UC Irvine

Digital markets are built on data, low-friction access, and user trust. This dissertation studies what happens when each of these pillars is disrupted. Across three essays, it shows that privacy regulation, registration-based access barriers, and public revelations of workplace misconduct can materially alter consumer purchasing, attention allocation, and platform choice. The first essay examines the unintended consequences of the California Consumer Privacy Act (CCPA) for consumer commerce. Using the CCPA as a natural experiment and combining billions of payment transactions with browsing data and firm-level measures of advertising technologies, the essay shows that Californians purchased less, returned more, and spent more time browsing after the law took effect. Firms subject to the regulation also reduced their use of ad-related technologies. These findings suggest that stronger privacy protections may unintentionally weaken personalization and product-consumer matching, reducing commercial activity even as they expand consumers' control over personal data. The second essay studies the New York Times' introduction of a registration wall as a first-party data collection strategy. Using web browsing data together with a regional measure of privacy salience based on cumulative exposure to data breaches, the essay finds that registration walls generate heterogeneous responses across user segments. News-singlehomers reduce engagement after the wall is introduced, whereas news-multihomers slightly increase engagement. Higher privacy salience attenuates the decline among singlehomers and strengthens the positive response among multihomers. When users disengage, their attention shifts primarily toward search engines and video-sharing platforms, revealing broader ecosystem consequences of registration-based access barriers for news consumption. The third essay investigates how public whistleblowing about workplace sexual harassment and discrimination affects consumer demand in platform competition. Exploiting a blog post by a former Uber employee describing sexual harassment, discrimination, and the company’s response—an account that quickly went viral on social media—as an exogenous reputational shock, the essay shows that Uber usage declines while Lyft usage rises among riders most able to switch between the two platforms. These effects are strongest in markets with greater local attention to the event and among riders predicted to be female, highlighting how misconduct toward minorities can translate into meaningful demand-side consequences in consumer markets. Taken together, the three essays demonstrate that consumer behavior in digital markets is shaped not only by prices and product attributes, but also by data governance, access design, and corporate legitimacy. The dissertation contributes to research on privacy regulation, digital media monetization, and platform competition, while offering practical implications for firms and policymakers seeking to balance personalization, access, and accountability in the digital economy.

The distributed training of machine learning models via gradient descent is generally conducted by iteratively computing the local gradients of a loss function and aggregating them across all processors. Communicating these gradients during aggregation is often a major cost but sparsification techniques can greatly improve efficiency. One such technique, Top-k gradient compression, ensures that only the k largest components of each local gradient are sent. However, effectively scaling this method can be challenging. The standard ring AllReduce algorithm, which is frequently used to aggregate dense gradients, lacks a counterpart that is optimized for sparse data. Notably, ring algorithms are contention-free, which generally make them easier to scale than other collective communication algorithms. Thus, in practice, the ring AllGather algorithm, which can be trivially adapted for sparse data, may be used instead, even though its bandwidth costs are proportional to the number of utilized processors (unlike ring AllReduce). To provide a more scalable contention-free alternative, we present a variant of ring AllReduce that has been better optimized for sparse data. We compare it to the standard dense ring AllReduce and ring AllGather algorithms, and we evaluate it empirically using gradients sampled from fine-tuning Llama 2 7b.

English academic writing is a critical yet challenging skill for learners in English as a Foreign Language (EFL) contexts. The rapid integration of digital tools, accelerated by the COVID-19 pandemic, has transformed writing instruction; however, evidence of its pedagogical effectiveness remains fragmented and often overlooks teacher perceptions and genre-specific impacts. This three-study dissertation addresses these gaps by investigating the role of digital tools in EFL writing instruction within the EFL higher education context, employing a multi-method approach to triangulate evidence from student outcomes, meta-analytic synthesis, and teacher experiences.Study 1 conducted a classroom experiment with 111 Chinese undergraduates, comparing infographic-based pre-writing to traditional outlining. Results showed that infographic creation significantly improved summary-writing quality, source use, and self-efficacy, but not opinion writing, highlighting the genre-sensitive nature of tool effectiveness.Study 2 synthesized 17 experimental and quasi-experimental studies (N = 1,085) through a meta-analysis. It found a statistically significant, moderate overall effect favoring web-based collaborative writing (WBCW) over non-technological collaboration on writing quality (Hedges’ *g* = 0.51). Moderator analyses indicated that training was a significant factor enhancing outcomes.Study 3 explored the perceptions of seven Chinese university writing instructors through qualitative interviews. Grounded in the TPACK framework, the findings revealed that while digital tools are deeply embedded in instruction, their use is shaped by teachers' knowledge, institutional support, and access. Teachers exhibited cautious and limited integration of emerging Generative Artificial Intelligence (GenAI) tools due to concerns over academic integrity and a lack of institutional guidance.Collectively, the findings demonstrate that technology enhances EFL academic writing most effectively when tools are aligned with genre demands, collaboration is scaffolded by design and training, and implementation is supported by teacher knowledge and institutional context. The dissertation concludes by advocating for a situated, tool-genre-task alignment perspective in future research and practice, moving beyond technocentric adoption to support equitable and effective writing instruction in diverse EFL settings.

