Large language models (LLMs) can leak sensitive training data through memorization and membership inference attacks. Prior work has primarily focused on strong adversarial assumptions, including attacker access to entire samples or long, ordered prefixes, leaving open the question of how vulnerable LLMs are when adversaries have only partial, unordered sample information. For example, if an attacker knows a patient has "hypertension," under what conditions can they query a model fine-tuned on patient data to learn the patient also has "osteoarthritis?" In this paper, we introduce a more general threat model under this weaker assumption and show that fine-tuned LLMs are susceptible to these fragment-specific extraction attacks. To systematically investigate these attacks, we propose two data-blind methods: (1) a likelihood ratio attack inspired by methods from membership inference, and (2) a novel approach, PRISM, which regularizes the ratio by leveraging an external prior. Using examples from medical and legal settings, we show that both methods are competitive with a data-aware baseline classifier that assumes access to labeled in-distribution data, underscoring their robustness.

Spurious correlations occur when models rely on non-essential features that coincidentally co-vary with target labels, leading to incorrect reasoning under distribution shift. We consider spurious correlations in multi-modal Large Vision Language Models (LVLMs) pretrained on extensive and diverse datasets without explicit task supervision. We develop a benchmark by sourcing GPT-4o errors on real-world visual-question-answering (VQA) benchmarks, then curating a subset through LVLM-human annotation and synthetic counterfactual evaluation to identify errors caused by spurious correlations. This process yields SpuriVerse, a novel benchmark comprised of 124 distinct types of spurious correlations extracted from real-world datasets, each containing 1 realistic and 10 synthetic VQA samples for a total of 1364 multiple choice questions. We evaluate 15 open and closed-source LVLMs on SpuriVerse, finding that even state-of-the-art closed-source models struggle significantly, achieving at best only 35.0\% accuracy. Fine-tuning on synthetic examples that emphasize the spurious correlation improves performance to 78.4\%, suggesting that training on diverse spurious patterns generalizes to unseen situations: models appear to learn to avoid "shortcuts" and attend to the overall image context.

Abstract Abstention, the refusal of large language models (LLMs) to provide an answer, is increasingly recognized for its potential to mitigate hallucinations and enhance safety in LLM systems. In this survey, we introduce a framework to examine abstention from three perspectives: the query, the model, and human values. We organize the literature on abstention methods, benchmarks, and evaluation metrics using this framework, and discuss merits and limitations of prior work. We further identify and motivate areas for future research, such as whether abstention can be achieved as a meta-capability that transcends specific tasks or domains, and opportunities to optimize abstention abilities in specific contexts. In doing so, we aim to broaden the scope and impact of abstention methodologies in AI systems.1

Popular and news media often portray teenagers with sensationalism, as both a risk to society and at risk from society. As AI begins to absorb some of the epistemic functions of traditional media, we study how teenagers in two countries speaking two languages: 1) are depicted by AI, and 2) how they would prefer to be depicted. Specifically, we study the biases about teenagers learned by static word embeddings (SWEs) and generative language models (GLMs), comparing these with the perspectives of adolescents living in the U.S. and Nepal. We find English-language SWEs associate teenagers with societal problems, and more than 50% of the 1,000 words most associated with teenagers in the pretrained GloVe SWE reflect such problems. Given prompts about teenagers, 30% of outputs from GPT2-XL and 29% from LLaMA-2-7B GLMs discuss societal problems, most commonly violence, but also drug use, mental illness, and sexual taboo. Nepali models, while not free of such associations, are less dominated by social problems. Data from workshops with N=13 U.S. adolescents and N=18 Nepalese adolescents show that AI presentations are disconnected from teenage life, which revolves around activities like school and friendship. Participant ratings of how well 20 trait words describe teens are decorrelated from SWE associations, with Pearson's rho=.02, n.s. in English FastText and rho=.06, n.s. GloVe; and rho=.06, n.s. in Nepali FastText and rho=-.23, n.s. in GloVe. U.S. participants suggested AI could fairly present teens by highlighting diversity, while Nepalese participants centered positivity. Participants were optimistic that, if it learned from adolescents, rather than media sources, AI could help mitigate stereotypes. Our work offers an understanding of the ways SWEs and GLMs misrepresent a developmentally vulnerable group and provides a template for less sensationalized characterization.

Multimodal AI models capable of associating images and text hold promise for numerous domains, ranging from automated image captioning to accessibility applications for blind and low-vision users. However, uncertainty about bias has in some cases limited their adoption and availability. In the present work, we study 43 CLIP vision-language models to determine whether they learn human-like facial impression biases, and we find evidence that such biases are reflected across three distinct CLIP model families. We show for the first time that the the degree to which a bias is shared across a society predicts the degree to which it is reflected in a CLIP model. Human-like impressions of visually unobservable attributes, like trustworthiness and sexuality, emerge only in models trained on the largest dataset, indicating that a better fit to uncurated cultural data results in the reproduction of increasingly subtle social biases. Moreover, we use a hierarchical clustering approach to show that dataset size predicts the extent to which the underlying structure of facial impression bias resembles that of facial impression bias in humans. Finally, we show that Stable Diffusion models employing CLIP as a text encoder learn facial impression biases, and that these biases intersect with racial biases in Stable Diffusion XL-Turbo. While pretrained CLIP models may prove useful for scientific studies of bias, they will also require significant dataset curation when intended for use as general-purpose models in a zero-shot setting.

This research introduces the Multilevel Embedding Association Test (ML-EAT), a method designed for interpretable and transparent measurement of intrinsic bias in language technologies. The ML-EAT addresses issues of ambiguity and difficulty in interpreting the traditional EAT measurement by quantifying bias at three levels of increasing granularity: the differential association between two target concepts with two attribute concepts; the individual effect size of each target concept with two attribute concepts; and the association between each individual target concept and each individual attribute concept. Using the ML-EAT, this research defines a taxonomy of EAT patterns describing the nine possible outcomes of an embedding association test, each of which is associated with a unique EAT-Map, a novel four-quadrant visualization for interpreting the ML-EAT. Empirical analysis of static and diachronic word embeddings, GPT-2 language models, and a CLIP language-and-image model shows that EAT patterns add otherwise unobservable information about the component biases that make up an EAT; reveal the effects of prompting in zero-shot models; and can also identify situations when cosine similarity is an ineffective metric, rendering an EAT unreliable. Our work contributes a method for rendering bias more observable and interpretable, improving the transparency of computational investigations into human minds and societies.

