Aligning Explanations for Recommendation with Rating and Feature via Maximizing Mutual Information

TL;DR

The paper tackles the problem of misalignment between natural language explanations for recommendations and the model-predicted ratings or salient item features. It introduces a model-agnostic Maximizing Mutual Information (MMI) framework that uses Mutual Information Neural Estimation (MINE) to provide a reward signal for reinforcement-learning-based fine-tuning of existing explanation generators, with additional KL and Entropy regularizers and a Dynamic Weight Average mechanism to manage multi-objective optimization. Empirical results on three real-world datasets show that MMI-enhanced explanations achieve superior alignment with predicted ratings or features while largely preserving the quality and fluency of the original explanations, with human studies confirming improved decision support. The work advances practical, knowledge-grounded explanation generation for recommender systems and offers a flexible path to balance interpretability and linguistic quality in explanations.

Abstract

Providing natural language-based explanations to justify recommendations helps to improve users' satisfaction and gain users' trust. However, as current explanation generation methods are commonly trained with an objective to mimic existing user reviews, the generated explanations are often not aligned with the predicted ratings or some important features of the recommended items, and thus, are suboptimal in helping users make informed decision on the recommendation platform. To tackle this problem, we propose a flexible model-agnostic method named MMI (Maximizing Mutual Information) framework to enhance the alignment between the generated natural language explanations and the predicted rating/important item features. Specifically, we propose to use mutual information (MI) as a measure for the alignment and train a neural MI estimator. Then, we treat a well-trained explanation generation model as the backbone model and further fine-tune it through reinforcement learning with guidance from the MI estimator, which rewards a generated explanation that is more aligned with the predicted rating or a pre-defined feature of the recommended item. Experiments on three datasets demonstrate that our MMI framework can boost different backbone models, enabling them to outperform existing baselines in terms of alignment with predicted ratings and item features. Additionally, user studies verify that MI-enhanced explanations indeed facilitate users' decisions and are favorable compared with other baselines due to their better alignment properties.

Figures (2)

• Figure 1: The overview of the MMI framework ( The backbone model in this figure belongs to post-hoc generation models. )
• Figure 2: The training curve of Entropy Reward Value under w./w.o. DWA settings. The value of y-axis represents $\gamma_{Entropy}(t).\pi_{Entropy}(\hat{e})$ in equation (14)