Self-renewing hematopoietic stem cells (HSCs) produce billions of blood cells a day to maintain peripheral blood and immune cells in circulation. Inflammation can alter the balance of steady-state hematopoiesis, and disrupted hematopoiesis can lead to blood disorders or cancers. Over a lifetime of cell divisions, HSCs may acquire somatic mutations that provide a competitive advantage over their wild-type counterparts. The clonal expansion of a mutant HSC population is termed “clonal hematopoiesis” (CH). CH is linked to an increased overall risk of mortality due to incidences of cardiovascular disease and transformation into hematological malignancy. Tobacco and nicotine use remain the leading preventable drivers of cancer risk, and both direct and secondhand exposure to combustible cigarettes or electronic nicotine devices perturbs immune function and hematopoiesis. The World Health Organization estimates more than 100 million people across the world are using electronic cigarettes, or e-cigarettes, but the health impacts of these “safer” e-cigarette alternative have yet to be fully elucidated. E-cigarettes have been associated with inflammation and oxidative stress, which can provide selective pressures for the outgrowth of CHIP mutant cells.Here, we evaluate the impact of e-cigarette vapor and combustible cigarette smoke on in vitro cell inflammatory responses and in vivo long-term hematopoietic differentiation. The overarching goal of this project was to understand further how an inflammatory lifestyle stressor, such as smoking, contributes to aberrant hematopoiesis in wild-type normal stem cells, Tet2- deficient cells, and JAK2V617F mutant cells. In cell-based studies, cigarette smoke extract (CSE) and e-cigarette vapor extract (EVE) consistently suppress LPS-induced TNF-α secretion across macrophage/monocyte models, including primary mouse and human cells and complementary cell lines, indicating a reproducible immunosuppressive effect on mature myeloid cells. To assess consequences of smoking behavior in vivo, we used a custom nose-cone inhalation system to deliver controlled exposures to combustible cigarette smoke or e-cigarette aerosol to mice. Chronic exposure increased myeloid proliferation, a hallmark of HSC aging. Taken together, these results support a model in which tobacco exposures blunt innate immune responsiveness while simultaneously driving myeloid expansion conditions that accelerate hematopoietic aging and promote the expansion of mutant hematopoietic cells.

One-dimensional (1D) and quasi-1D (q-1D) van der Waals (vdW) materials, which have strong bonding along only one crystallographic axis, have emerged as powerful class of solids that hosts novel electronic, optical, and quantum properties useful for next-generation electronics. These materials can, in theory, be thinned to the angstrom scale due to the ideal vdW surfaces along two axes. In practice, achieving atomically precise single chains of these materials poses a significant challenge, as conventional top-down exfoliation and bottom-up growth techniques consistently retain interchain bonding, especially in q-1D materials with anisotropic interchain bonding motifs. The question then arises: Is there a synthetic approach that would enable the suppression of all inter-chain interactions, leaving only intra-chain covalent bonding in one dimension? Using the model q-1D pnictogen chalcogenides (Pn2Ch3; Pn = Sb, Bi; Ch = S, Se, Te), chosen for the highly anisotropic structural complexity and strong inter-chain bonding combined with distinct photophysical properties, we explore encapsulation within ultranarrow nanotube growth templates to precisely define the growth of the material in sub-nanometer length scales. The deployment of these nanotube templates as encapsulants overcomes the inter-chain interactions in these phases to isolate single q-1D chains by affording physical space that directly matches the size of a single chain. More importantly, we can also probe many chains simultaneously in a collective using conventional spectroscopic techniques used in ensemble samples, allowing us to access nanoscale properties through bulk measurements. Herein, not only do we gain insight into structural and electronic properties of these single chains, but we also further our understanding of anisotropic bonding in the bulk structure by understanding how it affects single chain accessibility. As technological devices advance through shrinking and densification, understanding material properties at the atomic limit becomes increasingly vital. Our results demonstrate a powerful tool for synthesizing and studying precisely defined single inorganic chains that approach the atomic limit.

