Abstract
The emergence of online social networks has provided a wealth of social activity data, enabling microscopic investigation into product adoption, usage, and competition. We propose a continuous-time probabilistic model based on temporal point processes to track the adoption and frequency of use of competing products. Our model incorporates social influence, recency, and competitive inhibition. We develop an efficient inference method using convex optimization that scales to large networks. Validation on synthetic and real Twitter data demonstrates that our model provides superior fit and predictive accuracy while offering interpretable insights into the drivers of user behavior.
Introduction
Traditional models of product adoption, such as the Bass diffusion model, typically focus on macroscopic population trends. However, these models often overlook the "microscopic" reality of individual behavior: the fine-grained, recurrent usage patterns and the constant competition between similar products (e.g., URL shorteners like Bitly vs. TinyURL). Existing microscopic models often treat adoption as a binary, one-time event or fail to account for the complex interplay of social network structures. There is a critical need for a model that handles large-scale, continuous-time data while simultaneously accounting for social influence, personal usage history, and the inhibitory effects of competition.
Methodology: Multivariate Hawkes Processes with Inhibition
The core of our approach is a modified Multivariate Hawkes Process. While standard Hawkes processes are "self-exciting"—where one event increases the likelihood of future events—we extend this framework to include both excitation (positive influence) and inhibition (negative influence).
Our model accounts for three primary factors:
- Recency (Habit): Recent use of a product increases the probability of subsequent use.
- Social Influence: Adoption by neighbors in a social network increases a user's likelihood of adoption.
- Competition: The use of a competing product by a user or their neighbors exerts an inhibitory effect, decreasing the intensity of use for the original product.
To ensure scalability, we formulated parameter estimation as a convex optimization problem that "decouples" into independent sub-problems for each user-product pair. This allows for massive parallelization across hundreds of thousands of nodes. The model's ability to capture fine-grained temporal patterns is illustrated below.
Experimental Evaluation
We validated the model using both synthetic and real-world datasets:
- Synthetic Networks: We utilized Kronecker graph models to simulate various social structures, including core-periphery, hierarchical, and random networks. This allowed us to test the model's robustness across different topologies.
- Twitter URL Shorteners: We tracked the usage of seven services (e.g., Bitly, TinyURL) among 3,000 active users.
- Twitter Retweet Conventions: We analyzed how users transitioned between different retweeting styles (e.g., "RT", "via") prior to the introduction of the official retweet button.
Key Insights and Results
Our analysis revealed several critical dynamics in social diffusion:
- Habit vs. Influence: Popular products like Bitly are primarily driven by spontaneous use and habit. In contrast, niche products require a "social nudge" from neighbors to trigger adoption, after which personal habit becomes the dominant driver.
- Asymmetric Competition: The model successfully captured market dominance. For instance, Bitly exerted a strong inhibitory effect on TinyURL, while the reverse was not true, indicating Bitly's established dominance even before becoming a platform default.
- Predictive Performance: On real Twitter data, our model achieved a prediction probability of 81-89%, significantly outperforming Poisson (32-45%) and Weibull (58-76%) baselines.
- Detecting Market Shifts: The model's goodness-of-fit (log-likelihood) dropped sharply when Twitter changed its default URL shortener in May 2009. This demonstrates the model's utility as a diagnostic tool for identifying external interventions or "black swan" events.
Conclusion and Future Work
The proposed model offers a scalable and interpretable framework for understanding product competition in social media. While highly effective, future work could explore non-parametric kernels for more complex temporal patterns and account for varying levels of influence between different types of social ties. Additionally, incorporating intrinsic product quality metrics could further enhance the model's explanatory power.
Summary updated