Zero-shot learning for tabular data?

Question

Can anyone point me to methods for zero-shot learning on tabular data? There is some very cool work being done for zero-shot learning on images and text, but I'm struggling to find work being done to extend these techniques to tabular data.

Answer 1

Zero-shot learning is a type of machine learning that allows a model to make predictions on previously unseen data. It is typically used in natural language processing and computer vision tasks, where it can be difficult or impractical to manually label all of the possible classes that a model might encounter. Zero-shot learning on tabular data is an active area of research, but it can be challenging because tabular data often lacks the intrinsic structure that makes zero-shot learning easier for other types of data, such as images and text.

One approach to zero-shot learning on tabular data is to use a multi-task learning framework, in which the model is trained to predict multiple related target variables simultaneously. The idea is that by training on related tasks, the model can learn a common representation that can be transferred to the unknown target variable.

Another approach is to use meta-learning, in which the model is trained to learn new tasks quickly by adapting its parameters based on a small number of examples. In the context of zero-shot learning on tabular data, the model would be trained on a variety of related tasks, and then fine-tuned on a small number of examples from the unknown target variable in order to make predictions on that variable.

Both of these approaches have been applied to various types of tabular data, such as medical records, customer data, and financial data. You can find more information and references to specific papers in the following article: A Survey on Zero-shot Learning.

---

• In the paper Meta-Learning for Zero-Shot Classification (Ravi and Larochelle, 2017), the authors propose a meta-learning algorithm for zero-shot classification on tabular data that learns to map data points to a latent space, where similar points are mapped to nearby locations. The authors demonstrate that their method outperforms other approaches on a number of benchmark datasets.

• In the paper Zero-Shot Learning with Knowledge Graphs (Zhang et al., 2018), the authors propose a method for zero-shot learning on tabular data that uses a knowledge graph to represent the data and its labels. They demonstrate that their method outperforms other approaches on a number of benchmark datasets.