In this thesis, we study efficient attention in the dominant transformer architecture from a theoretical perspective. We then use this theoretical understanding to guide the design of models for problems in time-series forecasting, image processing, and video understanding. In Part I, we begin with the motivation for and an introduction to modern machine learning, starting with recurrent neural networks for processing sequential data and continuing to the current transformer-based approach. We discuss how image processing has progressed from relying mainly on convolutional neural networks, which respect image geometry, to the current paradigm of transformers in computer vision. We provide an overview of the chapters in this dissertation and explain how they connect to the core research question. We then present a unified theoretical analysis of the attention mechanism, laying the theoretical groundwork for the remainder of the dissertation. In Part II, we consider applications to time-series forecasting. Chapter 3 presents the main work on developing FWin as an efficient model and comparing it with many state-of-the-art methods. In particular, we combine window attention with the Fourier transform as an efficient global mixing operator. Chapter 4 then validates the proposed FWin method on a more challenging dataset involving dengue cases in Singapore, where the data are nonstationary and quite limited. This demonstrates that FWin is robust enough to adapt to other applications rather than only perform well on standard benchmarks with well-known properties. In Part III, we develop a mathematically inspired efficient attention mechanism for image processing. We use arithmetic averaging as a global approximation of full attention while using window attention to mimic the local and recurrent nature of Mamba. This allows our proposed SEMA method to combine two well-known mechanisms in computer vision, retain their advantageous properties, and complement their weaknesses. We verify our design against state-of-the-art models on several challenging benchmarks. Following this success, we extend SEMA into a video model in the next chapter. We design a split space-time transformer in which SEMA processes the spatial component and standard softmax attention processes the temporal component. This design is efficient and scales to high-resolution videos. We conclude the dissertation by discussing the limitations of the current line of research and the importance of the works presented in this thesis. We also point toward future research directions in efficient machine learning and discuss their importance for real-world applications.

One-dimensional (1D) van der Waals (vdW) materials provide a unique platform that enables the understanding of how anisotropic bonding governs crystal growth and enables access to complex morphologies and emergent physical states. However, predictive synthetic strategies for directing their crystallization remain underdeveloped. This work establishes chemical design principles for directing the bottom-up crystallization of 1D vdW materials by controlling directional bonding, composition, and defect-mediated growth. Using modified chemical vapor transport, strong intrachain bonding—particularly Peierls-like metal dimerization—is shown to promote catalyst-free growth of ultralong, high-aspect-ratio nanowires in NbS3-I and MoI3. Building on this framework, compositional alloying and chalcogen deficiency emerge as complementary chemical design parameters that modify lattice flexibility and growth pathways that enable the formation of selfcoiled nanorings and screw-dislocation-mediated spiral hillocks while preserving the intrinsic anisotropic properties of the parent crystals. Microscopy, spectroscopy, and first-principles calculations collectively reveal the interplay between bonding anisotropy, lattice perturbations, and crystal growth. Together, these studies establish generalizable chemical design principles for programming crystal growth in burgeoning classes of 1D vdW materials and provide new synthetic routes to structurally complex crystals that serve as platforms for exploring structure–property relationships and emergent physical phenomena.

This dissertation investigates three of the most elusive phenomena in modern astrophysics: black holes, dark matter, and neutrinos, studied through theoretical modeling and large-scale numerical simulation to confront a new generation of gravitational-wave, electromagnetic, and neutrino observations.The first part addresses the astrophysical population of merging binary black holes (BBHs) detected by LIGO-Virgo-KAGRA. Using the population-synthesis code SEVN combined with a galaxy-evolution framework spanning cosmic star formation, metallicity, and galaxy stellar mass, I construct a model for the volumetric BBH merger rate density and its dependence on primary mass, secondary mass, and mass ratio, finding that the observed primary mass distribution requires either a top-heavy initial mass function in low-metallicity dwarf galaxies or a substantial dynamical-capture contribution above ∼ 30 M⊙. I then examine mass ratio reversal (MRR), binaries in which the initially less massive star forms the more massive compact object, across the SEVN and COMPAS codes. Both codes show that MRR leaves a distinct imprint on the merger-rate landscape, though the size and character of that imprint depends on the treatment of mass transfer and supernova physics. This makes clear that the primary-mass distribution inferred by LVK is not a direct tracer of the initially more massive stars, but a superposition of physically distinct evolutionary populations, an effect that must be accounted for to robustly connect gravitational-wave observations to massive binary evolution.The second part turns to the nature of dark matter. Motivated by JWST’s discovery of unexpectedly massive, rapidly assembled galaxies in the early Universe, I use cosmological N-body simulations to test whether self-interacting dark matter (SIDM) can enhance early star formation relative to cold dark matter (CDM). I find that SIDM leaves the halo mass function unchanged but produces systematically rounder halos and suppressed sub halo survival by z ≈ 6, with the total accretion rate onto hosts remaining comparable to CDM — indicating that self-interactions primarily enhance post-infill sub halo disruption rather than large-scale structure growth. Using TNG50-1, I estimate that this disruption liberates gas that could plausibly boost star-formation efficiency. I further show that gravothermal collapse, though analytically expected for a sizable portion of the simulated population, is not realized in these simulations, an absence attributable to the merger-dominated environment of early halo assembly.The third part develops a model-independent framework for extracting the total emitted energy and net deleptonization of the next Galactic core-collapse supernova from its neutrino signal, using B-spline spectral unfolding applied jointly across Super-Kamiokande, DUNE, and JUNO. I show that this approach recovers the total emitted energy with percent-level precision without assuming a parametric spectral shape, that a nonzero net deleptonization can be robustly established despite its larger reconstruction uncertainty, that DUNE’s dedicated νe channel is indispensable to the reconstruction, and that the resulting energy measurement is precise enough to serve as a direct, model-independent probe of the photoneutron star’s mass.Together, these studies demonstrate how upcoming and current observational facilities, spanning gravitational waves, deep imaging, and neutrino detection, can be used to test and constrain fundamental physics across some of the most extreme environments in the Universe

