Abstract
Network science plays a big role in the representation of real-world phenomena such as user-item bipartite networks presented in e-commerce or social media platforms. It provides researchers with tools and techniques to solve complex real-world problems. Identifying and predicting future popularity and importance of items in e-commerce or social media platform is a challenging task. Some items gain popularity repeatedly over time while some become popular and novel only once. This work aims to identify the key-factors: popularity and novelty. To do so, we consider two types of novelty predictions: items appearing in the popular ranking list for the first time; and items which were not in the popular list in the past time window, but might have been popular before the recent past time window. In order to identify the popular items, a careful consideration of macro-level analysis is needed. In this work we propose a model, which exploits item level information over a span of time to rank the importance of the item. We considered ageing or decay effect along with the recent link-gain of the items. We test our proposed model on four various real-world datasets using four information retrieval based metrics.
Introduction
In the rapidly evolving landscape of social media and e-commerce, users are often overwhelmed by an abundance of information. Recommendation systems are designed to mitigate this by surfacing relevant content. However, many existing algorithms fall prey to the "Matthew Effect," where items that are already popular continue to dominate the rankings, effectively burying new or "novel" content. This research addresses the challenge of predicting which items will become the "next big thing" by balancing historical popularity with recent momentum and novelty.
Defining and Predicting Novelty
The authors distinguish between two critical types of novelty:
- Absolute Novelty: Items that enter the "popular" list for the very first time.
- Temporal Novelty: Items that may have been popular in the distant past but have recently re-emerged after a period of relative obscurity.
Predicting these requires moving beyond simple link-counting (Degree) models, which tend to favor "stale" popular items.
The RBAE Model: Recent Behaviour with Aging Effect
To address these challenges, the authors propose the Recent Behaviour with Aging Effect (RBAE) model. This model is built on three core pillars:
- Parametric Interpolation: It utilizes a parameter ($\lambda$) to strike a balance between an item's long-term historical popularity and its immediate recent gains.
- Exponential Decay: It incorporates an aging factor ($\gamma$), based on the assumption that the influence of a past interaction (e.g., a rating or a share) decays exponentially over time. This ensures that newer interactions carry more weight in the ranking process.
- Ordinal Ranking Minimization: Rather than attempting to predict the exact number of future interactions, the model is optimized using gradient descent to minimize ranking errors, ensuring that items are correctly ordered relative to one another.
Experimental Setup and Datasets
The RBAE model was rigorously tested against four diverse real-world datasets:
- Netflix: Movie ratings spanning 2000–2005.
- Movielens 10M: A standard benchmark for recommendation systems.
- Facebook Wall Posts: Bipartite user-to-user interactions from 2004–2009.
- Twitter Re-tweet Data: A sample of 5,000 tweets used to analyze sharing dynamics.
Performance was evaluated using four metrics: Temporal Novelty ($TN$), Precision ($P$), Novelty ($Q$), and Area Under the Curve ($AUC$).
Results and Discussion
The experimental results demonstrate that the RBAE model consistently outperforms standard benchmarks, such as the Degree model and the Popularity-Based Predictor (PBP).
As shown in the figure above, RBAE (represented by the solid black line) maintains superior performance in identifying Temporal Novelty ($TN_{100}$) and achieving higher $AUC_{100}$ scores across the Movielens, Netflix, and Facebook datasets. The inclusion of the exponential aging factor is critical; it allows the model to "forget" older, less relevant popularity and focus on rising stars.
Furthermore, the model exhibits remarkable stability. While benchmark models often see their ability to detect novelty decline as the size of the "top $k$" list increases, RBAE's performance remains robust. On the Twitter dataset, while specialized models like the Reinforced Poisson Process Model (RPPM) were slightly better at predicting absolute retweet counts, RBAE was significantly more effective at identifying which tweets would become novel hits.
Conclusion and Future Directions
The RBAE model provides a generic, purely temporal framework for ranking items in evolving networks without requiring content-specific features. While highly effective at ranking, future work could improve the model's ability to predict absolute interaction counts. Additionally, integrating macro-level factors such as community structure and sentiment analysis could further refine its predictive power.
Summary updated