Oh That Looks Familiar: A Novel Similarity Measure for Spreadsheet Template Discovery

The introduction of a novel hybrid distance metric for spreadsheet similarity marks a significant advancement in the field of data management. Traditional methods often overlooked the spatial layouts and type patterns that define templates, leading to inefficiencies in identifying similar spreadsheets. The new approach utilizes semantic embeddings, data type information, and spatial positioning to create cell-level embeddings, allowing for more accurate comparisons. Experiments demonstrated its superior performance, achieving an Adjusted Rand Index of 1.00 compared to the 0.90 of the graph-based Mondrian baseline. This breakthrough not only facilitates large-scale automated template discovery but also opens doors for downstream applications such as retrieval-augmented generation over tabular collections, model training, and bulk data cleaning, thus enhancing the overall efficiency of data handling processes.