Emb-GAM: an Interpretable and Efficient Predictor using Pre-trained Language Models

Chandan Singh, Jianfeng Gao

Keywords: Interpretability, Explainability, Additivity, Generalized additive model, Linearity

TL;DR: By using pre-trained language models to extract fixed-size representations, we can learn much more effective linear ngram models without sacrificing interpretability

Abstract:

Deep learning models have achieved impressive prediction performance but often sacrifice interpretability, a critical consideration in high-stakes domains such as healthcare or policymaking. In contrast, generalized additive models (GAMs) can maintain interpretability, but often suffer from poor prediction performance due to their inability to effectively capture feature interactions. In this work, we aim to bridge this gap by using pre-trained large-language models to extract embeddings for each input before learning a linear model in the embedding space. The final model (which we call Emb-GAM) is a transparent, linear function of its input features and feature interactions.
Leveraging the language model allows Emb-GAM to learn far fewer linear coefficients, model larger interactions, and generalize well to novel inputs (e.g. unseen ngrams in text). Across a variety of natural-language-processing datasets, Emb-GAM achieves strong prediction performance without sacrificing interpretability. All code for using Emb-GAM and reproducing our results is made available on github.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)

Community Implementations: [ 2 code implementations](https://www.catalyzex.com/paper/emb-gam-an-interpretable-and-efficient/code)