Abstract
In the era of information overload, human attention serves as a critical bottleneck for content consumption on social media. Traditional recommendation systems often overlook this psychological constraint, assuming users have infinite capacity to process content from their social circles. This paper introduces LA-CTR (Limited Attention Collaborative Topic Regression), a model that explicitly incorporates attention allocation across friends and topics. By modeling how users non-uniformly divide their finite attention, LA-CTR significantly outperforms state-of-the-art collaborative filtering and topic modeling hybrids in predicting user engagement.
Introduction
Social media platforms act as massive information filters, yet the ultimate gatekeeper is the user's own limited attention. We do not engage with every post from every connection; instead, we subconsciously prioritize specific "influential" friends for particular topics. Jeon-Hyung Kang and Kristina Lerman argue that modeling this "social filter" is essential for accurate recommendations. Their proposed model, LA-CTR, treats attention not as a byproduct of interest, but as a primary constraint that shapes social discovery.
The LA-CTR Framework
Traditional models like Collaborative Topic Regression (CTR) combine user behavior with content analysis but fail to account for the social-psychological layer of attention allocation. LA-CTR addresses this by introducing three key latent variables:
- Interests ($u$): The user's general topical preferences.
- Influence ($s$): The baseline weight a user assigns to specific friends.
- Attention ($\phi$): A fine-grained, topic-specific mapping of how much attention a user pays to a friend regarding a particular subject.
The model leverages Latent Dirichlet Allocation (LDA) to extract topics from content and uses an Expectation-Maximization (EM) style algorithm to learn the "social filter." By calculating the probability of adoption based on the attention paid to the recommender (the friend) for a specific topic, LA-CTR captures the reality that a user might trust one friend for technology news but ignore their political commentary.
Experimental Evaluation
The researchers evaluated LA-CTR using two large-scale datasets from Digg, a social news aggregator:
- Digg 2009: 70,000 users, 3,500 stories, 2.1 million votes.
- Digg 2010: 12,000 users, 48,000 stories, 1.9 million votes.
To isolate "social discovery," the study focused on votes cast before stories reached the front page, ensuring that engagement was driven by social connections rather than global trends.
Key Results and Insights
LA-CTR was compared against standard CTR and CTR-smf (which includes social matrix factorization). The results demonstrated several key advantages:
- Superior Predictive Accuracy: LA-CTR consistently achieved higher Recall@X scores compared to baseline models.
- Information Efficiency: LA-CTR with only 50 topics outperformed baseline models using 200 topics, suggesting that psychological grounding makes the model more robust.
- The Role of Divergence: The model includes a parameter ($\lambda_\phi$) to balance personal interest with social influence. Optimal performance was found when the model allowed for a degree of divergence between a user's long-term interests and the specific topics they attend to via their social network.
Conclusion and Future Work
By modeling the "science of limited attention," LA-CTR provides a more interpretable and accurate framework for social media recommendations. It successfully distinguishes between what a user likes (interest) and who they listen to (attention). Future work aims to address the computational complexity ($N^2D$) of the model to enable deployment on massive-scale networks and to validate the learned "attention" parameters against actual eye-tracking or click-stream data.
This summary is based on the paper "LA-CTR: A Limited Attention Collaborative Topic Regression for Social Media" by Jeon-Hyung Kang and Kristina Lerman.
Summary updated