This repository serves as the central hub for DeepTab — Effective Neural Networks on Tabular Data, a research project focused on developing and investigating effective neural approaches for learning from tabular data.
DeepTab explores methods for improving how neural models represent, retrieve, and exploit information from tabular datasets, with the goal of making neural approaches more effective and competitive across a wide range of tabular learning problems.
The code and experiments associated with individual research works are maintained in dedicated repositories listed below.
Beyond Single-Space Representation: Adaptive Multi-View Neighborhood Retrieval for Tabular Learning
An adaptive neighborhood retrieval method that combines multiple representation spaces to improve nearest-neighbor selection for tabular learning.
Reference:
Maia, B.; Sousa, E.; Buzatto, B.; Cunha, V.; Mantovani, T.; Nazaré, T.; Egydio, J.; Jordão, A.; Costa, A. H. R. Beyond Single-Space Representation: Adaptive Multi-View Neighborhood Retrieval for Tabular Learning. In: Structural, Syntactic, and Statistical Pattern Recognition — Joint IAPR International Workshops, S+SSPR 2026, Bern, Switzerland, August 24–26, 2026, Revised Selected Papers, Lecture Notes in Computer Science, vol. 16799, Springer, 2026. In press.
Repository:
https://github.com/c2d-usp/Adaptive-Multi-view-Neighborhood-Retrieval
On the Role of Neighborhood Purity in Local Test-Time Adaptation for Tabular Data
A local test-time adaptation approach that uses Neighborhood Component Analysis (NCA) to retrieve training neighbors in a supervised metric space, improving neighborhood purity for query-specific adaptation.
Reference:
Cunha, V.; Maia, B.; Buzatto, B.; Sousa, E.; Mantovan, T.; Nazaré, T.; Egydio, J.; Costa, A. H. R.; Jordão, A. On the Role of Neighborhood Purity in Local Test-Time Adaptation for Tabular Data. In: Conference on Graphics, Patterns and Images (SIBGRAPI), 2026.
Repository:
https://github.com/c2d-usp/deeptab-nca-ttlt
A plug-and-play coreset selection method that selects training samples by minimizing covariance mismatch with a reference set.
Status: Ongoing research.
Repository:
https://github.com/Erick-DAS/Covariance-Matched-Coreset-Selection
Additional DeepTab research repositories will be added here as the project evolves.
DeepTab is developed within the Centro de Ciência de Dados (C2D), Escola Politécnica, Universidade de São Paulo (USP).