Introduction
Social media platforms have become central hubs for information exchange and community building. Predicting a user's future topic participation is a critical task for enhancing recommendation systems, optimizing targeted advertising, and monitoring public sentiment. Traditional approaches often suffer from a narrow focus, either prioritizing a user's historical content (intrinsic preference) or focusing solely on network dynamics (extrinsic preference). This paper introduces NTP³, a joint learning framework designed to bridge this gap by integrating both personal history and social influence into a unified predictive model.
The NTP³ Framework
The NTP³ framework treats next topic participation as a multi-label classification problem, utilizing a four-stage pipeline to process social media data efficiently.
- Intrinsic Preference Learning: The model aggregates a user's posting history into a single document, which is then transformed into a low-dimensional vector using paragraph embeddings. This captures the user's long-term, personal interests.
- Extrinsic Preference Learning: To model social influence, an Inductive Graph Convolutional Network (GCN) is employed. This allows the framework to generalize to new users by aggregating features from their immediate social neighborhood.
- Binary Representation via Locality-Sensitive Hashing (LSH): A key innovation of NTP³ is the use of LSH to convert continuous embedding vectors into binary Hamming codes. This transformation not only accelerates computation through integer arithmetic but also enhances the model's ability to distinguish between users with subtle differences in interest.
- Multi-label CNN Classifier: The binary vectors are reshaped into a 2D matrix—effectively treating the user's interest profile as an "image"—and processed by a Convolutional Neural Network (CNN) to predict multiple potential topics.
Social Influence and Herding
A core assumption of the NTP³ framework is that users are significantly influenced by their social circles, a phenomenon often referred to as "herding." To validate this, the authors analyzed the topic distributions of users compared to their friends using KL-divergence.
As shown in the figure above, there is a clear correlation between the topics a user engages with and those popular among their friends. This empirical evidence justifies the inclusion of extrinsic preference features in the predictive model.
Experimental Evaluation
The NTP³ model was rigorously tested against 17 baseline methods across four major social media datasets: Twitter, Weibo, Zhihu, and Douban.
- Performance Gains: The model consistently outperformed state-of-the-art baselines like DeepWalk and standard GCNs. Notably, on the Weibo dataset, NTP³ achieved a Macro-Precision improvement of up to 65%. On Twitter and Douban, gains ranged from 5% to 22%.
- Efficiency and Scalability: The LSH-based binary representation allowed the model to maintain low computational overhead even as data dimensionality increased, making it suitable for large-scale real-world applications.
- Interpretability: By utilizing influence functions, the authors were able to trace predictions back to specific training samples. Their analysis revealed that users 2 to 3 "hops" away in the social graph often provide highly influential signals for predicting behavior.
Conclusion and Future Work
NTP³ demonstrates that combining intrinsic and extrinsic preferences through a binary-coded CNN framework significantly improves topic participation prediction. While highly effective on platforms like Twitter and Weibo, the model faced challenges on Q&A sites like Zhihu, where high volumes of shared content can obscure unique user preferences. Future research will focus on incorporating dynamic temporal signals, such as real-time trending topics, and demographic data to further refine the model's predictive accuracy.
Summary updated