Date of Award

12-11-2025

Degree Type

Thesis

Degree Name

Master of Science (M.S.)

Department

Computer Science

First Advisor

Manar D. Samad

Abstract

Transfer learning is a cornerstone of artificial intelligence (AI), enabling the learning of new domains with limited labeled data by reusing knowledge from large-scale foundation models. Transfer learning has achieved remarkable success with vision and language models due to the homogeneity in image and text data. In contrast, a heterogeneous feature space, structured in rows and columns as tabular data, remains underexplored in the transfer-learning literature. The inductive bias of tabular data with heterogeneous feature spaces from disparate application domains is very different from image and text data, which complicates transfer learning. Several recent studies have attempted within-domain and limited cross-domain transfer learning, assuming common features between tables. This thesis proposes a novel approach to achieve transfer learning between tabular data of distinct domains without requiring any shared features. A cross-attention framework is presented to learn cross-domain attention between two transformer models trained on distinct tabular data sets. Instead of learning attention scores on key, query, and value representations of a transformer, this thesis proposes a data-agnostic approach to enable cross-domain transfer learning. The proposed approach reuses some selective key and value projection weights of a source pre-trained transformer in a new transformer to learn target tabular data. Extensive evaluation against state-of-the-art machine/deep learning and transfer learning methods for tabular data demonstrate the superiority of the proposed method in numerous experimental scenarios. The proposed method paves the way for developing foundation models for tabular data from disparate domains and reusing them for transfer learning with limited sample data.

Download

Share

COinS