Research
My research activity lies at the intersection of data management and artificial intelligence, with a focus on methods for integrating, understanding, querying, and explaining heterogeneous structured and semi-structured data.
Data Integration and Matching
This research area addresses the integration of heterogeneous data at the record, schema, and table levels. It includes the automatic configuration of rule-based Entity Matching systems through TuneR, the integration of noisy data sources in real-world settings, and the combination of transformer representations with Automated Machine Learning for entity deduplication. c1 · c2 · c4 · c10
A further component concerns the analysis of BERT-based Entity Matching models, with particular attention to their behaviour, performance, and sensitivity to data characteristics. j3 · j6 · c20
Additional contributions cover synonymous attribute identification, privacy-preserving schema matching, and unsupervised evaluation of data integration in the absence of ground truth. At the table level, Armadillo learns graph-based representations for estimating table overlap and identifying duplicate, related, or evolving tables in large collections. c5 · c6 · c14 · c22 · j7
Explainable and Trustworthy AI
This research area focuses on the transparency, interpretability, and actionability of machine learning systems. A substantial part of the work concerns Entity Matching models, for which explanations identify the attributes, words, or combinations of terms that influence matching decisions.
The proposed approaches include post-hoc explanations based on input perturbations through Landmark, intrinsically interpretable architectures such as WYM, and CREW, which organizes influential terms into compact and understandable clusters. These contributions investigate the relationship between explanation fidelity, interpretability, and usability. c9 · c11 · c13 · c15 · c17 · c19 · c24
The same research direction extends to decision-support systems in other domains. CREDIX studies expert-guided counterfactual explanations for credit-risk models, with the objective of producing realistic and actionable alternatives. Privacy-aware integration and transparent analysis of corporate information provide further connections with trustworthy data-centric AI. c22 · c25
Language Models for Tabular and Document Intelligence
This research area investigates the use of language models for understanding, querying, and reasoning over structured and semi-structured information.
GRI-QA provides a benchmark for question answering over complex environmental tables extracted from corporate sustainability reports. It covers numerical reasoning, heterogeneous table structures, and questions requiring the combination of information distributed across multiple tables and documents. c23
CLARIESG is an end-to-end framework for analysing ESG information in corporate reports. It combines table extraction, structured prompting, unit normalization, table selection, and program-based reasoning to support transparent and verifiable analyses of sustainability indicators and potential ESG-washing practices. c25
Related work addresses natural language understanding and educational applications, including argument-relation classification through discourse information and adversarial training, and CLAIRE, a controllable multi-agent framework for guided reading comprehension. c21 · c26
In-DBMS Machine Learning Inference
This research area studies the execution of machine learning pipelines directly within relational database management systems. The objective is to reduce data movement and simplify deployment by translating preprocessing operations and predictive models into SQL.
MASQ transforms machine learning pipelines into standard SQL queries, allowing inference to be executed where the data reside. The corresponding evaluation analyses relational databases as prediction-serving platforms, considering execution efficiency, deployment complexity, and compatibility with different models and preprocessing operations. c7 · j5 · c18
Additional Research Contributions
Anomaly Detection
Novelty and anomaly detection in industrial and network environments.
Financial Machine Learning
Machine learning methods and evaluation challenges in financial applications.
Parallel Data Integration
Parallel algorithms for relational integration and full-disjunction computation.