Going Beyond Local: Global Graph-Enhanced Personalized News Recommendations
“News Recommendation (NR) is the process of recommending news articles to users to satisfy their need for information by optimizing the accuracy of predicting relevance between news articles and users. Compared with making recommendations in other domains, news recommendation is more challenging because of the highly dynamic environment caused by the natural characteristics of news articles, such as timeliness, novelty, etc., for example, rapid changes occur in the relevance of news articles [6]. Therefore, unsurprisingly, content-based recommendation methods [1, 17, 22, 28–30] have proven their effectiveness. Typically, these are achieved via harnessing various natural language processing (NLP) and machine learning (ML) technologies to extract user interest representations by analyzing news articles that they have read in the past, to build representations of candidate news articles by studying their content, and then match users with candidate news articles.
Recently, deep learning-based technologies have been developed rapidly and brought new opportunities for enhancing content-based solutions due to their strong ability to deal with textual data, such as news content, news title, etc., and capture sequential dependencies in user reading history. For example, [30] defines a news encoder and a user encoder based on multi-head self-attention mechanisms, which learn news representations from titles and learn user representation from news embeddings. [28] acquires additional semantic information by using both word-level and view-level attention mechanisms to the news encoder to select important words and views for learning informative news representations. Research on modeling user preferences has gained popularity. [22] employs a three-layer hierarchical structure to learn user interests from the subtopic and topic levels of news articles. Additionally, graph-based methods represent another direction in this field. For instance, [20] utilizes a semantic augmentation graph to enhance news items and leverages a dual-graph interaction method to learn both candidate news and user representations.
Whilst these solutions have achieved great success in news recommendation tasks, such as alleviating the serious item user behaviors sparse problem by extracting meaning representation from news content, these methods mainly focus on the reading history of each user individually (i.e., using a single user’s historical news click sequence to construct this user’s representation), which lack a global border view of news articles and clicking among multiple users, and might be insufficient for uncovering more implicit hidden user behaviors. For example, as illustrated in Fig. 1, assume that we want to recommend news articles for user U1. User U1’s reading list contains only an orange news article, but at the moment, there are no news articles in the candidate set similar to the orange one. If recommendation models merely consider U1’s reading history, it is difficult to make recommendations for U1 in this situation. But if we can harness multiple user histories together, for example, extracting a Global News Graph from user U3 and U2’s clicks, we may find that there are certain relationships between orange news articles with blue news articles and green news articles. Therefore, the model may consider recommending dotted blue and green candidates that are similar to the blue one and the green one respectively. The challenge is how to properly generate this global news graph based on clicking from multiple users and integrate it with news recommendation systems.
To address the mentioned issue, we propose a novel model, GLORY (Global-LOcal news Recommendation system). Because historical news interaction data can provide more extensive and implicit relational information than semantic relevance, we propose a global-aware historical news encoder, specifically, using a global news graph to provide information on global perspectives for historical news. At the same time, to address the user behaviors sparse problem of candidate news, we propose a global-aware candidate news encoder, which uses a global entity graph to provide more effective associations for candidate news. Next, we use the multi-head self-attention mechanism to extract user interests from historical news. The final matching score came from the user news vector and the candidate news vector. We conducted extensive experiments on the MIND news dataset and Adressa, and the results showed that our model outperformed existing methods.”