Zezhou Huang, Eugene Wu
Data profilers play a crucial role in the preprocessing phase of data analysis by identifying quality issues such as missing, extreme, or erroneous values. Traditionally, profilers have relied solely on statistical methods, which lead to high false positives and false negatives. For example, they may incorrectly flag missing values where such absences are expected and normal based on the data's semantic context. To address these, we introduce Cocoon, a data profiling system that integrates LLMs to imbue statistical profiling with semantics. Cocoon enhances traditional profiling methods by adding a three-step process: Semantic Context, Semantic Profile, and Semantic Review. Our user studies show that Cocoon is highly effective at accurately discerning whether anomalies are genuine errors requiring correction or acceptable variations based on the semantics for real-world datasets.
Zezhou Huang, Pavan Kalyan Damalapati, Eugene Wu
Analysts often struggle with analyzing data from multiple tables in a database due to their lack of knowledge on how to join and aggregate the data. To address this, data engineers pre-specify "semantic layers" which include the join conditions and "metrics" of interest with aggregation functions and expressions. However, joins can cause "aggregation consistency issues". For example, analysts may observe inflated total revenue caused by double counting from join fanouts. Existing BI tools rely on heuristics for deduplication, resulting in imprecise and challenging-to-understand outcomes. To overcome these challenges, we propose "weighing" as a core primitive to counteract join fanouts. "Weighing" has been used in various areas, such as market attribution and order management, ensuring metrics consistency (e.g., total revenue remains the same) even for many-to-many joins. The idea is to assign equal weight to each join key group (rather than each tuple) and then distribute the weights among tuples. Implementing weighing techniques necessitates user input; therefore, we recommend a human-in-the-loop framework that enables users to iteratively explore different strategies and visualize the results.
Zezhou Huang, Pavan Kalyan Damalapati, Eugene Wu
Text-to-SQL allows experts to use databases without in-depth knowledge of them. However, real-world tasks have both query and data ambiguities. Most works on Text-to-SQL focused on query ambiguities and designed chat interfaces for experts to provide clarifications. In contrast, the data management community has long studied data ambiguities, but mainly addresses error detection and correction, rather than documenting them for disambiguation in data tasks. This work delves into these data ambiguities in real-world datasets. We have identified prevalent data ambiguities of value consistency, data coverage, and data granularity that affect tasks. We examine how documentation, originally made to help humans to disambiguate data, can help GPT-4 with Text-to-SQL tasks. By offering documentation on these, we found GPT-4's performance improved by 28.9%.
Zezhou Huang
Entity matching is a critical challenge in data integration and cleaning, central to tasks like fuzzy joins and deduplication. Traditional approaches have focused on overcoming fuzzy term representations through methods such as edit distance, Jaccard similarity, and more recently, embeddings and deep neural networks, including advancements from large language models (LLMs) like GPT. However, the core challenge in entity matching extends beyond term fuzziness to the ambiguity in defining what constitutes a "match," especially when integrating with external databases. This ambiguity arises due to varying levels of detail and granularity among entities, complicating exact matches. We propose a novel approach that shifts focus from purely identifying semantic similarities to understanding and defining the "relations" between entities as crucial for resolving ambiguities in matching. By predefining a set of relations relevant to the task at hand, our method allows analysts to navigate the spectrum of similarity more effectively, from exact matches to conceptually related entities.
Zezhou Huang, Eugene Wu
Dashboards are vital in modern business intelligence tools, providing non-technical users with an interface to access comprehensive business data. With the rise of cloud technology, there is an increased number of data sources to provide enriched contexts for various analytical tasks, leading to a demand for interactive dashboards over a large number of joins. Nevertheless, joins are among the most expensive operations in DBMSes, making the support of interactive dashboards over joins challenging. In this paper, we present Treant, a dashboard accelerator for queries over large joins. Treant uses factorized query execution to handle aggregation queries over large joins, which alone is still insufficient for interactive speeds. To address this, we exploit the incremental nature of user interactions using Calibrated Junction Hypertree (CJT), a novel data structure that applies lightweight materialization of the intermediates during factorized execution. CJT ensures that the work needed to compute a query is proportional to how different it is from the previous query, rather than the overall complexity. Treant manages CJTs to share work between queries and performs materialization offline or during user "think-times." Implemented as a middleware that rewrites SQL, Treant is portable to any SQL-based DBMS. Our experiments on single node and cloud DBMSes show that Treant improves dashboard interactions by two orders of magnitude, and provides 10x improvement for ML augmentation compared to SOTA factorized ML system.
