Train Once, Use Flexibly: A Modular Framework for Multi-Aspect Neural News Recommendation

TL;DR

MANNeR addresses the rigidity of prior multi‑aspect neural news recommenders by introducing a modular framework that learns aspect‑specific encoders via metric‑based contrastive learning and combines their similarity scores linearly at inference. The final ranking score $s_{final}$ is a sum of content relevance $s_{CR}$ and weighted aspect scores $\sum_{A_p}\lambda_{A_p}s_{A_p}$, enabling ad‑hoc personalization or diversification without retraining. Empirical results on MIND and Adressa show superiority over state‑of‑the‑art baselines for both standard content personalization and single/multi‑aspect customization, with robust cross‑lingual transfer when using a multilingual PLM. The approach offers practical benefits in deployment flexibility and efficiency, since new aspects require only training a new A‑Module, while inference leverages modular, parallelizable components. Overall, MANNeR demonstrates that modular, linear‑combination scoring can achieve versatile, high‑quality recommendations across diverse objectives and languages.

Abstract

Recent neural news recommenders (NNRs) extend content-based recommendation (1) by aligning additional aspects (e.g., topic, sentiment) between candidate news and user history or (2) by diversifying recommendations w.r.t. these aspects. This customization is achieved by ``hardcoding`` additional constraints into the NNR's architecture and/or training objectives: any change in the desired recommendation behavior thus requires retraining the model with a modified objective. This impedes widespread adoption of multi-aspect news recommenders. In this work, we introduce MANNeR, a modular framework for multi-aspect neural news recommendation that supports on-the-fly customization over individual aspects at inference time. With metric-based learning as its backbone, MANNeR learns aspect-specialized news encoders and then flexibly and linearly combines the resulting aspect-specific similarity scores into different ranking functions, alleviating the need for ranking function-specific retraining of the model. Extensive experimental results show that MANNeR consistently outperforms state-of-the-art NNRs on both standard content-based recommendation and single- and multi-aspect customization. Lastly, we validate that MANNeR's aspect-customization module is robust to language and domain transfer.

Figures (11)

• Figure 1: Illustration of the MANNeR framework. ① We train aspect-specialized NEs (i.e. CR-Module for content-based personalization, A-Module for aspect-based similarity) with metric-based contrastive learning. ② Inference: we linearly aggregate aspect-specific similarity scores between candidate and clicked news for final ranking.
• Figure 2: Results of single-aspect customization for MANNeR and the best baseline on MIND.
• Figure 3: t-SNE plots of the news embeddings in the test set of MIND.
• Figure 4: Results of multi-aspect customization for MANNeR on MIND.
• Figure 5: XLT in single-aspect diversification, with modules trained on different (combinations of) source-language datasets and evaluated on the target-language dataset MIND. The line style indicates the metric, the color the source-language datasets used in training.