Deep neural networks are prone to capture correlations between spurious attributes and class labels, leading to low accuracy on some combinations of class labels and spurious attribute values. When a spurious attribute represents a protected class, these low-accuracy groups can manifest discriminatory bias. Existing methods attempting to improve worst-group accuracy assume the training data, validation data, or both are reliably labeled by the spurious attribute. But a model may be perceived to be biased towards a concept that is not represented by pre-existing labels on the training data. In these situations, the spurious attribute must be defined with external information. We propose Concept Correction, a framework that represents a concept as a curated set of images from any source, then labels each training sample by its similarity to the concept set to control spurious correlations. For example, concept sets representing gender can be used to measure and control gender bias even without explicit labels. We demonstrate and evaluate an instance of the framework as Concept DRO, which uses concept sets to estimate group labels, then uses these labels to train with a state of the art distributively robust optimization objective. We show that Concept DRO outperforms existing methods that do not require labels of spurious attributes by up to 33.1 % on three image classification datasets and is competitive with the best methods that assume access to labels. We consider how the size and quality of the concept set influences performance and find that even smaller, manually curated sets of noisy AI-generated images are effective at controlling spurious correlations, suggesting that high-quality, reusable concept sets are easy to create and effective in reducing bias.

Equitable urban transportation applications require high-fidelity digital representations of the built environment (streets, crossings, curb ramps and more). Direct inspections and manual annotations are costly at scale, while conventional machine learning methods require substantial annotated training data for adequate performance. This study explores vision language models as a tool for annotating diverse urban features from satellite images, reducing the dependence on human annotation. Although these models excel at describing common objects in human-centric images, their training sets may lack signals for esoteric built environment features, making their performance uncertain. We demonstrate a proof-of-concept using a vision language model and a visual prompting strategy that considers segmented image elements. Experiments on two urban features --- stop lines and raised tables --- show that while zero-shot prompting rarely works, the segmentation and visual prompting strategies achieve nearly 40% intersection-over-union accuracy. We describe how these results motivate further research in automatic annotation of the built environment to improve equity, accessibility, and safety at scale and in diverse environments.

Open data is frequently released spatially aggregated, usually to comply with privacy policies. However, coarse heterogeneous aggregations complicate learning and integration for downstream AI/ML systems. In this work, we consider models to disaggregate spatio-temporal data from a low-resolution, irregular partition (e.g., census tract) to a high-resolution, irregular partition (e.g., city block). We propose an overarching model named the Structurally-Aware Recurrent Network (SARN), which integrates structurally-aware spatial attention (SASA) layers into the Gated Recurrent Unit (GRU) model. The spatial attention layers capture spatial interactions among regions, while the gated recurrent module captures the temporal dependencies. Each SASA layer calculates both global and structural attention --- global attention facilitates comprehensive interactions between different geographic levels, while structural attention leverages the containment relationship between different geographic levels (e.g., a city block being wholly contained within a census tract) to ensure coherent and consistent results. For scenarios with limited historical training data, we explore transfer learning and show that a model pre-trained on one city variable can be fine-tuned for another city variable using only a few hundred samples. Evaluating these techniques on two mobility datasets, we find that on both datasets, SARN significantly outperforms other neural models (5.0% and 1.2%) and typical heuristic methods (40% and 14%), enabling us to generate realistic, high-quality fine-grained data for downstream applications.

Differential privacy (DP) data synthesizers are increasingly proposed to afford public release of sensitive information, offering theoretical guarantees for privacy (and, in some cases, utility), but limited empirical evidence of utility in practical settings. Utility is typically measured as the error on representative proxy tasks, such as descriptive statistics, multivariate correlations, the accuracy of trained classifiers, or performance over a query workload. The ability for these results to generalize to practitioners' experience has been questioned in a number of settings, including the U.S. Census. In this paper, we propose an evaluation methodology for synthetic data that avoids assumptions about the representativeness of proxy tasks, instead measuring the likelihood that published conclusions would change had the authors used synthetic data, a condition we call epistemic parity. Our methodology consists of reproducing empirical conclusions of peer-reviewed papers on real, publicly available data, then re-running these experiments a second time on DP synthetic data and comparing the results. We instantiate our methodology over a benchmark of recent peer-reviewed papers that analyze public datasets in the ICPSR social science repository. We model quantitative claims computationally to automate the experimental workflow, and model qualitative claims by reproducing visualizations and comparing the results manually. We then generate DP synthetic datasets using multiple state-of-the-art mechanisms, and estimate the likelihood that these conclusions will hold. We find that, for reasonable privacy regimes, state-of-the-art DP synthesizers are able to achieve high epistemic parity for several papers in our benchmark. However, some papers, and particularly some specific findings, are difficult to reproduce for any of the synthesizers. Given these results, we advocate for a new class of mechanisms that can reorder the priorities for DP data synthesis: favor stronger guarantees for utility (as measured by epistemic parity) and offer privacy protection with a focus on application-specific threat models and risk-assessment.

We describe a customizable benchmark for hierarchical and ontological multi-label classification, a task where labels are equipped with a graph structure and data items can be assigned multiple labels. We find that current benchmarks do not adequately represent the problem space, casting doubt on the generalizability of current results. We consider three dimensions of the problem space: context (availability of rich features on the data and labels), distribution of labels over data, and graph structure. For context, the lack of complex features on the labels (and in some cases, the data) artificially prevent the use of modern representation learning techniques as an appropriate baseline. For distribution, we find the long tail of labels over data constitute a few-shot learning problem that artificially confounds the results: for most common benchmarks, over 40% of the labels have fewer than 5 data points in the training set. For structure, we find that the correlation between performance and the height of the tree can explain some of the variation in performance, informing practical utility. In this paper, we demonstrate how the lack of diversity in benchmarks can confound performance analysis, then present a declarative query system called Ontologue for generating custom benchmarks with specific properties, then use this system to design 4 new benchmarks extracted from DBPedia that better represent the problem space. We evaluate state-of-the-art algorithms on both existing and new benchmarks and show that the performance conclusions can vary significantly depending on the dimensions we consider. We intend the system and derived benchmarks to improve the analysis of generalizability for these problems.

