Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

InsightBench: Evaluating Business Analytics Agents Through Multi-Step Insight Generation

Gaurav Sahu, Abhay Puri, Juan Rodriguez, Amirhossein Abaskohi, Mohammad Chegini, Alexandre Drouin, Perouz Taslakian, Valentina Zantedeschi, Alexandre Lacoste, David Vazquez, Nicolas Chapados, Christopher Pal, Sai Rajeswar Mudumba, Issam Hadj Laradji

TL;DR

InsightBench presents the first automated benchmark for evaluating end-to-end, multi-step data analytics by LLM-based agents, using 100 synthetic ServiceNow datasets with planted insights across Descriptive, Diagnostic, Predictive, and Prescriptive categories. The framework employs a two-way evaluation with LLaMA-3-Eval, comparing agent-generated insights and summaries against expert-ground-truth notebooks, and demonstrates that AgentPoirot outperforms the Pandas Agent, especially with open-source backbones like LLaMA-3-70b approaching GPT-4-level performance. The authors show that goal specificity (SMART goals) and diverse questioning are crucial for robust analytics, and they provide extensive ablation results on goal types, question diversity, and evaluation strategies. InsightBench, with its reproducible data-generation pipeline and open evaluator, aims to push forward practical, end-to-end automated data analytics across enterprise domains and invites community contributions to broaden coverage.

Abstract

Data analytics is essential for extracting valuable insights from data that can assist organizations in making effective decisions. We introduce InsightBench, a benchmark dataset with three key features. First, it consists of 100 datasets representing diverse business use cases such as finance and incident management, each accompanied by a carefully curated set of insights planted in the datasets. Second, unlike existing benchmarks focusing on answering single queries, InsightBench evaluates agents based on their ability to perform end-to-end data analytics, including formulating questions, interpreting answers, and generating a summary of insights and actionable steps. Third, we conducted comprehensive quality assurance to ensure that each dataset in the benchmark had clear goals and included relevant and meaningful questions and analysis. Furthermore, we implement a two-way evaluation mechanism using LLaMA-3 as an effective, open-source evaluator to assess agents' ability to extract insights. We also propose AgentPoirot, our baseline data analysis agent capable of performing end-to-end data analytics. Our evaluation on InsightBench shows that AgentPoirot outperforms existing approaches (such as Pandas Agent) that focus on resolving single queries. We also compare the performance of open- and closed-source LLMs and various evaluation strategies. Overall, this benchmark serves as a testbed to motivate further development in comprehensive automated data analytics and can be accessed here: https://github.com/ServiceNow/insight-bench.

InsightBench: Evaluating Business Analytics Agents Through Multi-Step Insight Generation

TL;DR

InsightBench presents the first automated benchmark for evaluating end-to-end, multi-step data analytics by LLM-based agents, using 100 synthetic ServiceNow datasets with planted insights across Descriptive, Diagnostic, Predictive, and Prescriptive categories. The framework employs a two-way evaluation with LLaMA-3-Eval, comparing agent-generated insights and summaries against expert-ground-truth notebooks, and demonstrates that AgentPoirot outperforms the Pandas Agent, especially with open-source backbones like LLaMA-3-70b approaching GPT-4-level performance. The authors show that goal specificity (SMART goals) and diverse questioning are crucial for robust analytics, and they provide extensive ablation results on goal types, question diversity, and evaluation strategies. InsightBench, with its reproducible data-generation pipeline and open evaluator, aims to push forward practical, end-to-end automated data analytics across enterprise domains and invites community contributions to broaden coverage.

Abstract

Data analytics is essential for extracting valuable insights from data that can assist organizations in making effective decisions. We introduce InsightBench, a benchmark dataset with three key features. First, it consists of 100 datasets representing diverse business use cases such as finance and incident management, each accompanied by a carefully curated set of insights planted in the datasets. Second, unlike existing benchmarks focusing on answering single queries, InsightBench evaluates agents based on their ability to perform end-to-end data analytics, including formulating questions, interpreting answers, and generating a summary of insights and actionable steps. Third, we conducted comprehensive quality assurance to ensure that each dataset in the benchmark had clear goals and included relevant and meaningful questions and analysis. Furthermore, we implement a two-way evaluation mechanism using LLaMA-3 as an effective, open-source evaluator to assess agents' ability to extract insights. We also propose AgentPoirot, our baseline data analysis agent capable of performing end-to-end data analytics. Our evaluation on InsightBench shows that AgentPoirot outperforms existing approaches (such as Pandas Agent) that focus on resolving single queries. We also compare the performance of open- and closed-source LLMs and various evaluation strategies. Overall, this benchmark serves as a testbed to motivate further development in comprehensive automated data analytics and can be accessed here: https://github.com/ServiceNow/insight-bench.
Paper Structure (33 sections, 8 figures, 7 tables)

This paper contains 33 sections, 8 figures, 7 tables.

Table of Contents

  1. Introduction
  2. InsightBench -- An Enterprise Data Analysis Benchmark
  3. Creating InsightBench Datasets
  4. Data Tables
  5. Planting Insights
  6. Synthetic Data Generation
  7. Ground-Truth Analysis Notebooks
  8. Evaluating Agents on InsightBench
  9. Dataset Statistics
  10. Quality Assurance
  11. Experimental Setup
  12. Baselines
  13. Implementation Details
  14. Benchmark Results
  15. Main Results.
  16. ...and 18 more sections

Figures (8)

  • Figure 1: Existing benchmarks (top) assess the agents' ability to solve a highly specific data analytics query with pre-defined output templates. They often require an expert user to ask the questions. InsightBench (bottom), on the other hand, evaluates the LLM-based agents on the complete comprehensive data analytics processes. This includes evaluating the agents' ability to answer high-level questions by a general user, recommend the best specific tasks to address a specific goal, extracting insights across descriptive, diagnostic, predictive, and prescriptive categories, and summarizing both findings and recommending next steps (as demonstrated by AgentPoirot).
  • Figure 2: Stages of Benchmark Creation:(a) Shows a demo incidents table from ServiceNow for the Incident Management theme, detailing schema and data fields. (b) Demonstrates the process of creating one of the datasets in the benchmark that embeds a linear increasing trend in incident resolution times, also highlighting the role of the slope parameter in dictating the trend's strength. (c) Displays the multi-step annotation analysis, with each step involving a question, corresponding plot, insightful description, and classification of the insight type.
  • Figure 3: InsightBench breakdown across five key thematic areas.
  • Figure 4: Performance of different agents on InsightBench grouped by difficulty and dataset category.
  • Figure 5: Screenshots of the Interface used for Dataset Quality-check:
  • ...and 3 more figures