Source-linked AI summary
Graph Neural News Recommendation with Long-term and Short-term Interest Modeling
Linmei Hu, Chen Li, Chuan Shi, Cheng Yang, Chao Shao
TL;DR
News recommendation methods remain limited by sparse user-item interactions and insufficient use of high-order structure. The paper proposes GNewsRec, which models users, news, and topics in a heterogeneous graph while combining graph-based long-term interests with attention-based LSTM short-term interests. On real-world datasets, GNewsRec significantly outperforms state-of-the-art news recommendation methods.
Problem
Existing news recommendation methods suffer from data sparsity because they do not extensively exploit high-order structure, while frequently replaced news items create cold-start problems.
Method
GNewsRec builds a heterogeneous user-news-topic graph and uses graph neural networks for high-order representations, combining complete-history and recent-history interest modeling.
Results
GNewsRec significantly outperforms state-of-the-art methods on news recommendation in experiments on real-world datasets.
Takeaways & Limitations
Modeling topics and both long-term and short-term user interests provides the paper’s supported approach to news recommendation.
Takeaways & Limitations
The paper identifies frequent news replacement as a cold-start limitation of existing collaborative-filtering methods.
Abstract
from arXiv · showhide
With the information explosion of news articles, personalized news recommendation has become important for users to quickly find news that they are interested in. Existing methods on news recommendation mainly include collaborative filtering methods which rely on direct user-item interactions and content based methods which characterize the content of user reading history. Although these methods have achieved good performances, they still suffer from data sparse problem, since most of them fail to extensively exploit high-order structure information (similar users tend to read similar news articles) in news recommendation systems. In this paper, we propose to build a heterogeneous graph to explicitly model the interactions among users, news and latent topics. The incorporated topic information would help indicate a user's interest and alleviate the sparsity of user-item interactions. Then we take advantage of graph neural networks to learn user and news representations that encode high-order structure information by propagating embeddings over the graph. The learned user embeddings with complete historic user clicks capture the users' long-term interests. We also consider a user's short-term interest using the recent reading history with an attention based LSTM model. Experimental results on real-world datasets show that our proposed model significantly outperforms state-of-the-art methods on news recommendation.
1. Introduction
News recommendation must handle information overload, sparse interactions, and users’ distinct long-term and short-term interests. GNewsRec addresses these issues with a heterogeneous user-news-topic graph, graph neural networks, and attention-based LSTM modeling, and significantly outperforms state-of-the-art methods.
- Motivation: News recommendation helps users find relevant articles amid the growing volume of online news.Personalized systems automatically recommend a small set of articles to satisfy users’ preferences.
- Limitations of existing methods: Collaborative filtering and content-based methods have complementary limitations, including cold start, ignored collaborative signals, and data sparsity.Hybrid methods combine collaborative filtering and content, but existing approaches still fail to extensively exploit high-order user-behavior similarity.
- Motivation: Latent topic information can indicate user interests and alleviate sparse user-item interactions by helping news with few clicks aggregate more information.Topics provide a bridge for incorporating additional information into interactions.
- Method: GNewsRec constructs a heterogeneous user-news-topic graph and applies graph neural networks to propagate embeddings and encode high-order relationships.The resulting user and news representations model interactions among users, news, and topics.
- Method: Complete click histories represent long-term interests, while recent reading histories are modeled with an attention-based LSTM for short-term interests.The model combines both interests and compares the resulting user representation with candidate news for prediction.
- Results: GNewsRec significantly outperforms state-of-the-art methods on news recommendation in experiments on real-world datasets.The paper presents this result as an experimental demonstration of the proposed model’s performance.
2. Related Work
News recommendation methods include collaborative filtering, content-based, and hybrid approaches, but existing methods have limitations in exploiting graph structure and handling frequently replaced news items. The paper proposes GNewsRec to model high-order user–item structure with a heterogeneous graph while incorporating long- and short-term interests.
- Collaborative filtering methods model user and item preferences from historical interactions, commonly using parameterized embeddings or matrix factorization.
- Most CF-based methods use descriptive features such as IDs and attributes without considering higher-order information in the user–item interaction graph.
- CF methods also suffer from cold-start problems because news items are frequently substituted.
- Content-based and hybrid methods incorporate news content, feature interactions, semantic or knowledge representations, sequential patterns, or combinations of recommendation algorithms.
- GNewsRec combines user–item interactions and news content, constructing a heterogeneous graph and applying graph convolutional networks to exploit high-order structure.
