Tabular Learnwares Can Be Repurposed for Seemingly Irrelevant New Tasks

The learnware paradigm aims to help users solve new tasks by reusing existing well-trained models instead of starting from scratch, where a learnware consists of a model and the specification describing its capabilities. Numerous learnwares are accommodated by the learnware dock system. Note that it is very likely that when a new task passed by a user has never been tackled before, and there is no model that can be directly taken to address the user task. In this paper, we focus on the tabular task, and propose a method for reusing tabular learnwares for classification tasks with significantly different feature and label spaces, exploiting the potential of numerous existing specialized tabular models developed for various tasks. We find tabular learnwares that seem semantically irrelevant can be beneficial with new user tasks sometimes. The proposed method relies solely on model-predicted probabilities and does not require gradient information, making it applicable to a wide range of tabular models. Experiments demonstrate that tabular learnwares can be reused beyond their original purpose across heterogeneous tasks.