Zezhou Huang, Rathijit Sen, Jiaxiang Liu, Eugene Wu
Although dominant for tabular data, ML libraries that train tree models over normalized databases (e.g., LightGBM, XGBoost) require the data to be denormalized as a single table, materialized, and exported. This process is not scalable, slow, and poses security risks. In-DB ML aims to train models within DBMSes to avoid data movement and provide data governance. Rather than modify a DBMS to support In-DB ML, is it possible to offer competitive tree training performance to specialized ML libraries...with only SQL? We present JoinBoost, a Python library that rewrites tree training algorithms over normalized databases into pure SQL. It is portable to any DBMS, offers performance competitive with specialized ML libraries, and scales with the underlying DBMS capabilities. JoinBoost extends prior work from both algorithmic and systems perspectives. Algorithmically, we support factorized gradient boosting, by updating the $Y$ variable to the residual in the non-materialized join result. Although this view update problem is generally ambiguous, we identify addition-to-multiplication preserving, the key property of variance semi-ring to support rmse, the most widely used criterion. System-wise, we identify residual updates as a performance bottleneck. Such overhead can be natively minimized on columnar DBMSes by creating a new column of residual values and adding it as a projection. We validate this with two implementations on DuckDB, with no or minimal modifications to its internals for portability. Our experiment shows that JoinBoost is 3x (1.1x) faster for random forests (gradient boosting) compared to LightGBM, and over an order magnitude faster than state-of-the-art In-DB ML systems. Further, JoinBoost scales well beyond LightGBM in terms of the # features, DB size (TPC-DS SF=1000), and join graph complexity (galaxy schemas).
Zezhou Huang, Eugene Wu
Recent query explanation systems help users understand anomalies in aggregation results by proposing predicates that describe input records that, if deleted, would resolve the anomalies. However, it can be difficult for users to understand how a predicate was chosen, and these approaches are limited to errors that can be resolved through deletion. In contrast, data errors may be due to group-wise errors, such as missing records or systematic value errors. This paper presents Reptile, an explanation system for hierarchical data. Given an anomalous aggregate query result, Reptile recommends the next drill-down attribute,and ranks the drill-down groups based on the extent repairing the group's statistics to its expected values resolves the anomaly. Reptile efficiently trains a multi-level model that leverages the data's hierarchy to estimate the expected values, and uses a factorised representation of the feature matrix to remove redundancies due to the data's hierarchical structure. We further extend model training to support factorised data, and develop a suite of optimizations that leverage the data's hierarchical structure. Reptile reduces end-to-end runtimes by more than 6 times compared to a Matlab-based implementation, correctly identifies 21/30 data errors in John Hopkin's COVID-19 data, and correctly resolves 20/22 complaints in a user study using data and researchers from Columbia University's Financial Instruments Sector Team.
Zezhou Huang, Jiaxiang Liu, Haonan Wang, Eugene Wu
Modern dataset search platforms employ ML task-based utility metrics instead of relying on metadata-based keywords to comb through extensive dataset repositories. In this setup, requesters provide an initial dataset, and the platform identifies complementary datasets to augment (join or union) the requester's dataset such that the ML model (e.g., linear regression) performance is improved most. Although effective, current task-based data searches are stymied by (1) high latency which deters users, (2) privacy concerns for regulatory standards, and (3) low data quality which provides low utility. We introduce Mileena, a fast, private, and high-quality task-based dataset search platform. At its heart, Mileena is built on pre-computed semi-ring sketches for efficient ML training and evaluation. Based on semi-ring, we develop a novel Factorized Privacy Mechanism that makes the search differentially private and scales to arbitrary corpus sizes and numbers of requests without major quality degradation. We also demonstrate the early promise in using LLM-based agents for automatic data transformation and applying semi-rings to support causal discovery and treatment effect estimation.
