MAINLE: a Multi-Agent, Interactive, Natural Language Local Explainer of Classification Tasks

European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD)

Paulo Bruno de Sousa Serafim¹   Rômulo Freire Férrer Filho²   Stenio Wagner Freitas²
Gizem Gezici³   Fosca Giannotti³   Franco Raimondi¹   Alexandre Magno Monteiro Santos²

¹Gran Sasso Science Institute (GSSI)  ²Federal University of Ceará (UFC)  ³Scuola Normale Superiore di Pisa (SNS)

---

Pre-Print: [PDF]  Paper: [ACM]

Abstract

There is an increasing need to explain machine learning decisions in an understandable way, even for non-expert users. In this paper, we introduce a multi-agent architecture to provide interactive explanations for classification tasks based on a range of machine learning algorithms, so that end-users can obtain answers in natural language. Our architecture is composed of four agents that are able to convert any classifier into a surrogate Decision Tree around the neighbourhood of a classification instance, which is then translated into a natural language explanation that can be further explored in an interactive way. We validate our approach against publicly available datasets using different classification methods, discussing the relevance of the architecture along five quality attributes, and performing a user study to evaluate the generated explanations. Our results show that the proposed architecture is able to generate simplified explanations that are more understandable for non-expert users in comparison to the ones given directly by a single explainer in all evaluated criteria.

Slides

[Slides] [PDF]

BibTeX

Download

@inproceedings{serafim2025mainle,
  author = {Serafim, Paulo Bruno and
    F\'{e}rrer Filho, R\^{o}mulo and
    Freitas, Stenio and
    Gezici, Gizem and
    Giannotti, Fosca and
    Raimondi, Franco and
    Santos, Alexandre},
  title = { {MAINLE}: A Multi-Agent, Interactive, Natural Language Local Explainer of Classification Tasks},
  booktitle = {Machine Learning and Knowledge Discovery in Databases. Research Track: European Conference, ECML PKDD 2025, Porto, Portugal, September 15--19, 2025, Proceedings, Part IV},
  publisher = {Springer-Verlag},
  address = {Berlin, Heidelberg},
  pages = {149--165},
  numpages = {17},
  year = {2025},
  isbn = {978-3-032-06077-8},
  doi = {10.1007/978-3-032-06078-5_9}
}