business intelligence

This paper addresses the inefficiency, tight coupling, and heavy manual programming burden in table joining and transformation during the data preparation phase of self-service business intelligence (BI). Drawing insights from an analysis of 2,000 real-world BI projects, we propose— for the first time—the joint modeling of these two interdependent tasks. Our method introduces a novel graph-based model inspired by Steiner trees, which unifies join and transformation prediction within a single framework. It integrates graph neural networks, Steiner tree approximation algorithms, BI workflow pattern mining, and empirically grounded rule-augmented reinforcement learning. The approach provides theoretical guarantees on solution quality and overcomes the limitations of conventional isolated modeling. Evaluated on real BI datasets, our method achieves over 70% accuracy—significantly outperforming state-of-the-art domain-specific algorithms and large language models including GPT-4.

Existing ETL pipelines heavily rely on manual, context-sensitive design of transformation logic, resulting in poor generalizability and low reusability. To address this, we propose an example-driven autonomous ETL framework: given user-provided target data examples, it constructs a paired-sample-based planning engine that automatically infers and synthesizes high-fidelity, context-adapted data transformation programs. Integrated with modular ETL components and runtime monitoring, the framework enables end-to-end automation for multi-format, multi-structured, and multi-scale data processing. Experiments across 14 real-world, cross-domain datasets demonstrate that our approach substantially reduces human intervention while achieving high-precision transformations (average F1 score of 0.92), strong generalization across diverse schemas and formats, and practical engineering deployability.

Large language models (LLMs) struggle to jointly perform data processing, predictive modeling, and visual analytics in multi-step tabular reasoning tasks. Method: This paper proposes a three-tier collaborative reasoning engine: a Planner that decomposes natural-language queries into subtasks; a Coder that generates and sandbox-executes Python code; and a Grapher that performs semantic parsing of charts and enables dynamic feedback-driven correction. It introduces the first end-to-end “Plan–Execute–Insight” closed-loop paradigm, supporting cross-modal joint reasoning and task-state backtracking. Contribution/Results: Evaluated on WikiTableQuestions and TabFact benchmarks, the approach achieves state-of-the-art accuracy—significantly outperforming existing methods—demonstrating robustness and strong generalization capability in complex tabular reasoning and multi-hop semantic insight generation.

Business professionals—non-technical domain experts—lack appropriate tools and methodologies for effective what-if analysis (WIA), hindering data-informed decision-making. Method: We conducted a two-phase mixed-methods user study—comprising contextual interviews and in-situ task-based evaluations—to systematically characterize their analytical behaviors for the first time. Contribution/Results: Based on empirical findings, we propose three domain-grounded design principles: business-contextual data preparation, risk-aware assessment, and domain-knowledge integration. We implemented and validated these principles in an interactive visual analytics prototype. The study identifies three critical support gaps, empirically confirms that six classes of what-if techniques significantly improve decision efficiency and confidence, and yields eight actionable design guidelines for commercial business intelligence systems. This work bridges a key theoretical and practical gap in WIA research concerning non-technical users.

This work addresses the challenge of evaluating commercial analytics agents in multi-step insight generation. We introduce InsightBench, the first end-to-end benchmark comprising 100 real-world business datasets and human-annotated ground-truth insights, requiring agents to autonomously execute the full pipeline: question formulation, data analysis, and derivation of actionable insights. We propose a novel multi-step insight generation evaluation paradigm and an open-source dual-path assessment framework built on LLaMA-3, incorporating a quality assurance process that jointly enforces goal clarity and analytical depth. Experiments demonstrate that our AgentPoirot—integrating Pandas, SQL, and natural language reasoning—significantly outperforms single-step baselines (e.g., Pandas Agent). Results validate the feasibility of open-weight large language models for complex commercial analytics tasks. All datasets, code, and evaluation tools are publicly released.

Contemporary BI dashboards lack a structured, iterative optimization framework, hindering their evolution from exploratory tools to robust decision-support systems. Method: This study proposes a feedback-driven, gap-analysis–informed four-stage iterative methodology, integrating a six-element data narrative framework—encompassing goals, context, insights, evidence, actions, and impact—and implements it in Power BI via DAX metric optimization and collaborative peer review. Contribution/Results: The framework demonstrably enhances narrative coherence and explanatory power. Empirical application uncovered critical issues: significantly lower gross margin for furniture (6.94% vs. 13.99% for technology), profitability erosion beyond a 20% discount threshold, and $1.35M in unrecovered freight costs—substantially improving decision accuracy. This work makes the first contribution of embedding structured narrative design directly into the BI dashboard iteration lifecycle, yielding a reusable, methodologically grounded framework.

This study addresses the lack of systematic evaluation of data quality tools with respect to their measurement capabilities and integration with large language models (LLMs). It presents the first multidimensional assessment framework grounded in real-world enterprise use cases, systematically evaluating six prominent tools—including open-source solutions such as Great Expectations and Deequ, as well as commercial platforms like Informatica and Experian—across dimensions including rule definition, duplicate detection, metric aggregation, and uncertainty handling, along with their LLM integration mechanisms. The findings reveal that commercial tools offer more comprehensive functionality and初步 support for LLM-assisted rule generation, whereas open-source tools provide greater flexibility at the cost of higher implementation effort. Notably, none of the evaluated tools currently enable direct LLM-based data validation. This work provides empirical guidance for selecting data quality tools and advancing their integration with LLMs.

Practitioners face significant challenges in effectively transforming customer feedback data into actionable software improvements. Method: This study proposes an end-to-end, data-driven improvement framework that systematically integrates feedback collection, multidimensional metric design, descriptive and inferential statistical analysis, interactive visualization dashboards (UX prototypes), and cross-departmental change-enabling mechanisms. Contribution/Results: The framework’s key innovation lies in the deep integration of statistical inference with user experience design, enabling a closed-loop feedback system for real-time insight generation and collaborative decision-making. Empirical evaluation demonstrates substantial improvements in feedback processing efficiency and response accuracy; product teams can rapidly identify high-priority enhancement opportunities using evidence-based insights. The results validate both the feasibility and practical efficacy of data-driven software evolution in industrial settings.

Imperative process models (e.g., Petri nets) exhibit semantic and execution-level incompatibility with structured process data in relational databases, hindering data-driven compliance analysis. To address this, we propose an automated model-to-query translation method that maps Petri net models to relaxed SQL queries, incorporating declarative techniques—such as behavioral footprints—to formally encode process constraints. Our approach ensures semantic traceability while unifying model-driven and data-driven analysis. It enables direct generation of executable, verifiable database queries from formal process models, and is empirically validated on real-world industrial datasets. The core contribution is a computationally grounded bridge between imperative process models and relational data, significantly enhancing the reusability and practical applicability of existing process models in data-intensive, compliance-critical scenarios.

This work addresses the state desynchronization between natural language queries and dashboard interactions in multi-step business intelligence (BI) analysis by proposing the first agent-based digital twin framework. The approach couples large language model (LLM) agents with executable dashboard states and reconstructs a shared analytical context through unified interaction logs, thereby ensuring consistency across dialogue, user actions, semantic alignment, and provenance tracking. Innovatively leveraging a digital twin mechanism, the framework enables state-aware analytical summarization and traceable context management. Experimental results demonstrate a 20.0% absolute improvement in exact-match accuracy (from 43.3% to 63.3%), a partial-match accuracy of 70.8%, and a reduced timeout rate of 10.0%. User studies further confirm high task accuracy and a positive interactive experience.