TablEye: Seeing small Tables through the Lens of Images
Seung-eon Lee, Sang-Chul Lee
TL;DR
TablEye introduces a novel framework that converts tabular data into images to enable few-shot learning, leveraging image-based algorithms to improve performance on tabular tasks.
Contribution
It proposes a domain transformation approach that generates tabular images, allowing the use of image-based few-shot learning methods for tabular data without dataset constraints.
Findings
Outperforms TabLLM with 0.11 AUC improvement in 4-shot tasks.
Achieves 3.17% higher accuracy on average in 1-shot settings.
Demonstrates effectiveness of image-based domain transfer for tabular data.
Abstract
The exploration of few-shot tabular learning becomes imperative. Tabular data is a versatile representation that captures diverse information, yet it is not exempt from limitations, property of data and model size. Labeling extensive tabular data can be challenging, and it may not be feasible to capture every important feature. Few-shot tabular learning, however, remains relatively unexplored, primarily due to scarcity of shared information among independent datasets and the inherent ambiguity in defining boundaries within tabular data. To the best of our knowledge, no meaningful and unrestricted few-shot tabular learning techniques have been developed without imposing constraints on the dataset. In this paper, we propose an innovative framework called TablEye, which aims to overcome the limit of forming prior knowledge for tabular data by adopting domain transformation. It facilitates…
Peer Reviews
Decision·Submitted to ICLR 2024
TablEye addresses a challenging problem in few-shot tabular learning using a unique approach. Few-shot tabular learning is a relatively new and underserved area, and the paper contributes to this domain. TablEye introduces an innovative approach to few-shot tabular learning by bridging the gap between tabular and image data domains. The paper provides evidence of TablEye’s effectiveness through a series of experiments. TablEye consistently outperforms existing methods in multiple scenarios, inc
While the paper presents positive results, it lacks detailed discussions regarding the differences in dataset performances. A more in-depth analysis of why certain structures perform better on specific datasets would provide a deeper understanding of TablEye's strengths. It seems that the improvement in performance is partly based on the variety of structures since none of them perform well in most of the experiments. The experiments of T-M-C2, T-M-C3, T-M-C4 on comparison with TabLLM and in t
One of the notable strengths of this paper is the novel idea of utilizing image domain priors for few-shot tabular learning. This approach capitalizes on the inherent structure and relationships within image data to address the challenges of tabular learning, demonstrating its effectiveness in transferring knowledge to few-shot scenarios.
While TablEye represents a promising approach, it is not without its limitations. One concern is the potential scalability issues that might arise when dealing with tabular data possessing a substantial number of features. The transformation of tabular data into an image format could lead to image dimensions that are impractically large, which may hinder the method's scalability and efficiency. Additionally, the authors acknowledge that for heterogeneous tabular data, establishing meaningful spa
### Originality & Significance: The paper explores an interesting underlying idea to leverage information from a well-explored area (in this case the image domain) and transfer both prior knowledge and existing/proven algorithmic methods; ### Quality: - Data: Authors experiment on different datasets and consider different important aspects: 1) feature diversity (categorical vs. numerical), 2) task diversity (n-way classification), 3) applicability to/relevance for ‘real-world’ applications
While I do like the general underlying idea, there are several severe weaknesses present in this work – leading me to lean towards rejection of the manuscript in its current form. The two main areas of concern are briefly listed here, with details explained in the ‘Questions’ part: ### 1) Lacking quality of the “Domain Transformation” part This is arguably the KEY part of the paper, and needs significant improvement in two points: Underlying intuition/motivation/justification, as well as tech
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsDomain Adaptation and Few-Shot Learning · Multimodal Machine Learning Applications · Advanced Neural Network Applications