arXiv
GRAFT: Graph-Matched Retrieval and Fusion of Tables in Data Lakes
Daomin Ji, Hui Luo, Zhifeng Bao, Shane Culpepper, Shazia Sadiq
Problem
Data agents answering queries over tabular data lakes need multiple related tables, but retrievers that score tables independently return fragmented, hard-to-join evidence.
Method
GRAFT casts retrieval as graph matching between a query intent graph and the data-lake graph, scoring subgraphs with a log-determinant reward (relevance, joinability, diversity) learned via implicit Q-learning.
Results
On Spider and BIRD adapted to a data-lake setting, GRAFT beats the strongest point-wise/greedy/structure-aware baselines by 7.8% F1 and 10.6% on a sufficiency metric, while staying search-efficient.
Limitations
Likely: reward and Q-learning require the compression-operator trajectories to be well-formed, and the approach depends on Spider/BIRD-style schemas rather than messier real enterprise data lakes.
arXiv
Contrastive-Collapsed Loss for Flexible and Geometrically Optimal Embeddings and Faster Convergence
Blanca Cano-Camarero, Ángela Fernández-Pascual, José R. Dorronsoro
Problem
Standard classification losses (cross-entropy, dot regression) don't explicitly optimize for tight, well-separated class representations, limiting convergence speed and embedding quality on tabular data.
Method
CoCo (Contrastive-Collapsed loss) jointly pushes same-class embeddings to collapse together and different-class embeddings apart, giving better-conditioned gradients than prior objectives.
Results
On OpenML-CC18 tabular benchmarks, CoCo matches kernel SVM, Random Forest, dot regression and cross-entropy neural nets on accuracy while producing tighter class clusters and converging faster.
Limitations
Likely: gains are 'competitive with' rather than clearly beating strong baselines, and the abstract reports no results on TFMs/ICL settings or larger-scale enterprise tabular data.
No new updates in the industry today.