Exploiting User Interest on Social Media for Aggregating Diverse Data and Predicting Interest

Nozomi Nori, Danushka Bollegala, Mitsuru Ishizuka

Abstract

More and more users have been taking various actions to diverse resources referred to by URLs such as news, web pages, images, products, movies as a result of the growth of social media. They are annotating, tweeting in Twitter, reblogging in Tumblr, and Liking in Facebook, etc. Analyses about these diverse actions will be useful for aggregating or integrating diverse resources. In this paper, we view users’ actions to resources as expressing their some interests, and by investigating how their interests are expressed in social media, we get suggestions for aggregations. Our results show that a certain kind of action (such as tagging on Delicious) can be used to make predictions on a different kind of action (such as favorite on Twitter). These analyses will be useful for aggregating or integrating diverse contents on multiple sources. In addition to some experimental analyses, we propose a novel method to predict users’ interests in social media, using time-evolving, multinomial relational data. Our experimental results show that the proposed method significantly outperforms standard tensor analysis and an existing state-of-the-art method (LDA) in prediction tasks.

Introduction

In the modern web ecosystem, user activity is fragmented across multiple platforms. A single individual might "Like" a product on Facebook, "Retweet" a news link on Twitter, and "Tag" a technical article on Delicious. This fragmentation presents significant hurdles for recommender systems, specifically regarding data aggregation, the cold-start problem for new users, and the need to capture shifting temporal dynamics in user interests.

Conceptual framework of user interests and behaviors toward resources.

The researchers aim to bridge these gaps by treating every social media action as a signal of interest. By identifying "hidden bridges" between different types of behavior across platforms, they propose a way to create a coherent profile of user interest from diverse, sparse data sources.

Methodology: The ActionGraph Model

The primary contribution of this work is ActionGraph, a bipartite graph representation designed to handle complex, multi-entity relationships. Unlike traditional models that rely on simple user-item pairs, ActionGraph utilizes:

  • Action Nodes: Representing a specific event at a specific time.
  • Object Nodes: Representing the user, the URL, the action type (e.g., "retweet"), or keywords.
  • Temporal Preservation: By assigning each action its own node, the model naturally preserves the temporal sequence of a user’s behavior without requiring a universal global timeline.
  • Regularized Laplacian Kernel: This graph kernel calculates similarity between nodes by considering indirect paths. For instance, if two users frequently perform the same action type, that shared behavior acts as a bridge, allowing the model to infer shared interests in specific resources even without direct interaction history.

Experimental Setup

The study utilized a custom-crawled dataset linking two major platforms: Twitter (tweets, retweets, favorites, following) and Delicious (social bookmarking/tagging). Users were linked across services using FriendFeed. The dataset covered one month (August 2010) and included over 200,000 tweets and 21,000 tagging instances.

Results and Discussion

A key finding of the research is the Cross-Domain Asymmetry in interest prediction. The experiments demonstrated that certain actions are better predictors of others. Specifically, tagging behavior on Delicious proved highly effective at predicting various actions on Twitter.

Precision of predicting various test actions based on training with tagging (TAG) actions.

As shown in the figure above, while tagging (TAG) is the best predictor for future tagging, it also maintains significant predictive power for other actions like following (Fol) and favoriting (Fav). The authors hypothesize that this is due to the rich intent captured by keywords in tags, which "low-signal" actions like retweets lack.

In comparative evaluations, ActionGraph significantly outperformed traditional methods:

  • Precision: ActionGraph achieved an R-Precision of 7.6%, nearly doubling the performance of PARAFAC (3.4%) and beating LDA (4.3%).
  • Coverage: ActionGraph achieved 99.8% coverage, whereas PARAFAC only managed 43.6%, demonstrating the graph model's robustness against the sparse nature of social media data.

Conclusion

ActionGraph provides a powerful framework for aggregating fragmented social media data to predict user interests. While the current study highlights the importance of keyword-rich actions like tagging, future work could further integrate social network structures (follower graphs) and explore the non-linear relationship between data volume and predictive performance.

Summary updated

Abstract

More and more users have been taking various actions to diverse resources referred to by URLs such as news, web pages, images, products, movies as a result of the growth of social media. They are annotating, tweeting in Twitter, reblogging in Tumblr, and Liking in Facebook, etc. Analyses about these diverse actions will be useful for aggregating or integrating diverse resources. In this paper, we view users’ actions to resources as expressing their some interests, and by investigating how their interests are expressed in social media, we get suggestions for aggregations. Our results show that a certain kind of action (such as tagging on Delicious) can be used to make predictions on a different kind of action (such as favorite on Twitter). These analyses will be useful for aggregating or integrating diverse contents on multiple sources. In addition to some experimental analyses, we propose a novel method to predict users’ interests in social media, using time-evolving, multinomial relational data. Our experimental results show that the proposed method significantly outperforms standard tensor analysis and an existing state-of-the-art method (LDA) in prediction tasks.
BibTeX
@inproceedings{nori2021exploiting,
  author = {Nozomi Nori and Danushka Bollegala and Mitsuru Ishizuka},
  title = {{Exploiting User Interest on Social Media for Aggregating Diverse Data and Predicting Interest}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/icwsm.v5i1.14114},
}
Figure