3. Problem Formulation
The task is to predict whether a user will be interested in an unseen news item from historical interactions, recent click sequences, and news features. News is represented using title and page-profile information, while clicks provide both aggregate and temporal signals.
- The dataset contains K users, M news items, and a user–item interaction matrix defined from implicit feedback.
- The interaction value y_u,d equals 1 when user u clicked news d and 0 otherwise.
- Timestamped click histories provide each user’s recent click sequence, whose elements identify the news items clicked in sequence.
- The goal is to predict whether user u has potential interest in an unseen news item from the interaction matrix and recent click sequences.
- Each news item is represented by its title words and a page-content profile containing entities and their types.
4. The Proposed Method
GNewsRec combines CNN-based news features with a heterogeneous user-news-topic graph, GNN propagation, and attention-based LSTM modeling to represent long- and short-term interests. Topic links address interaction sparsity, while multi-layer propagation captures high-order user-news structure for recommendation.
- Overall architecture: GNewsRec has three main components: CNN text extraction, GNN long-term interest and news modeling, and attention-based LSTM short-term interest modeling.CNNs extract features from news titles and profiles; the resulting user representation combines long- and short-term interests for candidate-news recommendation.
- Text information extractor: Two parallel CNNs separately encode news titles and profiles, then concatenate their representations into the final news text feature.The title and profile CNNs use separate weight parameters before their outputs are combined.
- Heterogeneous graph: Topic information helps indicate user interests, alleviate user-item sparsity, and connect new documents to the graph for embedding updates.At testing time, topics for new documents are inferred using the estimated LDA model, enabling graph connections for documents absent from the constructed graph.
- Heterogeneous graph: The model constructs a heterogeneous graph containing users, news items, and latent topics, with edges representing user clicks and news-topic connections.Topics are mined through LDA, and the graph explicitly models interactions among the three node types.
- Long-term interest modeling: GNNs propagate embeddings across multiple layers so user and news representations encode relationships beyond immediate neighbors.A one-layer GNN uses immediate neighbors, whereas multiple layers aggregate information from broader neighborhoods up to H hops away.
- Interest modeling: Complete historical clicks support relatively stable long-term interest representations, while recent click sequences are modeled with attention over content and sequential information.The attention-based LSTM processes clicked-news embeddings and produces a short-term interest embedding.
5. Experiments
Experiments evaluate GNewsRec on Adressa datasets against several neural recommendation baselines and analyze its components and parameter settings. The model consistently improves the baselines, while ablations support the contributions of graph propagation, topic information, and both long- and short-term interest modeling.
- Baselines: The comparison includes DMF, DeepWide, DeepFM, DKN, and DAN, covering collaborative-filtering, hybrid, and content-based neural approaches.DMF uses implicit-feedback rating vectors, whereas the other baselines incorporate news content or feature interactions.
- Model comparison: More than 10.67% on F1 and 2.37% on AUC, GNewsRec improves all baselines on both datasets.The comparison reports results in terms of AUC and F1 metrics.
- Ablation analysis: Removing the GNN module causes a great performance decline, indicating that graph propagation captures high-order relationships useful for user and news representations.The model constructs a heterogeneous graph and applies GNN-based embedding propagation.
- Ablation analysis: Removing short-term interest modeling reduces performance by around 2% on both AUC and F1, supporting the use of both long-term and short-term interests.The comparison also finds the no-topic variant slightly better than the variant without short-term interest modeling.
- Ablation analysis: Topic information improves GNewsRec over the variant without topics and lets low-click or newly arriving news aggregate neighboring information through topics.The heterogeneous graph uses topics to connect news and propagate information despite sparse user-item interactions.
6. Conclusion
The paper concludes with GNewsRec, a graph neural news recommender that models user-news-topic interactions and combines long-term and short-term user interests. On a real-world dataset, it significantly outperforms state-of-the-art news recommendation methods.
- Model: GNewsRec constructs a heterogeneous user-news-topic graph and propagates embeddings with graph convolutional networks to encode high-order information.The graph models user-item interactions and supports learned user and news embeddings.
- User interest modeling: Complete click histories represent long-term interest, while an attention-based LSTM over recent reading history models short-term interest.The two interest representations are combined for user modeling and compared with candidate news representations for prediction.
- Results: Experimental results on a real-world dataset show that GNewsRec significantly outperforms state-of-the-art methods on news recommendation.