Lakemlb Eval

lakemlb-eval

LakeMLB: Data Lake Machine Learning Benchmark — Pan et al. (2026) (arXiv:2602.10441, 2026)

What this evaluates

Evaluates tabular machine learning models in data lake environments by leveraging auxiliary tables to improve prediction on a target table. It probes two integration paradigms: table unionability (vertical concatenation to increase training samples) and table joinability (horizontal enrichment to add features).

Datasets

• LakeMLB — total ?; splits: Union (-1), Join (-1); repo https://github.com/zhengwang100/LakeMLB

Metrics

• predictive performance (primary) — range: [0, 1]
  • Standard classification metrics (e.g., accuracy, F1-score) computed on the target table's label column. The exact metric is not specified in the provided section.

Input / output format

Input: A target table with a specified label column to predict, and a single auxiliary table (either unionable or joinable) retrieved from the data lake.

Output: Predicted class labels for each row in the target table.

Scoring recipe

def score(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return correct / len(gold_labels)

Common pitfalls

• Union tasks require semantic column alignment rather than exact name matching.
• Join tasks rely on fuzzy value overlap on join keys, which can cause alignment errors if not carefully validated.
• The benchmark assumes a single auxiliary table per task, though real data lakes may contain multiple.

Evidence (verbatim from paper)

The objective is to jointly leverage the target table and the auxiliary table to train a machine learning model that achieves improved predictive performance on the target table.

Citation

@misc{pan2026lakemlb,
  title={LakeMLB: Data Lake Machine Learning Benchmark},
  author={Pan et al. (2026)},
  year={2026},
  note={arXiv:2602.10441}
}

• arXiv: 2602.10441