Prediction Focused Topic Models via Feature Selection

TL;DR

This paper introduces a prediction-focused supervised topic model that uses feature selection to retain only relevant vocabulary terms, resulting in more coherent topics without sacrificing prediction accuracy.

Contribution

It proposes a novel approach that leverages the supervisory signal to select features, improving topic coherence while maintaining predictive performance.

Findings

More coherent topics compared to existing methods

Maintains competitive prediction accuracy

Effective feature selection improves interpretability

Abstract

Supervised topic models are often sought to balance prediction quality and interpretability. However, when models are (inevitably) misspecified, standard approaches rarely deliver on both. We introduce a novel approach, the prediction-focused topic model, that uses the supervisory signal to retain only vocabulary terms that improve, or at least do not hinder, prediction performance. By removing terms with irrelevant signal, the topic model is able to learn task-relevant, coherent topics. We demonstrate on several data sets that compared to existing approaches, prediction-focused topic models learn much more coherent topics while maintaining competitive predictions.