Topic-Centric Explanations for News Recommendation

TL;DR

This work tackles the lack of explainability in news recommendations by introducing Topic-Centric Explanations for News Recommendation, a BATM-based framework that jointly learns news representations and latent topics to generate interpretable explanations. It employs a Bi-level Attention-based Topical Model with K-topic attention heads to produce topic vectors and document representations, and uses additive attention (or a GRU variant) to build user representations, scoring relevance via $s_i = u^\top d_i$ and training with $\,\mathcal{L}_{NCE}$. The approach achieves state-of-the-art or competitive recommendation performance on the MIND dataset while providing quantitative topic-coherence measures (NPMI and Word2Vec) and case-study demonstrations illustrating interpretability. The results suggest that topic-aware explanations can enhance trust and transparency without sacrificing accuracy, and the authors provide avenues for future work including leveraging topic features to improve performance and conducting user studies.

Abstract

News recommender systems (NRS) have been widely applied for online news websites to help users find relevant articles based on their interests. Recent methods have demonstrated considerable success in terms of recommendation performance. However, the lack of explanation for these recommendations can lead to mistrust among users and lack of acceptance of recommendations. To address this issue, we propose a new explainable news model to construct a topic-aware explainable recommendation approach that can both accurately identify relevant articles and explain why they have been recommended, using information from associated topics. Additionally, our model incorporates two coherence metrics applied to assess topic quality, providing measure of the interpretability of these explanations. The results of our experiments on the MIND dataset indicate that the proposed explainable NRS outperforms several other baseline systems, while it is also capable of producing interpretable topics compared to those generated by a classical LDA topic model. Furthermore, we present a case study through a real-world example showcasing the usefulness of our NRS for generating explanations.

Figures (8)

• Figure 1: A simplified illustration of a standard recommender making a prediction based on the user embedding and the candidate news embedding from the respective encoders. Most existing NRSs only generate latent embeddings, making them difficult to understand. We aim to interpret these embeddings using latent topics. We map the latent feature of the user embedding to the corresponding history article and identify the topic descriptors of this article based on the corresponding news embedding.
• Figure 2: An example of the history sequence of news article clicks by a given user $U91836$, which reflects the broad interests of the user involving topics around travel, food, finance, and health. We highlight some topic indicators, such as "seafood" and "restaurant", for the food-related topic.
• Figure 3: The architecture of the proposed news recommendation framework, which consists of a model that encodes both news articles and users.
• Figure 4: An illustration of the process of using topics as recommendation explanations, where different colors represent different topics. The more saturated color indicates a higher topic weight for a given word.
• Figure 5: Box plot for the distribution of $NPMI$ scores and $W2V$ similarity scores for our $BATM\text{-}ATT$ model on the news classification (NC) and news recommendation (NR) tasks.