Neural methods are state-of-the-art for urban prediction problems such as transportation resource demand, accident risk, crowd mobility, and public safety. Model performance can be improved by integrating exogenous features from open data repositories (e.g., weather, housing prices, traffic, etc.), but these uncurated sources are often too noisy, incomplete, and biased to use directly. We propose to learn integrated representations, called EquiTensors, from heterogeneous datasets that can be reused across a variety of tasks. We align datasets to a consistent spatio-temporal domain, then describe an unsupervised model based on convolutional denoising autoencoders to learn shared representations. We extend this core integrative model with adaptive weighting to prevent certain datasets from dominating the signal. To combat discriminatory bias, we use adversarial learning to remove correlations with a sensitive attribute (e.g., race or income). Experiments with 23 input datasets and 4 real applications show that EquiTensors could help mitigate the effects of the sensitive information embodied in the biased data. Meanwhile, applications using EquiTensors outperform models that ignore exogenous features and are competitive with "oracle" models that use hand-selected datasets.

Emerging transportation modes, including car-sharing, bike-sharing, and ride-hailing, are transforming urban mobility yet have been shown to reinforce socioeconomic inequity. These services rely on accurate demand prediction, but the demand data on which these models are trained reflect biases around demographics, socioeconomic conditions, and entrenched geographic patterns. To address these biases and improve fairness, we present FairST, a fairness-aware demand prediction model for spatiotemporal urban applications, with emphasis on new mobility. We use 1D (time-varying, space-constant), 2D (space-varying, time-constant) and 3D (both time- and space-varying) convolutional branches to integrate heterogeneous features, while including fairness metrics as a form of regularization to improve equity across demographic groups. We propose two spatiotemporal fairness metrics, region-based fairness gap (RFG), applicable when demographic information is provided as a constant for a region, and individual-based fairness gap (IFG), applicable when a continuous distribution of demographic information is available. Experimental results on bike share and ride share datasets show that FairST can reduce inequity in demand prediction for multiple sensitive attributes (i.e. race, age, and education level), while achieving better accuracy than even state-of-the-art fairness-oblivious methods.

We propose JECL, a method for clustering image-caption pairs by training parallel encoders with regularized clustering and alignment objectives, simultaneously learning both representations and cluster assignments. These image-caption pairs arise frequently in high-value applications where structured training data is expensive to produce, but free-text descriptions are common. JECL trains by minimizing the Kullback-Leibler divergence between the distribution of the images and text to that of a combined joint target distribution and optimizing the Jensen-Shannon divergence between the soft cluster assignments of the images and text. Regularizers are also applied to JECL to prevent trivial solutions. Experiments show that JECL outperforms both single-view and multi-view methods on large benchmark image-caption datasets, and is remarkably robust to missing captions and varying data sizes.

The need for responsible data management intensifies with the growing impact of data on society. One central locus of the societal impact of data are Automated Decision Systems (ADS), socio-legal-technical systems that are used broadly in industry, non-profits, and government. ADS process data about people, help make decisions that are consequential to people's lives, are designed with the stated goals of improving efficiency and promoting equitable access to opportunity, involve a combination of human and automated decision making, and are subject to auditing for legal compliance and to public disclosure. They may or may not use AI, and may or may not operate with a high degree of autonomy, but they rely heavily on data. In this article, we argue that the data management community is uniquely positioned to lead the responsible design, development, use, and oversight of ADS. We outline a technical research agenda that requires that we step outside our comfort zone of engineering for efficiency and accuracy, to also incorporate reasoning about values and beliefs. This seems high-risk, but one of the upsides is being able to explain to our children what we do and why it matters.

We contribute user-centered prefetching and indexing methods that provide low-latency interactions across linked visualizations, enabling cold-start exploration of billion-record datasets. We implement our methods in Falcon, a web-based system that makes principled trade-offs between latency and resolution to optimize brushing and view switching times. To optimize latency-sensitive brushing actions, Falcon reindexes data upon changes to the active view a user is brushing in. To limit view switching times, Falcon initially loads reduced interactive resolutions, then progressively improves them. Benchmarks show that Falcon sustains real-time interactivity of 50fps for pixel-level brushing and linking across multiple visualizations with no costly precomputation. We show constant brushing performance regardless of data size on datasets ranging from millions of records in the browser to billions when connected to a backing database system.

Using inappropriate datasets for data science tasks can be harmful, especially for applications that impact humans. Targeting data ethics, we demonstrate MithraLabel, a system for generating task-specific information about a dataset, in the form of a set of visual widgets, as a flexible "nutritional label" that provides a user with information to determine the fitness of the dataset for the task at hand.

We present a fairness-aware model for predicting demand for new mobility systems. Our approach, called FairST, consists of 1D, 2D and 3D convolutions to learn the spatial-temporal dynamics of a mobility system, and fairness regularizers that guide the model to make equitable predictions. We propose two fairness metrics, region-based fairness gap (RFG) and individual-based fairness gap (IFG), that measure equity gaps between social groups for new mobility systems. Experimental results on two real-world datasets demonstrate the effectiveness of the proposed model: FairST not only reduces the fairness gap by more than 80%, but achieves better accuracy than state-of-the-art but fairness-oblivious methods including LSTMs, ConvLSTMs, and 3D CNN.

Publishers are increasingly using graphical abstracts to facilitate scientific search, especially across disciplinary boundaries. They are presented on various media, easily shared and information rich. However, very small amount of scientific publications are equipped with graphical abstracts. What can we do with the vast majority of papers with no selected graphical abstract? In this paper, we first hypothesize that scientific papers actually include a "central figure" that serve as a graphical abstract. These figures convey the key results and provide a visual identity for the paper. Using survey data collected from 6,263 authors regarding 8,353 papers over 15 years, we find that over 87% of papers are considered to contain a central figure, and that these central figures are primarily used to summarize important results, explain the key methods, or provide additional discussion. We then train a model to automatically recognize the central figure, achieving top-3 accuracy of 78% and exact match accuracy of 34%. We find that the primary boost in accuracy comes from figure captions that resemble the abstract. We make all our data and results publicly available at https://github.com/viziometrics/centraul_figure. Our goal is to automate central figure identification to improve search engine performance and to help scientists connect ideas across the literature.

