Enhancing The Recommendation of High-impact Rare-event Business News for Professionals with LLM-based Augmentation

Abstract

Personalized news recommendation has become an essential tool for professionals around the world to keep track of news events matching their interests and alleviate information overload. Beyond personalization, an essential aspect of useful news recommendations for professional use is that they highlight events that are more significant and of higher impact. However, we find that state-of-the-art recommenders struggle to identify and recommend news about significant events. In this paper, we address this gap as follows. To mitigate the relative scarcity of news about significant events, we use an LLM to create a synthetic dataset of significant news seeded from business-relevant news in the MIND dataset. We train four state-of-the-art recommendation models (MINER, UNBERT, UniTRec, Fastformer) with synthetically enhanced versions of a subset of the MIND dataset. We find that this successfully improves the performance of two of the recommendation models on the MIND-large dataset restricted to news about significant events in terms of the MRR, NDCG@5 and Hit@5 metrics and the performance of UNBERT on the AUC metric. The contribution of this paper is three-fold: we highlight news significance as an important aspect of useful news recommendation, we demonstrate the use of generative LLMs to create synthetic datasets for training on rare data and lastly, we demonstrate that augmenting some recommendation models with more significant news improves news recommendation performance on the MIND dataset.