Reproduce “Auto-Formula: Recommend Formulas in Spreadsheets using Contrastive Learning for Table Representations” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sibei Chen, Yeye He, Weiwei Cui, Ju Fan, Song Ge, Haidong Zhang, Dongmei Zhang, Surajit Chaudhuri
Auto-Formula introduces a system for recommending spreadsheet formulas by identifying structural and stylistic similarities across an organizational corpus using contrastive learning for table representations. The system achieved an F1-score of 0.70 across diverse enterprise datasets, outperforming existing context-based models like SpreadsheetCoder (F1-score 0.064) and various GPT-based approaches (best RAG F1 ~0.25), while providing sub-second online prediction latency.
Key insights
- Organizational spreadsheets often share underlying computational logic despite superficial differences, making 'similar-sheet' discovery a powerful approach for formula recommendation.
- Learning separate coarse-grained (CNN-based for sheets) and fine-grained (fully-connected for regions) representations is critical for balancing the efficiency of global similarity search with the precision needed for formula and parameter identification.
- Combining both content (semantic and syntactic) and style features for cell representation is essential for robustly capturing spreadsheet characteristics and achieving high prediction quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.