normal BMI of less than 25 by the conclusion of the project. Frequency analysis for nutrition and exercise tracking revealed a decline in tracking.
Implications for Practice
The use of a mobile application can help promote adherence to necessary lifestyle modifications for weight loss. A weight loss of 5-10% of baseline weight can improve quality of life, reduce complications and cost related to obesity. Medical care costs obese people $1,861 more than a person at a healthy weight.
Keywords: Arabic-speaking, obesity, weight loss, diet and exercise, coaching health coaching
D.N.P.Includes bibliographical referencesIncludes vit
Screening for Unhealthy Alcohol Use in a Primary Care Setting
No full textPurpose of Project: The U.S. Preventive Services Task Force (USPSTF) recommends screening all adults for unhealthy alcohol use in the primary care setting and providing those who screen positive with brief intervention. Despite these recommendations, patients are seldomly screened or treated in primary care. This quality improvement project sought to increase the number of people in a primary care setting who were screened for unhealthy alcohol use.Methodology: Clinicians at a primary care practice in NJ (n=4) were educated on the significance of unhealthy alcohol use as well as best practices for screening and brief intervention for those who screen positive. Surveys assessing provider knowledge, attitudes, and practices were administered prior to and immediately following the training. A third survey was done 4 weeks after the training to assess for short-term impact. Additionally, AUDIT-C questionnaires were administered to and collected from patients seen for an annual visit (n=104).
Results: Provider likelihood to use a validated screening tool increased by 85.7%, provider comfort level on screening increased by 16.7%, provider comfort level on providing brief intervention increased by 62.5%, and the number of patients that were screened using a validated screening tool increased by 71.4% after education and implementation.
Implications for Practice: A validated screening tool should be used to screen patients for unhealthy alcohol use in the primary care setting and brief intervention should be provided to those who screen positive. The practice change measured in this project occurred after efforts to enhance provider knowledge and comfort level were provided. Therefore, increasing the number of those that are screened for unhealthy alcohol use starts with education and awareness.D.N.P.Includes bibliographical referencesIncludes vit
Efficient and secure deep learning systems for resource-constrained environments
No full textWe outlines a comprehensive approach towards establishing secure and efficient deep learning systems, addressing key challenges in model compactness, robustness, and defense against emerging threats. Deep neural networks (DNNs) are increasingly integral in various applications, necessitating enhanced security measures. The proposed research encompasses several novel methodologies to fortify DNNs against adversarial and backdoor attacks, while simultaneously ensuring model efficiency and compactness. Firstly, we introduce CSTAR, an innovative solution that enforces Compactness, high Structuredness, and Adversarial Robustness in DNN models. CSTAR's unique framework allows compressed DNNs to achieve high compression performance alongside strong adversarial robustness, significantly improving both benign and robust accuracy over existing methods. In response to the rising concern of backdoor attacks on DNNs, particularly on compressed models, our research proposes IEBA and RIBAC. IEBA introduces invisible and efficient backdoor attacks for compressed deep neural networks. RIBAC goes one step further to leverages systematic analysis to optimize trigger patterns, model parameters, and pruning masks, achieving high trigger stealthiness, attack success rate, and model efficiency. Addressing the need for effective backdoor defense in resource-constrained environments, we present Clean & Compact (C&C), a data-free backdoor defense mechanism. C&C capitalizes on the rank-level sensitivity to trigger patterns, offering model purification and efficiency without relying on training data. This approach is particularly suited for real-world scenarios where access to clean data is limited. Lastly, the thesis explores the security implications of diffusion models in generative AI, particularly text-to-image generation. We introduce the Latent Diffusion Shield (LDS), a novel protection method operating within the latent space of diffusion models. LDS effectively counters unauthorized diffusion-based image synthesis, establishing new benchmarks in image protection with significant improvements in processing speed and memory usage. Overall, this thesis aims to advance the field of deep learning by developing a suite of solutions that assure the security and efficiency of DNNs, thereby enabling their safe and effective deployment in various critical applications.Ph.D.Includes bibliographical reference
Essays on Financial Intermediaries and Institutional Shareholders
No full textThis dissertation explores the roles of financial intermediaries and institutional shareholders in the context of advancing technology and emerging governance challenges. The first chapter investigates the impact of 3G mobile internet access on auditing process, demonstrating how enhanced communication and data transmission capabilities can reduce audit costs. The second chapter examines the role of mutual funds in regulating politically-motivated charitable giving (PCG) by corporations, a controversial corporate practice identified in the past decade. The new empirical evidence presented in this dissertation contributes to our understanding of how effectively auditors and mutual funds adapt to technological or governance shifts.Ph.D.Includes bibliographical referencesIncludes vit