Fairness is increasingly recognized as a critical component of machine learning systems. However, it is the underlying data on which these systems are trained that often reflect discrimination, suggesting a database repair problem. Existing treatments of fairness rely on statistical correlations that can be fooled by statistical anomalies, such as Simpson's paradox. Proposals for causality-based definitions of fairness can correctly model some of these situations, but they require specification of the underlying causal models. In this paper, we formalize the situation as a database repair problem, proving sufficient conditions for fair classifiers in terms of admissible variables as opposed to a complete causal model. We show that these conditions correctly capture subtle fairness violations. We then use these conditions as the basis for database repair algorithms that provide provable fairness guarantees about classifiers trained on their training labels. We evaluate our algorithms on real data, demonstrating improvement over the state of the art on multiple fairness metrics proposed in the literature while retaining high utility.

We present GraviTIE (Global Representation and Visualization of Text and Image Embeddings, pronounced ”gravity”), an interactive visualization system for large-scale image datasets. GraviTIE operates on datasets consisting of images equipped with unstructured and semi-structured text, relying on multi-modal unsupervised learning methods to produce an interactive similarity map. Users interact with the similarity map through pan and zoom operations, as well as keyword-oriented queries. GraviTIE makes no assumptions about the form, scale, or content of the data, allowing it to be used for exploratory analysis, assessment of unsupervised learning methods, data curation and quality control, data profiling, and other purposes where flexibility and scalability are paramount. We demonstrate GraviTIE on three real datasets: 500k images from the Russian misinformation dataset from Twitter, 2 million art images, and 5 million scientific figures. A screencast video is available at https://vimeo.com/310511187.

There exists a gap between visualization design guidelines and their application in visualization tools. While empirical studies can provide design guidance, we lack a formal framework for representing design knowledge, integrating results across studies, and applying this knowledge in automated design tools that promote effective encodings and facilitate visual exploration. We propose modeling visualization design knowledge as a collection of constraints, in conjunction with a method to learn weights for soft constraints from experimental data. Using constraints, we can take theoretical design knowledge and express it in a concrete, extensible, and testable form: the resulting models can recommend visualization designs and can easily be augmented with additional constraints or updated weights. We implement our approach in Draco, a constraint-based system based on Answer Set Programming (ASP). We demonstrate how to construct increasingly sophisticated automated visualization design systems, including systems based on weights learned directly from the results of graphical perception experiments.

Many real-world applications require automated data annotation, such as identifying tissue origins based on gene expressions and classifying images into semantic categories. Annotation classes are often numerous and subject to changes over time, and annotating examples has become the major bottleneck for supervised learning methods. In science and other high-value domains, large repositories of data samples are often available, together with two sources of organic supervision: a lexicon for the annotation classes, and text descriptions that accompany some data samples. Distant supervision has emerged as a promising paradigm for exploiting such indirect supervision by automatically annotating examples where the text description contains a class mention in the lexicon. However, due to linguistic variations and ambiguities, such training data is inherently noisy, which limits the accuracy in this approach. In this paper, we introduce an auxiliary natural language processing system for the text modality, and incorporate co-training to reduce noise and augment signal in distant supervision. Without using any manually labeled data, our EZLearn system learned to accurately annotate data samples in functional genomics and scientific figure comprehension, substantially outperforming state-of-the-art supervised methods trained on tens of thousands of annotated examples.

No abstract available.

Algorithmic decisions often result in scoring and ranking individuals to determine credit worthiness, qualifications for college admissions and employment, and compatibility as dating partners. While automatic and seemingly objective, ranking algorithms can discriminate against individuals and protected groups, and exhibit low diversity. Furthermore, ranked results are often unstable -- small changes in the input data or in the ranking methodology may lead to drastic changes in the output, making the result uninformative and easy to manipulate. Similar concerns apply in cases where items other than individuals are ranked, including colleges, academic departments, or products. Despite the ubiquity of rankers, there is, to the best of our knowledge, no technical work that focuses on making rankers transparent.

Recently, there has begun a movement towards Fairness, Accountability, and Transparency (FAT) in algorithmic decision making, and in data science more broadly. The database community has not been significantly involved in this movement, despite "owning" the models, languages, and systems that produce the (potentially biased) input to the machine learning applications. What role should the database community play in this movement? Do the objectives of fairness, accountability and transparency give rise to core data management issues that can drive new research questions and new systems, or are these "soft topics" that are best left to be managed with policy? Will emphasis on these topics dilute our core competency in techniques and technologies for data, or can it reinforce our central role in technology stacks ranging from startups to the enterprise, and from local non-profits to the federal government? The goal of this panel is to debate these questions, and to whet the appetite of the data management community for research in this important emerging area.

Scientific results are communicated visually in the literature through diagrams, visualizations, and photographs. These information-dense objects have been largely ignored in bibliometrics and scientometrics studies when compared to citations and text. In this paper, we use techniques from computer vision and machine learning to classify more than 8 million figures from PubMed into five figure types and study the resulting patterns of visual information as they relate to scholarly impact. We find that the distribution of figures and figure types in the literature has remained relatively constant over time, but can vary widely across field and topic. Remarkably, we find a significant correlation between scientific impact and the use of visual information, where higher impact papers tend to include more diagrams, and to a lesser extent more plots. To explore these results and other ways of extracting this visual information, we have built a visual browser to illustrate the concept and explore design alternatives for supporting viziometric analysis and organizing visual information. We use these results to articulate a new research agenda-viziometrics-to study the organization and presentation of visual information in the scientific literature.

Visual data analysis involves both open-ended and focused exploration. Manual chart specification tools support question answering, but are often tedious for early-stage exploration where systematic data coverage is needed. Visualization recommenders can encourage broad coverage, but irrelevant suggestions may distract users once they commit to specific questions. We present Voyager 2, a mixed-initiative system that blends manual and automated chart specification to help analysts engage in both open-ended exploration and targeted question answering. We contribute two partial specification interfaces: wildcards let users specify multiple charts in parallel, while related views suggest visualizations relevant to the currently specified chart. We present our interface design and applications of the CompassQL visualization query language to enable these interfaces. In a controlled study we find that Voyager 2 leads to increased data field coverage compared to a traditional specification tool, while still allowing analysts to flexibly drill-down and answer specific questions.

General Purpose computing on Graphics Processing Units (GPGPU) has become an increasingly popular option for accelerating database queries. However, GPUs are not well-suited for all types of queries as data transfer costs can often dominate query execution. We develop a methodology for quantifying how well databases utilize GPU architectures using proprietary profiling tools. By aggregating various profiling metrics, we break down the different aspects that comprise occupancy on the GPU across the runtime of query execution. We show that for the Alenka GPU database, only a small minority of execution time, roughly 5% is spent on the GPU. We further show that even on queries with seemingly good performance, a large portion of the achieved occupancy can actually be attributed to stalls and scalar instructions.

