Large-scale Data Management and AI with Applications to Oncology

BigLab! works on developing novel scalable AI technology to identify best practices for cancer treatment and care. Four threads of current work are below, followed by earlier lines of research. Each lists its papers, pulled from the publication database.

01

Large Number Model & data curation

The Large Number Model (LNM) is the centerpiece of the group’s current work, supported by an NSF CAREER Award (2026–2031). It is a new kind of LLM that comprehends and handles complex numerical and structured data accurately, together with the scalable data-curation infrastructure that supports it.

Language models capture language semantics well but struggle with tasks involving numbers. Treating numerical or structured data as ordinary terms is inadequate, because their semantics have to be understood and encoded by the model. CONE, a hybrid transformer encoder, embeds numbers, ranges, and gaussians into a vector space that preserves distance, while keeping unit and variable semantics.

The curation side turns documents into the structured data an LNM needs. Hemolix.Extract.V studies LLM-based information extraction: LLMs are more general than classic extractors but prone to hallucinations that hurt accuracy, so it adds AI-based plan selection. Hemolix.TabGen generates tables from documents.

The main application is cancer, but the LNM may also be useful for other quantitative domains such as finance.

Key ideas

  • Numbers, ranges, and gaussians embedded so that distance is preserved
  • Unit and variable semantics kept inside the embedding
  • LLM-based extraction made more reliable with AI-based plan selection
  • Table generation directly from documents

Papers

    02

    Cancer knowledge graphs

    CancerKG.ORG is a web-scale, interactive, and verifiable knowledge graph–LLM hybrid built to assist with optimal cancer treatment and care (ACM CIKM 2024). Its successor, CancerAIKG, is funded by the Florida Department of Health through the Casey DeSantis Florida Cancer Innovation Fund (about $1.2 million, 2024–current). It is a web-scale, trustworthy AI–knowledge graph–LLM hybrid, constructed and interrogated for bias using deep learning.

    The work is done with clinical partners. The consortium has subawards to Moffitt Cancer Center and Research Institute and to the University of South Florida in Tampa. In 2023 the group briefed the Executive Office of the Governor of Florida on this research with Moffitt.

    The approach was first developed on COVID-19. COVIDKG.ORG (EDBT 2023, with an NSF I-Corps award) is an interactive, trustworthy COVID-19 knowledge graph, also interrogated for bias with deep learning. In both domains tables on one topic, such as vaccine efficacy or side effects, appear under hundreds of different schemas across papers, which makes them hard to find and query. The group learns topical structured interfaces that standardize access to them.

    Key ideas

    • Knowledge graph and LLM combined, so answers are interactive and verifiable
    • Graphs constructed and interrogated for bias with deep learning
    • Standardized topical interfaces over tables with hundreds of schemas

    Papers

      03

      Tabular embeddings & profiling

      Large structured collections such as Web Data Commons (millions of tables from hundreds of thousands of sources) or CORD-19 are hard to use. Even for a single topic, tables from different sources come under hundreds of different schemas, which complicates both finding and querying them.

      The group builds scalable profilers and learned representations that make such corpora accessible. A meta-profile summarizes the metadata of an entire topic and can drive visualization and querying. Contrastive-learning classifiers identify multi-level hierarchical metadata in complex tables, and tabular embeddings encode the two-dimensional context of a table: horizontal and vertical hierarchical metadata, nesting, and units.

      Key ideas

      • Bi-dimensional tabular coordinates with a visibility matrix that separates horizontal metadata, vertical metadata, and data context
      • Unsupervised, scalable contrastive learning for hierarchical metadata classification, compared against the state of the art and recent LLMs
      • Meta-profiles as a standardized interface to a topic across a whole dataset

      Papers

        04

        Polystores & linear-relational systems

        Variety is one of the hardest problems in Big data. Data arrives in different shapes from many sources, and hundreds of data management systems expose it through their own APIs, storage schemes, and data models. Hybrid, the group’s family of systems, gives users one interface regardless of the shape, format, size, or backing store of the data.

        A core result is that a parallel or distributed relational database can be an excellent platform for scalable linear algebra. Relational systems already have cost-based optimization, which is vital to scaling linear algebra, and are known to scale. With a few changes, such a system becomes a competitive platform for linear-algebra workloads. This work received the IEEE ICDE 2017 Best Paper Award, an ACM SIGMOD Research Highlight Award (2018), and a CACM Research Highlight (2020).

        Hybrid.Poly extends this into an interactive, in-memory analytical polystore, with components for JSON ingest (Hybrid.JSON) and video ingestion (Hybrid.media).

        Key ideas

        • One interface to data of any shape, format, or storage system
        • Relational optimizers and scale-out reused for linear algebra
        • In-memory ingest and analytics for JSON, images, and video

        Papers

          +

          Earlier work

          Before the current threads, the group’s work centered on search over text and structured data, data integration, and health applications. Many of those ideas carry into today’s systems.

          Search & retrieval

          ReadFast and MedReadFast bring structural information retrieval to big text, especially biomedical and clinical text. Type-aware Web search argues that a system specialized to a domain or data type can outperform general keyword web search.

          Papers

            Data integration

            DataTamer fuses text and structured data at scale, and DataXFormer finds value mappings in web tables for semantic transformations. UFO (Unified Famous Objects), including the IBM UFO Repository, addresses the structural mismatch that arises when the same semantic objects are stored differently across systems.

            Papers

              Medicine & health sensing

              Working with Anna Pyayt, the group built an mHealth dipstick analyzer for monitoring pregnancy complications (IEEE Sensors 2016 Best Paper), EEG-based fear and emotion detection, and rapid antibiotic susceptibility analysis using microscopy and machine learning. This line was supported in part by an NSF subcontract (2017–2019).

              Papers