Are You Here to Stay? Unraveling the Dynamics of Stable and Curious Audiences in Web Systems

Rodrigo Alves

Mar 03

Why do influencers frequently request their subscribers to enable all notifications for their channels? This practice stems from their awareness that not all subscribers are regular viewers of their content. While the number of subscriptions is important, the true determinant of sustained interest lies in more than just subscriber count. The popularity dynamics of online content are influenced by various internal and external factors, including content quality, metadata, age, recommendation algorithms, keyword rankings, promotions, and social network effects. For instance, some videos maintain significant viewership for hundreds of weeks after their initial posting, while others experience the majority of their views within the first few hours. This indicates a concentrated viewership during the early stages.

Understanding the popularity of items in online systems, especially in recommendation scenarios, is crucial for the success of most of the online businesses. However, accurately predicting the enduring appeal of online content is challenging due to the distinct patterns exhibited in the popularity dynamics of online items. But why? These dynamics often involve one or more peaks of popularity bursts that intermingle with the regular and stable audience of the content.

To address this challenge, one must focus on differentiating two types of audiences: the curious and the loyal. The curious audience is attracted by external and viral events, such as gossip, while the loyal audience represents stable viewership. Empirical evidence reveals that content like keyword-discovered videos, popular TV episodes, and music videos maintain steady popularity over time, dominated by loyal audiences. In contrast, news, sports, and movie content often undergo rapid popularity surges followed by quick declines, mostly due to temporally limited events (e.g., breaking news). In these cases, the curious audience tends to prevail over the loyal audience. To better understand how loyal and curious audiences can help to improve recommendation systems key performance indicators (KPIs), we recommend our previous blog post here.

The primary challenge in distinguishing between stable and curious audiences arises from the lack of individual labels that differentiate these viewer types. Instead, we typically only have the total number of viewers for an observed series of events, which is a combination of both audience types. Disentangling these two audiences poses difficulties, particularly because during viral events, curious users tend to dominate the channel's activity, leading to a burst in the overall series of events. These bursts can become so prominent that they obscure the presence of stable users during these periods.

Another significant challenge arises from the potential for the stable audience to modify its typical behavior in response to bursts or external events. For instance, the unexpected death of Michael Jackson in June 2009 triggered a surge in media and web activity, leading to increased music sales, video views, and posts discussing him. During this period, both existing and new fans engaged with Jackson's work, transforming him into an enduring musical icon. This behavior is illustrated in the left-hand side of the figure above, where the blue line represents the cumulative web activity associated with Michael Jackson's YouTube searches by an American audience. Initially, it displayed a relatively constant growth rate until his death (vertical black line), followed by a sudden spike in interest. Over time, the blue line returned to a consistent growth rate.

We will illustrate how to distinguish between two types of web activities during such events: (1) regular stable activity (yellow line) and (2) activity driven by unexpected events generated by the curious audience (green line). Notably, the yellow curve changed its slope after Jackson's unexpected death, marking a significant transition event that not only led to a short-term burst of activity but also consistently altered the stable audience. The revival of his songs, tribute notes, and the younger generation's discovery of Jackson's work contributed to a sustained increase in interaction. Conversely, the end of Barack Obama's presidential term (right-hand side of the figure) resulted in reduced political activity and mentions.

Thus, understanding the change in popularity can be beneficial for recommender systems that can anticipate and respond to the current situation of an item. By accurately distinguishing between stable and curious audiences, these systems can better tailor recommendations, predict future trends, and optimize content delivery to maintain engagement and maximize viewer satisfaction. Fast reaction is critical, for instance, in the news industry, where audiences can quickly turn to competitors' websites for information, and losing the opportunity to fully leverage trending content could be detrimental. However, many events occur frequently and may only interest niche audiences (e.g., the injury of a volleyball player). Thus, the solution is not to classify every such event as "breaking news" and broadcast it to everyone, as this can negatively impact the loyalty of users who are not interested. Therefore, it is important to identify "curious" users who might be interested in these niche events, allowing for more personalized and relevant content delivery.

But how do we detect the intensity of both audiences if these systems are often only aware of the temporal dynamics of the items throughout the observed interactions? Notably, events like Michael Jackson's death and the end of Obama's term are observed through interactions, and if the system is not aware that such events can happen, it cannot harness the potential of a burst. This could impact revenue, retention, and other KPIs for recommenders.

To address this and other questions, our research proposes a novel methodology called BPoP (Burst-induced Poisson Process). BPoP is a mix of point processes that form a statistical framework to learn and infer about multiple series of events. Our research was accepted for publication in ACM The Web Conference, one of the most prestigious data mining/machine learning conferences and the premier event for web research. I had the pleasure of contributing to this research with top-class researchers from prestigious institutions located in four countries across three continents.

More formally, our model is able to flexibly incorporate dependencies between two hidden and underlying point processes involving the stable fanbase and the curious audience. Therefore, BPoP is capable of disentangling the observed audience (interactions, represented by blue dots) into two different stochastic point processes: one representing the curious audience (green) and another one the stable audience (yellow dots). The transitions of the stable audience (white dots), which we never observe, are influenced by the bursts observed in the curious audience. For that, we assume a generative process where we can define the parameters of the models and treat the labels of the observed points as latent variables, enabling recovery of the parameters with the traditional EM (Expectation Maximization) Algorithm. For detailed derivation, we refer to our paper, which may be more interesting to a technical audience. We show that BPoP mimics the bursts of events seen in real data and is also able to efficiently capture the time-varying background rates that realistically represent the fan base.

If you like our research, are interested in understanding more, have questions, or see opportunities to collaborate with us, please feel free to contact us. We look forward to discussing our research with you. You can also cite our paper as follows if any of this content is useful. Understanding, explaining, and measuring audiences of items - broadly known as item popularity - is key to producing more accurate recommendation algorithms.

References

Rodrigo Alves, Antoine Ledent, Renato Assunção, Pedro Vaz-De-Melo, and Marius Kloft. "Unraveling the Dynamics of Stable and Curious Audiences in Web Systems." In Proceedings of the ACM on Web Conference 2024, pp. 2464-2475. 2024.