We consider a new approach to extracting information from dendrograms in the biological literature representing phylogenetic trees. Existing algorithmic approaches to extract these relationships rely on tracing tree contours and are very sensitive to image quality issues, but manual approaches require significant human effort and cannot be used at scale. We introduce PhyloParser, a fully automated, end-to-end system for automatically extracting species relationships from phylogenetic tree diagrams using a multi-modal approach to digest diverse tree styles. Our approach automatically identifies phylogenetic tree figures in the scientific literature, extracts the key components of tree structure, reconstructs the tree, and recovers the species relationships. We use multiple methods to extract tree components with high recall, then filter false positives by applying topological heuristics about how these components fit together. We present an evaluation on a real-world dataset to quantitatively and qualitatively demonstrate the efficacy of our approach. Our classifier achieves 89% recall and 99% precision, with a low average error rate relative to previous approaches. We aim to use PhyloParser to build a linked, open, comprehensive database of phylogenetic information that covers the historical literature as well as current data, and then use this resource to identify areas of disagreement and poor coverage in the biological literature.

In the first wave of data science education programs, data engineering topics (systems, scalable algorithms, data management, integration) tended to be de-emphasized in favor of machine learning and statistical modeling. The anecdotal evidence suggests this was a mistake: data scientists report spending most of their time grappling with data far upstream of modeling activities. A second wave of data science education is emerging, one with increased emphasis on practical issues in ethics, legal compliance, scientific reproducibility, data quality, and algorithmic bias. The data engineering community has a second chance to influence these programs beyond just providing a set of tools. In this panel, we'll discuss the role of data engineering in data science education programs, and how best to capitalize on emerging opportunities in this space.

To facilitate collaboration over sensitive data, we present DataSynthesizer, a tool that takes a sensitive dataset as input and generates a structurally and statistically similar synthetic dataset with strong privacy guarantees. The data owners need not release their data, while potential collaborators can begin developing models and methods with some confidence that their results will work similarly on the real dataset. The distinguishing feature of DataSynthesizer is its usability --- the data owner does not have to specify any parameters to start generating and sharing data safely and effectively.

Issues of responsible data analysis and use are coming to the forefront of the discourse in data science research and practice, with most significant efforts to date on the part of the data mining, machine learning, and security and privacy communities. In these fields, the research has been focused on analyzing the fairness, accountability and transparency (FAT) properties of specific algorithms and their outputs. Although these issues are most apparent in the social sciences where fairness is interpreted in terms of the distribution of resources across protected groups, management of bias in source data affects a variety of fields. Consider climate change studies that require representative data from geographically diverse regions, or supply chain analyses that require data that represents the diversity of products and customers. Any domain that involves sparse or sampled data has exposure to potential bias.

Community detection is an increasingly popular approach to uncover important structures in large networks. Flow-based community detection methods rely on communication patterns of the network rather than structural properties to determine communities. The Infomap algorithm in particular optimizes a novel objective function called the map equation and has been shown to outperform other approaches in third-party benchmarks. However, Infomap and its variants are inherently sequential, limiting their use for large-scale graphs. In this article, we propose a novel algorithm to optimize the map equation called RelaxMap. RelaxMap provides two important improvements over Infomap: parallelization, so that the map equation can be optimized over much larger graphs, and prioritization, so that the most important work occurs first, iterations take less time, and the algorithm converges faster. We implement these techniques using OpenMP on shared-memory multicore systems, and evaluate our approach on a variety of graphs from standard graph clustering benchmarks as well as real graph datasets. Our evaluation shows that both techniques are effective: RelaxMap achieves 70% parallel efficiency on eight cores, and prioritization improves algorithm performance by an additional 20--50% on average, depending on the graph properties. Additionally, RelaxMap converges in the similar number of iterations and provides solutions of equivalent quality as the serial Infomap implementation.

We analyze the workload from a multi-year deployment of a database-as-a-service platform targeting scientists and data scientists with minimal database experience. Our hypothesis was that relatively minor changes to the way databases are delivered can increase their use in ad hoc analysis environments. The web-based SQLShare system emphasizes easy dataset-at-a-time ingest, relaxed schemas and schema inference, easy view creation and sharing, and full SQL support. We find that these features have helped attract workloads typically associated with scripts and files rather than relational databases: complex analytics, routine processing pipelines, data publishing, and collaborative analysis. Quantitatively, these workloads are characterized by shorter dataset "lifetimes", higher query complexity, and higher data complexity. We report on usage scenarios that suggest SQL is being used in place of scripts for one-off data analysis and ad hoc data sharing. The workload suggests that a new class of relational systems emphasizing short-term, ad hoc analytics over engineered schemas may improve uptake of database technology in data science contexts. Our contributions include a system design for delivering databases into these contexts, a description of a public research query workload dataset released to advance research in analytic data systems, and an initial analysis of the workload that provides evidence of new use cases under-supported in existing systems.

MOTIVATION: Recent technological innovations in flow cytometry now allow oceanographers to collect high-frequency flow cytometry data from particles in aquatic environments on a scale far surpassing conventional flow cytometers. The SeaFlow cytometer continuously profiles microbial phytoplankton populations across thousands of kilometers of the surface ocean. The data streams produced by instruments such as SeaFlow challenge the traditional sample-by-sample approach in cytometric analysis and highlight the need for scalable clustering algorithms to extract population information from these large-scale, high-frequency flow cytometers. RESULTS: We explore how available algorithms commonly used for medical applications perform at classification of such a large-scale, environmental flow cytometry data. We apply large-scale Gaussian mixture models to massive datasets using Hadoop. This approach outperforms current state-of-the-art cytometry classification algorithms in accuracy and can be coupled with manual or automatic partitioning of data into homogeneous sections for further classification gains. We propose the Gaussian mixture model with partitioning approach for classification of large-scale, high-frequency flow cytometry data. AVAILABILITY AND IMPLEMENTATION: Source code available for download at https://github.com/jhyrkas/seaflow_cluster, implemented in Java for use with Hadoop. CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

General visualization tools typically require manual specification of views: analysts must select data variables and then choose which transformations and visual encodings to apply. These decisions often involve both domain and visualization design expertise, and may impose a tedious specification process that impedes exploration. In this paper, we seek to complement manual chart construction with interactive navigation of a gallery of automatically-generated visualizations. We contribute Voyager, a mixed-initiative system that supports faceted browsing of recommended charts chosen according to statistical and perceptual measures. We describe Voyager's architecture, motivating design principles, and methods for generating and interacting with visualization recommendations. In a study comparing Voyager to a manual visualization specification tool, we find that Voyager facilitates exploration of previously unseen data and leads to increased data variable coverage. We then distill design implications for visualization tools, in particular the need to balance rapid exploration and targeted question-answering.