Zezhou Huang, Pranav Subramaniam, Raul Castro Fernandez, Eugene Wu
AutoML services provide a way for non-expert users to benefit from high-quality ML models without worrying about model design and deployment, in exchange for a charge per hour ($21.252 for VertexAI). However, existing AutoML services are model-centric, in that they are limited to extracting features and searching for models from initial training data-they are only as effective as the initial training data quality. With the increasing volume of tabular data available, there is a huge opportunity for data augmentation. For instance, vertical augmentation adds predictive features, while horizontal augmentation adds examples. This augmented training data yields potentially much better AutoML models at a lower cost. However, existing systems either forgo the augmentation opportunities that provide poor models, or apply expensive augmentation searching techniques that drain users' budgets. Kitana is a data-centric AutoML system that also searches for new tabular datasets that can augment the tabular training data with new features and/or examples. Kitana manages a corpus of datasets, exposes an AutoML interface to users and searches for augmentation with datasets in the corpus to improve AutoML performance. To accelerate search, Kitana applies aggressive pre-computation to train a factorized proxy model and evaluate each candidate augmentation within 0.1s. Kitana also uses a cost model to limit the time spent on augmentation search, supports expressive data access controls, and performs request caching to benefit from past similar requests. Using a corpus of 518 open-source datasets, Kitana produces higher quality models than existing AutoML systems in orders of magnitude less time. Across different user requests, Kitana increases the model R2 from 0.16 to 0.66 while reducing the cost by >100x compared to the naive factorized learning and SOTA data augmentation search.
Shuo Zhang, Zezhou Huang, Eugene Wu
Data cleaning is a crucial yet challenging task in data analysis, often requiring significant manual effort. To automate data cleaning, previous systems have relied on statistical rules derived from erroneous data, resulting in low accuracy and recall. This work introduces Cocoon, a novel data cleaning system that leverages large language models for rules based on semantic understanding and combines them with statistical error detection. However, data cleaning is still too complex a task for current LLMs to handle in one shot. To address this, we introduce Cocoon, which decomposes complex cleaning tasks into manageable components in a workflow that mimics human cleaning processes. Our experiments show that Cocoon outperforms state-of-the-art data cleaning systems on standard benchmarks.
Zezhou Huang, Eugene Wu
Data analytics over normalized databases typically requires computing and materializing expensive joins (wide-tables). Factorized query execution models execution as message passing between relations in the join graph and pushes aggregations through joins to reduce intermediate result sizes. Although this accelerates query execution, it only optimizes a single wide-table query. In contrast, wide-table analytics is usually interactive and users want to apply delta to the initial query structure. For instance, users want to slice, dice and drill-down dimensions, update part of the tables and join with new tables for enrichment. Such Wide-table Delta Analytics offers novel work-sharing opportunities. This work shows that carefully materializing messages during query execution can accelerate Wide-table Delta Analytics by >10^5x as compared to factorized execution, and only incurs a constant factor overhead. The key challenge is that messages are sensitive to the message passing ordering. To address this challenge, we borrow the concept of calibration in probabilistic graphical models to materialize sufficient messages to support any ordering. We manifest these ideas in the novel Calibrated Junction Hypertree (CJT) data structure, which is fast to build, aggressively re-uses messages to accelerate future queries, and is incrementally maintainable under updates. We further show how CJTs benefit applications such as OLAP, query explanation, streaming data, and data augmentation for ML. Our experiments evaluate three versions of the CJT that run in a single-threaded custom engine, on cloud DBs, and in Pandas, and show 30x - 10^5x improvements over state-of-the-art factorized execution algorithms on the above applications.
Zezhou Huang, Jiaxiang Liu, Daniel Alabi, Raul Castro Fernandez, Eugene Wu
Recent data search platforms use ML task-based utility measures rather than metadata-based keywords, to search large dataset corpora. Requesters submit a training dataset and these platforms search for augmentations (join or union compatible datasets) that, when used to augment the requester's dataset, most improve model (e.g., linear regression) performance. Although effective, providers that manage personally identifiable data demand differential privacy (DP) guarantees before granting these platforms data access. Unfortunately, making data search differentially private is nontrivial, as a single search can involve training and evaluating datasets hundreds or thousands of times, quickly depleting privacy budgets. We present Saibot, a differentially private data search platform that employs Factorized Privacy Mechanism (FPM), a novel DP mechanism, to calculate sufficient semi-ring statistics for ML over different combinations of datasets. These statistics are privatized once, and can be freely reused for the search. This allows Saibot to scale to arbitrary numbers of datasets and requests, while minimizing the amount that DP noise affects search results. We optimize the sensitivity of FPM for common augmentation operations, and analyze its properties with respect to linear regression. Specifically, we develop an unbiased estimator for many-to-many joins, prove its bounds, and develop an optimization to redistribute DP noise to minimize the impact on the model. Our evaluation on a real-world dataset corpus of 329 datasets demonstrates that Saibot can return augmentations that achieve model accuracy within 50 to 90% of non-private search, while the leading alternative DP mechanisms (TPM, APM, shuffling) are several orders of magnitude worse.
