Graph Representations for Tabular Data Understanding

This direction is merged into another direction: structural information loss (table linearization). Its last version stays readable and citable here.

Version 1 of 118 papers25 conceptsfrozen 2026-10-04

The standard way to feed a table into a machine learning model is to "flatten" it—converting rows and columns into a long string of text. While this works for simple lookups, it destroys the very thing that makes a table useful: its two-dimensional structure. When a table is linearized, a cell’s relationship to its header or its neighbors in the same column becomes a matter of distant text offsets rather than direct adjacency.

To solve this, researchers are increasingly treating tables not as sequences, but as graphs. By representing cells as nodes and their structural relationships as edges, models can preserve the topological context necessary for complex reasoning and information extraction.

The Failure of Linearization

In a typical Large Language Model (LLM) pipeline, a table is converted into a format like Markdown or CSV. As Yuxiang Wang and colleagues point out in Beyond Linearization: Attributed Table Graphs for Table Reasoning, this approach causes models to lose track of information "hidden in the middle" of long prompts. More importantly, it makes it difficult for the model to align a specific cell value with its corresponding header if they are separated by hundreds of tokens.

By contrast, modeling a table as a graph allows the model to "see" the table’s layout. In these graphs, nodes typically represent cells, while edges represent relationships like "same row," "same column," or "header-of." This structure ensures that the semantic link between a header and a data cell is always a single "hop" away, regardless of how large the table is.

Building the Table Graph

The transition from a raw image or spreadsheet to a graph involves identifying the logical coordinates of every element. Baosheng Yu and the team behind TGRNet frame this as a graph reconstruction problem. Instead of just detecting boxes, TGRNet jointly predicts spatial boundaries and logical row/column indices, mapping each cell to a global coordinate system.

Once the nodes are established, the edges define the flow of information:

  • Structural Edges: Connect cells within the same row or column to preserve the 2D topology, as seen in the work of Rihui Jin in TabPrompt.
  • Semantic Edges: Link cells to their relevant headers or captions. In MatSKRAFT, Kausik Hira uses these edges to extract complex materials science data, where a single value might only make sense in the context of a specific row header and a column sub-header.
  • Similarity Edges: In tasks like missing data imputation, Indro Spinelli and the creators of GINN use graphs to connect similar data points across different rows, allowing the model to "borrow" information from related instances to fill in gaps.

Reasoning via Message Passing

The power of the graph representation lies in how models process it. Instead of reading the table from top-left to bottom-right, Graph Neural Networks (GNNs) use message passing to aggregate context.

In GraSSLM, Xingyao Zhang demonstrates that this graph-based reasoning allows models to capture "multi-dimensional context." For example, to answer a question about a specific entry, the model can simultaneously pull information from the row's primary key and the column's header. This is significantly more efficient than sequential processing, where the header might be thousands of words away from the relevant cell.

This structural awareness is particularly effective for "few-shot" learning. TabPrompt uses graph contrastive learning to pre-train on the inherent dependencies of tables (like column-wise semantic consistency). Because the model already understands how tables are "built" through their graph structure, it requires far fewer labeled examples to master new table-understanding tasks.

Applications in Complex Domains

The shift toward graph representations is most visible in fields with highly non-standardized data, such as scientific literature. Tanishq Gupta and Mausam developed DiSCoMaT to extract chemical compositions from tables. They found that by treating tables as graphs, they could navigate complex, nested headers that typically baffle standard NLP models.

Similarly, for question answering, the TAG-QA framework by Wenting Zhao uses a graph structure to "localize" relevant cells before generating a long-form answer. By reasoning over the table's topology, the model ensures that the generated answers are faithful to the spatial relationships of the original data, avoiding the "hallucinations" that occur when a model misaligns a value with the wrong row or column.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_graph_representations_for_tabular_data_u_7adf3be8,
  title = {Graph Representations for Tabular Data Understanding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-representations-for-tabular-data-understanding/rdr_7adf3be8cbee4e6b95491926c8004d7f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_graph_representations_for_tabular_data_u_7adf3be8_v1,
  title = {Graph Representations for Tabular Data Understanding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-representations-for-tabular-data-understanding/rdr_7adf3be8cbee4e6b95491926c8004d7f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}