Flow cytometers measure the optical properties of particles to classify microbes. Recent innovations have allowed oceanographers to collect flow cytometry data continuously during research cruises, leading to an explosion of data and new challenges for the classification task. The massive scale, time-varying underlying populations, and noisy measurements motivate the development of new classification methods. We describe the problem, the data, and some preliminary results demonstrating the difficulty with conventional methods.

In this paper, we describe a new distributed community detection algorithm for billion-edge directed graphs that, unlike modularity-based methods, achieves cluster quality on par with the best-known algorithms in the literature. We show that a simple approximation to the best-known serial algorithm dramatically reduces computation and enables distributed evaluation yet incurs only a very small impact on cluster quality.

Before scientists can analyze, publish, or share their data, they often need to determine how their datasets are related. Determining relationships helps scientists identify the most complete version of a dataset, detect versions of datasets that complement each other, and determine multiple datasets that overlap. In previous work, we showed how observable relationships between two datasets help scientists recall their original derivation connection. While that work helped with identifying relationships between two datasets, it is infeasible for scientists to use it for finding relationships between all possible pairs in a large collection of datasets. In order to deal with larger numbers of datasets, we are extending our methodology with a relationship-prediction system, ReDiscover, a tool to identify pairs from a collection of datasets that are most likely related and the relationship between them. We report on the initial design of ReDiscover, which uses machine-learning methods such as Conditional Random Fields and Support Vector Machines to the relationship-discovery problem. Our preliminarily evaluation shows that ReDiscover predicted relationships with an average accuracy of 87%.

Abstract Distributed database performance is often unpredictable due to issues such as system complexity, network congestion, or imbalanced data distribution. These issues are difficult for users to assess in part due to the opaque mapping between declaratively specified queries and actual physical execution plans. Database developers currently must expend significant time and effort scanning log files to isolate and debug the root causes of performance issues. In response, we present Perfopticon, an interactive query profiling tool that enables rapid insight into common problems such as performance bottlenecks and data skew. Perfopticon combines interactive visualizations of (1) query plans, (2) overall query execution, (3) data flow among servers, and (4) execution traces. These views coordinate multiple levels of abstraction to enable detection, isolation, and understanding of performance issues. We evaluate our design choices through engagements with system developers, scientists, and students. We demonstrate that Perfopticon enables performance debugging for real‐world tasks.

Data is increasingly being bought and sold online, and Web-based marketplace services have emerged to facilitate these activities. However, current mechanisms for pricing data are very simple: buyers can choose only from a set of explicit views, each with a specific price. In this article, we propose a framework for pricing data on the Internet that, given the price of a few views, allows the price of any query to be derived automatically. We call this capability query-based pricing . We first identify two important properties that the pricing function must satisfy, the arbitrage-free and discount-free properties. Then, we prove that there exists a unique function that satisfies these properties and extends the seller's explicit prices to all queries. Central to our framework is the notion of query determinacy, and in particular instance-based determinacy : we present several results regarding the complexity and properties of it. When both the views and the query are unions of conjunctive queries or conjunctive queries, we show that the complexity of computing the price is high. To ensure tractability, we restrict the explicit prices to be defined only on selection views (which is the common practice today). We give algorithms with polynomial time data complexity for computing the price of two classes of queries: chain queries (by reducing the problem to network flow), and cyclic queries. Furthermore, we completely characterize the class of conjunctive queries without self-joins that have PTIME data complexity, and prove that pricing all other queries is NP-complete, thus establishing a dichotomy on the complexity of the pricing problem when all views are selection queries.

This paper presents BigDAWG, a reference implementation of a new architecture for "Big Data" applications. Such applications not only call for large-scale analytics, but also for real-time streaming support, smaller analytics at interactive speeds, data visualization, and cross-storage-system queries. Guided by the principle that "one size does not fit all", we build on top of a variety of storage engines, each designed for a specialized use case. To illustrate the promise of this approach, we demonstrate its effectiveness on a hospital application using data from an intensive care unit (ICU). This complex application serves the needs of doctors and researchers and provides real-time support for streams of patient data. It showcases novel approaches for querying across multiple storage engines, data visualization, and scalable real-time analytics.

In this demonstration, we will showcase Myria, our novel cloud service for big data management and analytics designed to improve productivity. Myria's goal is for users to simply upload their data and for the system to help them be self-sufficient data science experts on their data -- self-serve analytics. Using a web browser, Myria users can upload data, author efficient queries to process and explore the data, and debug correctness and performance issues. Myria queries are executed on a scalable, parallel cluster that uses both state-of-the-art and novel methods for distributed query processing. Our interactive demonstration will guide visitors through an exploration of several key Myria features by interfacing with the live system to analyze big datasets over the web.

The Database Community has a unique perspective on the challenges and solutions of long-term management of data and the value of data as a resource. In current computer science curricula, however, these insights are typically locked up in the context of the traditional Intro to Databases class that was developed years (or in some cases, decades) before the modern concept of Data Science arose and embedded in the discussion of legacy data management systems. We consider how to bring these concepts front and center into the emerging wave of Data Science courses, degree programs and even departments.

It seems inevitable that the datasets associated with a research project proliferate over time: collaborators may extend datasets with new measurements and new attributes, new experimental runs result in new files with similar structures, and subsets of data are extracted for independent analysis. As these "residual" datasets begin to accrete over time, scientists can lose track of the derivation history that connects them, complicating data sharing, provenance tracking, and scientific reproducibility. In this paper, focusing on data in spreadsheets, we consider how observable relationships between two datasets can help scientists recall their original derivation connection. For instance, if dataset A is wholly contained in dataset B, B may be a more recent version of A and should be preferred when archiving or publishing.

