Research map / Large language models / Retrieval and knowledge
Question-guided Insights Generation for Automated Exploratory Data Analysis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes QUIS, a fully automated exploratory data analysis system that iteratively generates LLM-driven insight cards (QUGen) and then finds statistically significant insights via beam search over subspaces (ISGen), without training.
Problem
Automated exploratory data analysis (EDA) is limited by goal-oriented approaches that require human-specified objectives and by RL-based systems that need dataset-specific training and heavy compute, so unanticipated insights are missed.
Approach
Generate analysis questions from data semantics only (schema, descriptions, basic statistics), refining them iteratively by feeding earlier generated insight cards back as in-context examples to boost coverage and diversity, then score candidate views with fixed statistical pattern functions (trend, outstanding value, attribution, distribution difference) and beam search over subspaces, requiring
Key results
- Human evaluation (six evaluators, three per insight, 1-5 scale) shows QUIS beats OnlyStats on Sales and Employee Attrition in relevance, comprehensibility and informativeness, while OnlyStats is slightly better on Adidas Sales.
- QUIS achieves higher normalized insight scores than OnlyStats across all three datasets.
- Iterative QUGen produces a rising number of unique insight cards across successive iterations (1 to 11 iterations, total outputs fixed at 20).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use QUIS instead of OnlyStats | Automated EDA · Sales | human evaluation (relevance, comprehensibility, informativeness) | Helped |
| Use QUIS instead of OnlyStats | Automated EDA · Employee Attrition | human evaluation (relevance, comprehensibility, informativeness) | Helped |
| Use QUIS instead of OnlyStats | Automated EDA · Adidas Sales | human evaluation | Hurt |
| Use QUIS instead of OnlyStats | Automated EDA · Sales, Adidas Sale, and Employee Attrition | average normalized insight score | Helped |
| Iterate question generation and use earlier Insight Cards as few-shot examples (ablation) | QUGen · Insight Card generation | number of unique Insight Cards | Helped |
Surprising findings
- OnlyStats outperformed QUIS on the Adidas Sales dataset in human evaluation, which the authors attribute to dataset characteristics favoring simpler analysis.
Limitations
- Evaluated only on three tabular datasets with a small human study (six participants).
- Compared against only one self-constructed baseline (OnlyStats) because prior systems were not reproducible.
- Insight types are limited to four predefined patterns (trend, outstanding value, attribution, distribution difference).
Experimental setup
Datasets: Sales Adidas Sales Employee Attrition
Benchmarks: Sales Adidas Sales Employee Attrition
Metrics: relevance comprehensibility informativeness normalized insight score unique insight cards
Baselines: OnlyStats
Builds on: chain-of-thought prompting Mann-Kendall trend test Kruskal-Wallis test beam search
Closest papers by meaning
- TAIA: Large Language Models are Out-of-Distribution Data Learners (NeurIPS 2024 · similarity 0.82)
- LiveXiv - A Multi-Modal live benchmark based on Arxiv papers content (ICLR 2025 · similarity 0.80)
- DACO: Towards Application-Driven and Comprehensive Data Analysis via Code Generation (NeurIPS 2024 · similarity 0.80)
- Autofocus Retrieval: An Effective Pipeline for Multi-Hop Question Answering With Semi-Structured Knowledge (NeurIPS 2026 · similarity 0.80)
- From Complex to Atomic: Enhancing Augmented Generation via Knowledge-Aware Dual Rewriting and Reasoning (ICML 2025 · similarity 0.80)
- MindGYM: What Matters in Question Synthesis for Thinking-Centric Fine-Tuning? (NeurIPS 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.