TLDR: A new research paper introduces a unified framework and public datasets for modeling both individual user actions (personal events) and interactions between users (relational events). The study demonstrates that combining these two event types significantly improves prediction tasks in various domains, highlighting the limitations of models that only consider one type of event and opening avenues for future research in comprehensive user behavior modeling.
The way we understand and predict how users interact with online systems is crucial for many applications, from e-commerce recommendations to fraud detection in finance. Traditionally, these interactions, or “user events,” have been divided into two main categories: personal events and relational events. Personal events are individual actions, like searching for a product or viewing an item. Relational events involve interactions between two or more users, such as sending a gift or commenting on a review.
For a long time, these two types of events have been modeled separately. Personal events are often handled with sequence-based methods, which are good at capturing the order and timing of individual actions. Relational events, on the other hand, are typically modeled using graph-based methods, which excel at showing connections and interactions between users. However, real-world systems often involve both types of events, and ignoring one can lead to an incomplete picture of user behavior.
A new research paper addresses this challenge by proposing a unified approach to model both personal and relational user events together. The authors introduce a new way to formally describe user events that incorporates both individual actions and interactions between users. They also release a collection of public datasets and prediction tasks specifically designed to encourage further research in this area. These resources are vital because there has been a lack of public datasets that explicitly combine both personal and relational events.
The paper highlights that in many real-world scenarios, personal events occur much more frequently than relational events. For instance, in an e-commerce platform, users might view many products or add items to their cart far more often than they send gifts or write co-reviews. Despite their high volume, personal events are often overlooked in traditional graph-based models, which tend to prioritize relational structures. The researchers argue that both types of events carry valuable, complementary information, and combining them can significantly improve predictive tasks like item recommendation, fraud detection, and customer profiling.
To demonstrate the benefits of their unified approach, the researchers conducted experiments on five diverse datasets: Brightkite, Gowalla, Amazon Clothing, Amazon Electronics, and GitHub. These datasets cover various domains, from location-based social networks to product reviews and software development activities. They vary significantly in terms of the number of users, events, and the ratio of personal to relational events, presenting a wide range of modeling challenges.
The study defined two main prediction tasks: relational event prediction (e.g., recommending friends or predicting co-review relationships) and personal event prediction (e.g., predicting future check-ins or product reviews). They tested several baseline models, including static graph models, sequence models, and temporal graph models, some of which were adapted to incorporate both event types.
The experimental results showed a clear pattern: models that integrated both personal and relational events generally performed better than those that relied on only one type. Specifically, static graph models augmented with embeddings from personal event sequences consistently achieved strong results across various datasets. This suggests that capturing the temporal dynamics of individual user actions, alongside the structural relationships between users, provides a more comprehensive understanding of user behavior.
Interestingly, simply converting personal events into nodes within a graph structure was often less effective than using sequence models to generate embeddings from personal events and then feeding these into graph models. This indicates that the sequential nature and hierarchical information within personal events are crucial and can be lost when treated merely as additional nodes in a graph.
Also Read:
- EVP: Enhancing Dynamic Graph Learning with Event-Aware Prompting
- Human-Centered AI: A New Framework for Designing and Evaluating Explainable Systems
The authors conclude that while their work demonstrates the value of integrating both event types, existing models still have room for improvement. They invite the research community to use their released datasets and code to further explore and develop more sophisticated models that can effectively handle the complexities of unified user event modeling. This research paves the way for more accurate and holistic predictions of user behavior in various applications. You can find the full research paper here: Integrating Sequential and Relational Modeling for User Events: Datasets and Prediction Tasks.