We develop a new pricing system, QueryMarket, for flexible query pricing in a data market based on an earlier theoretical framework (Koutris et al., PODS 2012). To build such a system, we show how to use an Integer Linear Programming formulation of the pricing problem for a large class of queries, even when pricing is computationally hard. Further, we leverage query history to avoid double charging when queries purchased over time have overlapping information, or when the database is updated. We then present a technique that fairly shares revenue when multiple sellers are involved. Finally, we implement our approach in a prototype and evaluate its performance on several query workloads.

In this demonstration, we show-case a database management system extended with a new type of component that we call a Data Use Manager (DUM). The DUM enables DBAs to attach policies to data loaded into the DBMS. It then monitors how users query the data, flags potential policy violations, recommends possible fixes, and supports offline analysis of user activities related to data policies. The demonstration uses real healthcare data.

MOTIVATION: As industry and science are increasingly data-driven, the need for skilled data scientists is exceeding what our universities are producing. According to a Mckinsey report: "By 2018, the United States alone could face a shortage of 140,000 to 190,000 people with deep analytical skills". Similarly, the ability to extract knowledge from scientific data is accelerating discovery and we need the next generation of domain scientists to be experts not only in their domain but also in data management. At the same time, however, researchers in academia who focus on building instruments or data management tools are often less recognized for their contributions than researchers focusing purely on the actual science.

We consider a case study using SQL-as-a-Service to support "instant analysis" of weakly structured relational data at a multi-investigator science retreat. Here, "weakly structured" means tabular, rows-and-columns datasets that share some common context, but that have limited a priori agreement on file formats, relationships, types, schemas, metadata, or semantics. In this case study, the data were acquired from hundreds of distinct locations during a multi-day oceanographic cruise using a variety of physical, biological, and chemical sensors and assays. Months after the cruise when preliminary data processing was complete, 40+ researchers from a variety of disciplines participated in a two-day "data synthesis workshop." At this workshop, two computer scientists used a web-based query-as-a-service platform called SQLShare to perform "SQL stenography": capturing the scientific discussion in real time to integrate data, test hypotheses, and populate visualizations to then inform and enhance further discussion. In this "field test" of our technology and approach, we found that it was not only feasible to support interactive science Q&A with essentially pure SQL, but that we significantly increased the value of the "face time" at the meeting: researchers from different fields were able to validate assumptions and resolve ambiguity about each others' fields. As a result, new science emerged from a meeting that was originally just a planning meeting. In this paper, we describe the details of this experiment, discuss our major findings, and lay out a new research agenda for collaborative science database services.

We analyze Hadoop workloads from three di?erent research clusters from a user-centric perspective. The goal is to better understand data scientists' use of the system and how well the use of the system matches its design. Our analysis suggests that Hadoop usage is still in its adolescence. We see underuse of Hadoop features, extensions, and tools. We see significant diversity in resource usage and application styles, including some interactive and iterative workloads, motivating new tools in the ecosystem. We also observe significant opportunities for optimizations of these workloads. We find that job customization and configuration are used in a narrow scope, suggesting the future pursuit of automatic tuning systems. Overall, we present the first user-centered measurement study of Hadoop and find significant opportunities for improving its efficient use for data scientists.

As MapReduce/Hadoop grows in importance, we find more exotic applications being written this way. Not every program written for this platform performs as well as we might wish. There are several reasons why a MapReduce program can underperform expectations. One is the need to balance the communication cost of transporting data from the mappers to the reducers against the computation done at the mappers and reducers themselves. A second important issue is selecting the number of rounds of MapReduce. A third issue is that of skew. If wall-clock time is important, then using many different reduce-keys and many compute nodes may minimize the time to finish the job. Yet if the data is uncooperative, and no provision is made to distribute the data evenly, much of the work is done by a single node.

Data is increasingly being bought and sold online, and Web-based marketplace services have emerged to facilitate these activities. However, current mechanisms for pricing data are very simple: buyers can choose only from a set of explicit views, each with a specific price. In this paper, we propose a framework for pricing data on the Internet that, given the price of a few views, allows the price of any query to be derived automatically. We call this capability "query-based pricing." We first identify two important properties that the pricing function must satisfy, called arbitrage-free and discount-free. Then, we prove that there exists a unique function that satisfies these properties and extends the seller's explicit prices to all queries. When both the views and the query are Unions of Conjunctive Queries, the complexity of computing the price is high. To ensure tractability, we restrict the explicit prices to be defined only on selection views (which is the common practice today). We give an algorithm with polynomial time data complexity for computing the price of any chain query by reducing the problem to network flow. Furthermore, we completely characterize the class of Conjunctive Queries without self-joins that have PTIME data complexity (this class is slightly larger than chain queries), and prove that pricing all other queries is NP-complete, thus establishing a dichotomy on the complexity of the pricing problem when all views are selection queries.

We present VizDeck, a web-based tool for exploratory visual analytics of unorganized relational data. Motivated by collaborations with domain scientists who search for complex patterns in hundreds of data sources simultaneously, VizDeck automatically recommends appropriate visualizations based on the statistical properties of the data and adopts a card game metaphor to help organize the recommended visualizations into interactive visual dashboard applications in seconds with zero programming. The demonstration allows users to derive, share, and permanently store their own dashboard from hundreds of real science datasets using a production system deployed at the University of Washington.

We present an automatic skew mitigation approach for user-defined MapReduce programs and present SkewTune, a system that implements this approach as a drop-in replacement for an existing MapReduce implementation. There are three key challenges: (a) require no extra input from the user yet work for all MapReduce applications, (b) be completely transparent, and (c) impose minimal overhead if there is no skew. The SkewTune approach addresses these challenges and works as follows: When a node in the cluster becomes idle, SkewTune identifies the task with the greatest expected remaining processing time. The unprocessed input data of this straggling task is then proactively repartitioned in a way that fully utilizes the nodes in the cluster and preserves the ordering of the input data so that the original output can be reconstructed by concatenation. We implement SkewTune as an extension to Hadoop and evaluate its effectiveness using several real applications. The results show that SkewTune can significantly reduce job runtime in the presence of skew and adds little to no overhead in the absence of skew.

