We also introduce GDC-SM, a new benchmark for evaluating schema matching algorithms on real-world biomedical data. Labeling this dataset was a huge collaborative effort between our group and colleagues at the NYU School of Medicine. The benchmark is available on Zenodo: zenodo.org/records/1496...
Our new paper "Magneto: Combining Small and Large Language Models for Schema Matching" has just been published in the new issue of #PVLDB! The paper introduces a new framework that combines both small and large language models for effective schema matching.
www.vldb.org/pvldb/vol18/...
@sigmod2025.bsky.social #SIGMOD #SIGMID2025
We also introduce Harmonia, our proof-of-concept prototype that implements this vision. It orchestrates specialized data integration algorithms and works with the user to create reproducible pipelines, boosting schema matching F1-score from 0.78 to 1.00 in our preliminary evaluation! #AI #LLMAgents
Data harmonization is a major bottleneck in many scientific fields. In our new paper, we present a vision for using LLM-based agents to streamline this slow, manual process of reconciling mismatched schemas and terms.
📢 Tomorrow, I'll be presenting our new paper on LLM-based agents for interactive data integration at the #SIGMOD2025 NOVAS workshop. I'll also be in Berlin for the whole week, so please reach out if you'd like to chat or hang out!
Paper: arxiv.org/abs/2502.07132
Transformer-based neural networks achieve impressive performance on coding, math & reasoning tasks that require keeping track of variables and their values. But how can they do that without explicit memory?
📄 Our new ICML paper investigates this in a synthetic setting!
🎥 youtu.be/Ux8iNcXNEhw
🧵 1/13
Maybe you will find this one-page proof for priority sampling, with applications to distinct elements and inner product sketches convenient to cover in class: epubs.siam.org/doi/abs/10.1... (full disclosure: I'm co-author and proofs are mainly due to Daliri and Musco)
Models are sensitive to minor changes in format, e.g., simply repeating the column name multiple times led to improvements in zero-shot settings. However, fine tuning seems to decrease the performance differences.
Very interesting! We also experimented with different column serialization formats and found somewhat similar results in a schema matching task (arxiv.org/abs/2412.08194).
The Data Management for End-to-End Machine Learning workshop (@deem-workshop.bsky.social) will be back at #SIGMOD2025! ✨
🔗 Check out the CfP: deem-workshop.github.io
📝 Submission deadline: March 21
📢 Notifications: April 25
Join us for the 9th edition in Berlin!
#DEEM2025
a recent paper discusses this: db.cs.cmu.edu/papers/2024/... the main reason for graph DBs success is the limitation of SQL for querying graphs, although relational DBs seem to be catching up since the addition of property graphs in the latest SQL 2023 standard.
Slides for "Table Foundation Models"
I explain why these models can strongly outperform tree-based models, what are the intuitions,
hopefully pointing to ways forward for more improvement
speakerdeck.com/gaelvaroquau...
@madelonhulsebos.bsky.social kicking off the 3rd Table Representation Learning workshop (@trl-research.bsky.social) at NeurIPS 2024. First keynote by @gaelvaroquaux.bsky.social.
