Problem
A company has customer tables (transactions, credit history) and many separate prediction tasks on them: loan default, product recommendation, will a new product be liked. Today a separate "bridge" is built for each task to translate patterns in the customer history into the target. Every bridge means its own model, its own training and its own maintenance.
Example from the article: a person is 42, has only a master's degree and a low income. The model has to answer whether he earns more than $50,000.
Solution in brief
Write each table row as a short text and ask the question in text. A language model then predicts the answer, either after fine-tuning with ordinary language-modeling loss or from a handful of examples in the prompt. The speaker says that with up to 16 examples this beats boosting, logistic regression and neural networks.
Questions this case answers
- How to predict an outcome from a structured record with a language model by serializing each record as text, without training a task-specific model
- How to handle missing feature values in a text-based predictor without imputation
- How to run recommendation tasks as text generation using a language model
- How to predict an outcome from a table row with an LLM without training a task-specific model
Tools and methods
- Row serialization — a table row becomes a sentence or a prompt; a missing feature is simply left out
- Few-shot / zero-shot prompting — a few labeled rows in the prompt instead of training a model
- P5 (Recommendation as Language Processing) — rating prediction and next-item recommendation as text generation
- Toolformer, ViperGPT, Visual ChatGPT, Socratic Models — ways to connect a language model to APIs, code and vision models
Google or a one-time email code. No password.