Increasingly data is being bought and sold online. To facilitate such transactions, online data market-places have emerged to provide a service for sellers to price views on their data, and buyers to buy such views. These marketplaces neither support the sale of ad-hoc queries (that are not one of the specified views), nor do they support queries that join datasets. We present QueryMarket, a prototype data marketplace that automatically extrapolates prices to ad-hoc queries, including those with joins, from the manually priced views. We call this capability "query-based pricing" and describe how it is superior to existing pricing methods, and how it provides more flexible pricing for the sellers. We then show how QueryMarket implements query-based pricing and how it generates explanations for the prices it computes.

We demonstrate SkewTune, a system that automatically mitigates skew in user-defined MapReduce programs and is a drop-in replacement for Hadoop. The demonstration has two parts. First, we demonstrate how SkewTune mitigates skew in real MapReduce applications at runtime by running a real application in a public cloud. Second, through an interactive graphical interface, we demonstrate the details of the skew mitigation process using both real and synthetic workloads that represent various skew configurations.

We explore automatic generation of example queries from an ad hoc database. An ad hoc database is a collection of tables with unknown relationships gathered to serve a spe-cific, often transient, often urgent, purpose. Consider these

Scientists today have the ability to generate data at an unprecedented scale and rate and, as a result, they must increasingly turn to parallel data processing engines to perform their analyses. However, the simple execution model of these engines can make it difficult to implement efficient algorithms for scientific analytics. In particular, many scientific analytics require the extraction of features from data represented as either a multidimensional array or points in a multidimensional space. These applications exhibit significant computational skew, where the runtime of different partitions depends on more than just input size and can therefore vary dramatically and unpredictably. In this paper, we present SkewReduce, a new system implemented on top of Hadoop that enables users to easily express feature extraction analyses and execute them efficiently. At the heart of the SkewReduce system is an optimizer, parameterized by user-defined cost functions, that determines how best to partition the input data to minimize computational skew. Experiments on real data from two different science domains demonstrate that our approach can improve execution times by a factor of up to 8 compared to a naive implementation.

Advances in cyber infrastructure for virtual observatories are poised to allow scientists from disparate fields to conduct experiments together, monitor large collections of instruments, and explore extensive archives of observed and simulated data. Such systems, however, focus on the `plumbing' and frequently ignore the critical importance of rich, 3D interactive visualization, asset management, and collaboration necessary for interdisciplinary communication. The NSF Ocean Observatories Initiative (OOI) is typical of modern observatory-oriented projects-its goal is to transform ocean science from an expeditionary science to an observatory science. This paper explores the design of an interactive tool to support this new way of conducting ocean science. Working directly with teams of scientists, we designed and deployed the Collaborative Ocean Visualization Environment (COVE). We then carried out three field evaluations of COVE: a multi-month deployment with the scientists and engineers of an observatory design team and two deployments at sea as the primary planning and collaboration platform on expeditionary cruises to map observatory sites and study geothermal vents.

The growing demand for large-scale data mining and data analysis applications has led both industry and academia to design new types of highly scalable data-intensive computing platforms. MapReduce and Dryad are two popular platforms in which the dataflow takes the form of a directed acyclic graph of operators. These platforms lack built-in support for iterative programs, which arise naturally in many applications including data mining, web ranking, graph analysis, model fitting, and so on. This paper presents HaLoop, a modified version of the Hadoop MapReduce framework that is designed to serve these applications. HaLoop not only extends MapReduce with programming support for iterative applications, it also dramatically improves their efficiency by making the task scheduler loop-aware and by adding various caching mechanisms. We evaluated HaLoop on real queries and real datasets. Compared with Hadoop, on average, HaLoop reduces query runtimes by 1.85, and shuffles only 4% of the data between mappers and reducers.

As the datasets used to fuel modern scientific discovery grow increasingly large, they become increasingly difficult to manage using conventional software. Parallel database management systems (DBMSs) and massive-scale data processing systems such as MapReduce hold promise to address this challenge. However, since these systems have not been expressly designed for scientific applications, their efficacy in this domain has not been thoroughly tested. In this paper, we study the performance of these engines in one specific domain: massive astrophysical simulations. We develop a use case that comprises five representative queries. We implement this use case in one distributed DBMS and in the Pig/Hadoop system. We compare the performance of the tools to each other and to hand-written IDL scripts. We find that certain representative analyses are easy to express in each engine's highlevel language and both systems provide competitive performance and improved scalability relative to current IDL-based methods.

The concept of a mashup is gaining popularity as a rapid-development, reuse-oriented programming model to replace monolithic, bottom-up application development - a programming style well-suited to scientific data management applications. A variety of mashup development frameworks exist, but none fully address the needs of the scientific community. Specifically, scientists may be proficient programmers, but they use different tools than are usually supported in mashup frameworks (e.g., arrays and MATLAB rather than XML and Javascript). Further, typical scientific mashups may extract tens of thousands of records from terabyte-scale datasets. At this scale, visualization techniques must be employed to quickly convey the gross features of the data - paging through 10 records at a time is not sufficient. Finally, scientific mashups must accommodate users across a wide spectrum of skills and domain knowledge - the same mashup must be reconfigurable for use by specialists, researchers, policy makers, or K12 students. In this poster presentation, we articulate these requirements, describe an initial prototype scientific mashup system, provide visual examples of scientific mashups, and outline ongoing research directions.

Data analysis tasks at an Ocean Observatory require integrative and and domain-specialized use of database, workflow, visualization systems. We describe a platform to support these tasks developed as part of the cyberinfrastructure at the NSF Science and Technology Center for Coastal Margin Observation and Prediction integrating a provenance-aware workflow system, 3D visualization, and a remote query engine for large-scale ocean circulation models. We show how these disparate tools complement each other and give examples of real scientific insights delivered by the integrated system. We conclude that data management solutions for eScience require this kind of holistic, integrative approach, explain how our approach may be generalized, and recommend a broader, application-oriented research agenda to explore relevant architectures.

We demonstrate a Web service and client application for querying and visualizing datasets defined over a topological grid structure. The context for our interest in gridded datasets is CORIE, an environmental observation and forecasting system designed to support scientific and industrial interests in the Columbia River estuary. The CORIE system both measures and simulates the physical properties of the estuary, generating 5GB of data and thousands of data products for each simulation run, including visualizations, aggregated results and derived datasets. In the current production CORIE system, "canned" visualizations are produced eagerly for every run. Users cannot customize their data products nor access the data directly, inhibiting data sharing. The term e-science is used to connote global, distributed collaboration enabled by sharing of both data and compute resources.