← rodrigohuber.github.io← rodrigohuber.github.io

FOIS 2026 · Early Career Symposium · Vitória, BrazilBrasil

How Much Local Data Is Enough?

An ontology-grounded neurosymbolic research agenda for transferability in forecasting.

Uma agenda de pesquisa neurossimbólica, ancorada em ontologia, para a transferibilidade em previsão.

Rodrigo Huber Marques Moreira Mendes · Fernanda Baião · Reinaldo Castro Souza

Pontifícia Universidade Católica do Rio de Janeiro (PUC-Rio), Department of Industrial Engineering

Pontifícia Universidade Católica do Rio de Janeiro (PUC-Rio), Departamento de Engenharia Industrial

Scheduled presentationsApresentações programadas

Lightning talkThursday 24 September 2026 · 14:00–15:30 session, Early Career Symposium
Apresentação relâmpagoQuinta-feira, 24 de setembro de 2026 · sessão das 14h às 15h30, Early Career Symposium
PosterSame day, during the 15:30–16:00 coffee break
PôsterNo mesmo dia, no coffee break das 15h30 às 16h

Times are local (Vitória, UTC−3), as listed in the published FOIS 2026 program.

Horário local (Vitória, UTC−3), conforme o programa publicado do FOIS 2026.

DownloadBaixar

The paper is also available as the official FOIS 2026 satellite-events preprint on the conference site.

O artigo também está disponível como preprint oficial dos eventos satélite do FOIS 2026, no site da conferência.

Related talksOutras apresentações

From the same PhD research at PUC-Rio: the empirical work the FOIS agenda grows out of.

Da mesma pesquisa de doutorado na PUC-Rio: o trabalho empírico de onde a agenda do FOIS nasce.

AbstractResumo (original, em inglês)

Renewable-energy planning depends on reliable solar-irradiation forecasts, yet measurement infrastructure remains sparse across much of Brazil, leaving practitioners with abundant data at a few well-instrumented sites and almost nothing elsewhere. This paper addresses that scarcity through a neurosymbolic (NSAI) approach: rather than treating transfer learning as a purely statistical operation, I propose to embed a well-founded domain ontology inside the decision procedure that should govern whether a model trained at a data-rich source site may be deployed at a data-scarce target site, and if so, how much local data the target needs before adaptation is warranted. The central question is partly conceptual and partly empirical: when can a model trained in one context be trusted in another, and how much local evidence is needed before that trust is justified? Within this paradigm, symbolic, connectionist, and statistical components are combined; the distinguishing choice here is to ground the symbolic component in a well-founded domain ontology, which would formalize domain knowledge about site similarity (climate type, biome, geography, radiative regime) so that ontological commitments become explicit, inspectable, and testable. To this end, I am developing a Neurosymbolic Semantic Gate: a symbolic decision layer that will operationalize these commitments to govern transfer. This statement lays out the motivation, the core ideas, and the experimental program: a principled investigation of when ontological structure can govern transfer, and of how much local data a target site truly needs.

Where the work standsEm que pé está

20 Brazilian sites where the reference data budget n* has been measured under one fixed protocol locais brasileiros onde o orçamento de dados de referência n* foi medido sob um protocolo fixo Largely completeQuase concluído
5 rival criteria for what makes two sites the same kind of place critérios rivais para o que torna dois locais o mesmo tipo de lugar To be tested against those budgetsA testar contra esses orçamentos
1 Neurosymbolic Semantic Gate between a trained model and a new site Semantic Gate neurossimbólico entre um modelo treinado e um novo local Proposed · still under developmentProposto · ainda em desenvolvimento

Where I want pushbackOnde quero ser contestado

  1. How I individuate and operationalize the similarity relations.
  2. The epistemic status I assign to a gate’s verdict.
  3. How to frame the interface between ontology and learning so that it speaks to both communities.
  1. Como eu individuo e operacionalizo as relações de similaridade.
  2. O status epistêmico que atribuo ao veredito do gate.
  3. Como enquadrar a interface entre ontologia e aprendizado para que ela fale às duas comunidades.

How to citeComo citar

Mendes, R. H. M. M., Baião, F., & Souza, R. C. (2026). How Much Local Data Is Enough? An Ontology-Grounded Neurosymbolic Research Agenda for Transferability in Forecasting. In Proceedings of FOIS 2026 Satellite Events, co-located with the 16th International Conference on Formal Ontology in Information Systems (FOIS 2026), Vitória, Brazil. CEUR Workshop Proceedings (to appear) (no prelo).

@inproceedings{mendes2026fois,
  author    = {Mendes, Rodrigo Huber Marques Moreira and Bai{\~a}o, Fernanda and Souza, Reinaldo Castro},
  title     = {How Much Local Data Is Enough? An Ontology-Grounded Neurosymbolic
               Research Agenda for Transferability in Forecasting},
  booktitle = {Proceedings of FOIS 2026 Satellite Events, co-located with the 16th International
               Conference on Formal Ontology in Information Systems (FOIS 2026)},
  address   = {Vit{\'o}ria, Brazil},
  series    = {CEUR Workshop Proceedings},
  year      = {2026},
  url       = {https://foisconference.org/assets/pdfs/satellitepreprints/Paper-207.pdf},
  note      = {Early Career Symposium. To appear}
}

ContactContato