TTNet: Tabular Transfer Network for Few-samples Prediction

Abstract

Tabular learning has been widely used in practical scenarios to handle tabular data such as the type of data in a spreadsheet or a CSV file. In many applications, it is necessary to transfer knowledge from the abundant source tabular data to the few target data, since models on the few target data are always easy to overfit. However, existing tabular learning methods have not touched the problem of transfer learning between different tabular datasets based on deep neural networks. To this end, we propose in this paper a new Tabular Transfer Network (TTNet for short) to enable effective and efficient knowledge transferring on tabular data. To enable network structures adaptive for transferring learning, TTNet first designs a BucTab network which integrates multi-bucket network with imitated tree-based network structures. In addition, TTNet uses a meta-transfer learning strategy to enable fast adaption by pre-training in the source domain of tabular data. Experiments on benchmark and real-world datasets show that TTNet significantly outperforms the state-of-the-art approaches in terms of effectiveness and efficiency, e.g., a new-shop prediction task testing on the e-commerce platform Ele.me shows that TTNet brings 9% MAE reduction comparing with the baselines.