A Hotspot-Aware Personalized News Recommendation Mechanism Based on DistilBERT-TC-MA

News Ranking

We rank potential news based on the modeling of user and news interactions to achieve personalized displays according to the user's individual interests. Relevance-based news ranking methods (Wu et al., 2023) typically involve sorting candidate news based on the user's personalized relevance. Numerous approaches straightforwardly evaluate the relevance between users and news by comparing their ultimate representations. For instance, (Goossen et al., 2011) gauge relevance by computing the cosine similarity of the CF-IDF feature vectors for users and news. Meanwhile, (Okura et al., 2017) predict relevance scores by utilizing the inner product between the representations of news and users. However, these methods have a potential issue—they tend to suggest news articles that are similar to those previously clicked by users. Therefore, some methods attempt to address this issue by striving to recommend content different from news previously clicked by the user. (Li et al., 2011) starts by ranking news articles according to their relevance to user interests. Subsequently, they refine the ranking list by incorporating factors such as news popularity and recency to generate the ultimate recommended list. Hence, the paper modifies the ranking approach by incorporating elements like news novelty, popularity, and timeliness. This aims to boost recommendation diversity and mitigate the identified issue. This is done without compromising the user experience during the process of exploring potential user interests.