| Abstract: |
The abundance of data, coupled with recent advancements in computation, has revolutionized almost every aspect of modern life, and the central role of data-centric algorithms and data structures in this revolution is undeniable. As data-driven technologies grow roots in our lives, their drawbacks and potential harms when they make irresponsible decisions become increasingly evident.
To resolve such bias issues, this project revisits the pipeline of data management, including the classical approximate query processing data structures, through the lens of responsibility.
Our project envisions a future where responsibility metrics, such as reliability and trustworthiness, are a first-class citizen of database systems.
To effectively accomplish its objective, this project aims to design data-centric algorithms, indices, and data structures for AI and Big Data pipeline, with theoretical guarantees across domains such as entity matching, membership estimation, aggregate query estimation, and approximate ranking and selection query answering.
We plan to use the ACCESS resources for the evaluation of our proposed solution in practice on public benchmarks and to compare against STOA baselines.
GenAI, particularly LLMs and Foundation Models, has emerged as a versatile tool, extending its influence beyond traditional applications, including almost all aspects of Data Management. However, LLMs' randomized nature and historical stereotypes in data raise critical concerns regarding responsibility.
Therefore, we aim to study (a) the potential reliability issues of using these technologies within data management, and to develop effective resolutions for resolving these issues. For example, as an initial step, we developed a Monte Carlo method based on repeated sampling to find more reliable and unbiased outputs for an LLM task (NAACL 2024).
In the reverse direction, we plan to study (b) how GenAI technologies can help improve fairness in web database applications. For example, our system Chameleon (VLDB 2024) utilizes foundation models such as DALL.E to minimally augment a dataset with synthetically generated data to enhance the coverage of all groups in multi-modal datasets. |
0000-0002-5251-6186