Gagan Bansal, Wenyue Hua, Zezhou Huang, Adam Fourney, Amanda Swearngin, Will Epperson, Tyler Payne, Jake M. Hofman, Brendan Lucier, Chinmay Singh, Markus Mobius, Akshay Nambi, Archana Yadav, Kevin Gao, David M. Rothschild, Aleksandrs Slivkins, Daniel G. Goldstein, Hussein Mozannar, Nicole Immorlica, Maya Murad, Matthew Vogel, Subbarao Kambhampati, Eric Horvitz, Saleema Amershi
As LLM agents advance, they are increasingly mediating economic decisions, ranging from product discovery to transactions, on behalf of users. Such applications promise benefits but also raise many questions about agent accountability and value for users. Addressing these questions requires understanding how agents behave in realistic market conditions. However, previous research has largely evaluated agents in constrained settings, such as single-task marketplaces (e.g., negotiation) or structured two-agent interactions. Real-world markets are fundamentally different: they require agents to handle diverse economic activities and coordinate within large, dynamic ecosystems where multiple agents with opaque behaviors may engage in open-ended dialogues. To bridge this gap, we investigate two-sided agentic marketplaces where Assistant agents represent consumers and Service agents represent competing businesses. To study these interactions safely, we develop Magentic-Marketplace -- a simulated environment where Assistants and Services can operate. This environment enables us to study key market dynamics: the utility agents achieve, behavioral biases, vulnerability to manipulation, and how search mechanisms shape market outcomes. Our experiments show that frontier models can approach optimal welfare -- but only under ideal search conditions. Performance degrades sharply with scale, and all models exhibit severe first-proposal bias, creating 10-30x advantages for response speed over quality. These findings reveal how behaviors emerge across market conditions, informing the design of fair and efficient agentic marketplaces.
Eugene Wu, Yiru Chen, Haneen Mohammed, Zezhou Huang
Human-data interaction (HDI) presents fundamentally different challenges from traditional data management. HDI systems must meet latency, correctness, and consistency needs that stem from usability rather than query semantics; failing to meet these expectations breaks the user experience. Moreover, interfaces and systems are tightly coupled; neither can easily be optimized in isolation, and effective solutions demand their co-design. This dependence also presents a research opportunity: rather than adapt systems to interface demands, systems innovations and database theory can also inspire new interaction and visualization designs. We survey a decade of our lab's work that embraces this coupling and argue that HDI systems are the foundation for reliable, interactive, AI-driven applications.
Zezhou Huang, Krystian Sakowski, Hans Lehnert, Wei Cui, Carlo Curino, Matteo Interlandi, Marius Dumitru, Rathijit Sen
GPUs are uniquely suited to accelerate (SQL) analytics workloads thanks to their massive compute parallelism and High Bandwidth Memory (HBM) -- when datasets fit in the GPU HBM, performance is unparalleled. Unfortunately, GPU HBMs remain typically small when compared with lower-bandwidth CPU main memory. Besides brute-force scaling across many GPUs, current solutions to accelerate queries on large datasets include leveraging data partitioning and loading smaller data batches in GPU HBM, and hybrid execution with a connected device (e.g., CPUs). Unfortunately, these approaches are exposed to the limitations of lower main memory and host-to-device interconnect bandwidths, introduce additional I/O overheads, or incur higher costs. This is a substantial problem when trying to scale adoption of GPUs on larger datasets. Data compression can alleviate this bottleneck, but to avoid paying for costly decompression/decoding, an ideal solution must include computation primitives to operate directly on data in compressed form. This is the focus of our paper: a set of new methods for running queries directly on light-weight compressed data using schemes such as Run-Length Encoding (RLE), index encoding, bit-width reductions, and dictionary encoding. Our novelty includes operating on multiple RLE columns without decompression, handling heterogeneous column encodings, and leveraging PyTorch tensor operations for portability across devices. Experimental evaluations show speedups of an order of magnitude compared to state-of-the-art commercial CPU-only analytics systems, for real-world queries on a production dataset that would not fit into GPU memory uncompressed. This work paves the road for GPU adoption in a much broader set of use cases, and it is complementary to most other scale-out or fallback mechanisms.