Data quality represents a critical factor in analytical processes and machine learning workflows. The presence of missing values, outliers, duplicates, invalid placeholders, and semantic inconsistencies can significantly compromise the reliability of analyses and the performance of predictive models. In this context, data profiling and data cleaning are essential preliminary activities, although they often require substantial manual effort, multi-disciplinary expertise, and ad-hoc procedures that are difficult to generalize across different datasets and application scenarios. This thesis presents the design and implementation of an agent-based pipeline for data profiling, data cleaning, and quantitative evaluation applied to tabular datasets in CSV format. The objective of the work is to explore the potential of AI agents as decision-making components for supporting data preparation workflows, maintaining robustness, traceability, reproducibility, and operational safety. To achieve this objective, the proposed system combines Large Language Models (LLMs) with an SQLite-based grounding layer, structured intermediate artifacts, syntactic and semantic validation procedures, and constrained generation of cleaning plans that are subsequently translated into deterministic executable scripts to ensure reliable execution. The architecture of the pipeline is organized into three main components: a reporting workflow dedicated to profiling and data quality assessment, a workflow for generating a cleaned datasets, and a quantitative evaluation phase. The evaluation framework compares the original raw dataset with the cleaned version using a composite criterion that integrates dataset completeness and model-based predictive performance for automatically distinguishing between classification and regression tasks in order to adapt the evaluation. In addition, the system incorporates an interactive human-in-the-loop mode that allows users to review and modify agent-proposed imputation strategies before executing the cleaning process. Experimental results confirm that the proposed approach is capable of producing consistent profiling reports, generating plausible imputation strategies, and achieving measurable improvements in dataset quality in a substantial portion of the considered cases. The results also reveal that the effectiveness of cleaning process is strongly dependent on the characteristics of the dataset, the type of anomalies involved, and the quality of the decisions made by agents \cite{Dasu2012StatisticalDistortion}. The main contribution of the thesis lies not only in the application of AI agents to data-centric tasks, but also in the definition and implementation of a controlled and methodologically grounded pipeline. The proposed pipeline aims to make the use of AI agents within data preparation workflows more controllable, verifiable, and explicitly measurable.
La qualità dei dati rappresenta un fattore critico nei processi analitici e nei flussi di lavoro di machine learning. La presenza di valori mancanti, outlier, duplicati, placeholder non validi e inconsistenze semantiche può compromettere significativamente l’affidabilità delle analisi e le prestazioni dei modelli predittivi. In questo contesto, il data profiling e il data cleaning costituiscono attività preliminari essenziali, sebbene richiedano spesso un notevole sforzo manuale, competenze multidisciplinari e procedure ad-hoc difficili da generalizzare a dataset e scenari applicativi differenti. Questa tesi presenta la progettazione e l’implementazione di una pipeline basata su agenti per il data profiling, il data cleaning e la valutazione quantitativa applicata a dataset tabellari in formato CSV. L’obiettivo del lavoro è esplorare il potenziale degli agenti di intelligenza artificiale come componenti decisionali a supporto dei processi di preparazione dei dati, mantenendo al contempo robustezza, tracciabilità, riproducibilità e sicurezza operativa. Per raggiungere tale obiettivo, il sistema proposto combina Large Language Models (LLM) con un livello di grounding basato su SQLite, artefatti intermedi strutturati, procedure di validazione sintattica e semantica e una generazione controllata di piani di pulizia che vengono successivamente tradotti in script eseguibili deterministici, al fine di garantire un’esecuzione affidabile. L’architettura della pipeline è organizzata in tre componenti principali: un workflow di reporting dedicato al profiling e alla valutazione della qualità dei dati, un workflow per la generazione di dataset puliti e una fase di valutazione quantitativa. Il framework di valutazione confronta il dataset originale grezzo con la versione ripulita mediante un criterio composito che integra la completezza del dataset e le prestazioni predittive basate su modelli, distinguendo automaticamente tra problemi di classificazione e regressione per adattare il processo di valutazione. Inoltre, il sistema incorpora una modalità interattiva human-in-the-loop che consente agli utenti di esaminare e modificare le strategie di imputazione proposte dagli agenti prima dell’esecuzione del processo di pulizia. I risultati sperimentali confermano che l’approccio proposto è in grado di produrre report di profiling coerenti, generare strategie di imputazione plausibili e ottenere miglioramenti misurabili nella qualità dei dataset in una parte significativa dei casi considerati. I risultati evidenziano inoltre che l’efficacia del processo di pulizia dipende fortemente dalle caratteristiche del dataset, dal tipo di anomalie presenti e dalla qualità delle decisioni assunte dagli agenti \cite{Dasu2012StatisticalDistortion}. Il contributo principale della tesi non risiede soltanto nell’applicazione di agenti di intelligenza artificiale a compiti data-centrici, ma anche nella definizione e implementazione di una pipeline controllata e metodologicamente fondata. La pipeline proposta mira a rendere l’utilizzo degli agenti di intelligenza artificiale nei processi di preparazione dei dati più controllabile, verificabile ed esplicitamente misurabile.
AI agent for data preparation: a pipeline for profiling, validation, and cleaning
YE, ZIHENG
2025/2026
Abstract
Data quality represents a critical factor in analytical processes and machine learning workflows. The presence of missing values, outliers, duplicates, invalid placeholders, and semantic inconsistencies can significantly compromise the reliability of analyses and the performance of predictive models. In this context, data profiling and data cleaning are essential preliminary activities, although they often require substantial manual effort, multi-disciplinary expertise, and ad-hoc procedures that are difficult to generalize across different datasets and application scenarios. This thesis presents the design and implementation of an agent-based pipeline for data profiling, data cleaning, and quantitative evaluation applied to tabular datasets in CSV format. The objective of the work is to explore the potential of AI agents as decision-making components for supporting data preparation workflows, maintaining robustness, traceability, reproducibility, and operational safety. To achieve this objective, the proposed system combines Large Language Models (LLMs) with an SQLite-based grounding layer, structured intermediate artifacts, syntactic and semantic validation procedures, and constrained generation of cleaning plans that are subsequently translated into deterministic executable scripts to ensure reliable execution. The architecture of the pipeline is organized into three main components: a reporting workflow dedicated to profiling and data quality assessment, a workflow for generating a cleaned datasets, and a quantitative evaluation phase. The evaluation framework compares the original raw dataset with the cleaned version using a composite criterion that integrates dataset completeness and model-based predictive performance for automatically distinguishing between classification and regression tasks in order to adapt the evaluation. In addition, the system incorporates an interactive human-in-the-loop mode that allows users to review and modify agent-proposed imputation strategies before executing the cleaning process. Experimental results confirm that the proposed approach is capable of producing consistent profiling reports, generating plausible imputation strategies, and achieving measurable improvements in dataset quality in a substantial portion of the considered cases. The results also reveal that the effectiveness of cleaning process is strongly dependent on the characteristics of the dataset, the type of anomalies involved, and the quality of the decisions made by agents \cite{Dasu2012StatisticalDistortion}. The main contribution of the thesis lies not only in the application of AI agents to data-centric tasks, but also in the definition and implementation of a controlled and methodologically grounded pipeline. The proposed pipeline aims to make the use of AI agents within data preparation workflows more controllable, verifiable, and explicitly measurable.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_07_Ye_Tesi.pdf
non accessibile
Descrizione: tesi
Dimensione
11.36 MB
Formato
Adobe PDF
|
11.36 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/260182