Structured Embedding Models for Grouped Data

TL;DR

This paper introduces structured exponential family embeddings (S-EFE), a novel method for learning group-specific embeddings that leverage shared information across related groups, improving interpretability and predictive performance.

Contribution

The paper develops S-EFE, which incorporates hierarchical and amortized sharing strategies to model group-specific variations in embeddings, extending exponential family embeddings to structured data.

Findings

S-EFE enables interpretable group-specific embeddings.

S-EFE outperforms standard EFE in predictive tasks.

Shared information improves embedding quality across diverse data types.

Abstract

Word embeddings are a powerful approach for analyzing language, and exponential family embeddings (EFE) extend them to other types of data. Here we develop structured exponential family embeddings (S-EFE), a method for discovering embeddings that vary across related groups of data. We study how the word usage of U.S. Congressional speeches varies across states and party affiliation, how words are used differently across sections of the ArXiv, and how the co-purchase patterns of groceries can vary across seasons. Key to the success of our method is that the groups share statistical information. We develop two sharing strategies: hierarchical modeling and amortization. We demonstrate the benefits of this approach in empirical studies of speeches, abstracts, and shopping baskets. We show how S-EFE enables group-specific interpretation of word usage, and outperforms EFE in predicting held-out data.