Background. Black African immigrant mothers in the United States navigate postpartum recovery while adapting to unfamiliar cultural, social, and healthcare contexts. Acculturative stress, postpartum distress, maternal-infant bonding, and social support have each been examined in immigrant perinatal populations, but not together, and not in this group.Purpose. This study examined how Black African immigrant mothers understand and experience acculturative stress, postpartum distress, maternal-infant bonding, and social support.Methods. A qualitative design informed by hermeneutic phenomenology was used. Sixteen Black African immigrant mothers with a child aged 24 months or younger born in the United States completed semi-structured interviews. Data were analyzed using reflexive thematic analysis, supported by reflexive and analytic memos.Results. Six themes were developed: navigating cultural adaptation and acculturative stress, discrimination and identity-linked othering, the erosion and reconfiguration of communal support, psychosocial distress and coping, maternal-infant bonding amid postpartum and acculturative demands, and the spectrum of social support experiences. Three findings organized these themes. First, acculturative stress operated as a multidimensional, contextual process in which discrimination was embedded while also exerting a distinct effect. Second, postpartum depression, in particular, and maternal-infant bonding were interconnected, with mothers actively working to sustain connection with their infants rather than experiencing bonding as a passive outcome. Third, social support was conditional: it was experienced as effective when it both fit mothers' functional needs and was culturally congruent with their current cultural frame.Conclusion. Postpartum distress and maternal-infant bonding among Black African immigrant mothers were embedded within broader social, cultural, and structural conditions rather than being isolated individual experiences. These findings expand our understanding of postpartum social support for immigrant mothers and highlight the need for perinatal research and care systems to go beyond simply assessing the presence of support toward ensuring that care is accessible, culturally informed, and responsive to each mother’s unique circumstances.

Despite their inability to been seen with the naked eye, nanopores are ubiquitous to our cell membranes to regulate the passage of ions and molecules. Their superior selectivity, transport and sensitivity to external stimuli have inspired the innovation of solid-state nanopores for applications in single molecule sensing, label-free sequencing, filtration, energy harvesting, and more. Compared to macroscopic channels, nanopores exhibit these unique functionalities due to the large surface-to-volume ratio which facilitates the surface properties on the walls to dictate ion transport. It is imperative to understand the principles which govern transport at nanoscale interfaces to help solid-state systems replicate the superior capabilities observed in their biological inspirations. The solid-liquid interface in nanopores systems is well-understood for aqueous solutions which can often be described by a continuum model. However, standard descriptions fall short when applied to non-aqueous solvents which exhibit long-range ordering. In the first part of this dissertation, we aim to provide a molecular-level understanding of how the spatial organization of solvents at an interface influences electrochemical potential and, in turn, governs electrokinetic phenomena.Many separation platforms and energy-storage systems use non-aqueous solvents and rely on interactions at the interface for determining electrochemical properties. Here we probe the non-aqueous solvent, propylene carbonate, in the presence of salts using electrochemical measurements which rely on nanopores as a model system. We found that this solvent organizes at polar interfaces with a bilayer-like structure which dictates the position of ions in a concentration-dependent manner. We also take into consideration the chiral character of the solvent and show that the enantiomeric excess plays a role in determining interfacial ordering. This work aims to provide a complete description of the organization of solvent molecules at polar surfaces and underscores the need to consider the role of chirality. The second part of this dissertation focuses on preparing solid-state nanopore systems that fluctuate in diameter to better mimic biological channels which are believed to benefit from the ability to gate transport by undergoing conformation changes in pore shape. Here we propose a solid-state nanopore rendered dynamic by modifying a gold electromechanical gate with DNA. We show that the response of the DNA to an external electric field effects the ion transport within the pore and effective opening diameter. This result provides the first steps towards preparing non-equilibrium nanopore systems and achieving the complex functionality of biological nanopores.