[Springer 2017] TempoRec: Solving User Interest Drift in Social Networks via Temporal-Topic Hybrid Recommendation

TL;DR

Social network interests are rarely static—a football fan during the World Cup might pivot to tech news a month later. TempoRec addresses this "interest drift" by combining a socially-aware matrix factorization model with a temporal-topic extraction module. By weighting recent interests more heavily and incorporating demographic data, it achieves superior friend recommendation accuracy on real-world Weibo data.

Problem & Motivation: The Static Interest Trap

Most recommendation engines treat your profile as a frozen snapshot. However, in platforms like Weibo or Twitter, two major hurdles exist:

1. Dynamic Shifts: Interests change. Static models fail to distinguish between deep-seated hobbies and fleeting seasonal trends.
2. Limited Reach: Social-only recommendations (like "Friends of Friends") often trap users in an echo chamber of offline acquaintances, failing to find new like-minded strangers.

The authors' insight was to create a "hybrid" logic: use Social Links to establish trust and Temporal Topics to ensure relevance to a user's current state of mind.

Methodology: The Core Modules

1. Social Relation Module (SR-FR)

The system doesn't just look at who you follow. it maps users into a 44-dimensional vector based on Gender and Age, groups them, and applies Matrix Factorization. A key innovation here is the weighting of social influence: if User A and User B share many friends, the influence of their latent preferences on one another is increased proportionally.

2. Temporal-Topic Module (TLDA-FR)

To track shifting interests, the authors:

• Segment Content: Weibo posts are divided into time windows ($\delta$).
• Extract Topics: LDA identifies latent topics within these windows.
• Calculate Similarity: Instead of simple Cosine similarity, they use Jensen-Shannon (JS) Divergence to measure the distance between user-topic distributions.
• Apply Time Decay: A decay function $f(t)=\exp (-\frac{n-t}{\gamma })$ ensures that interests from 6 months ago have less weight than interests from last week.

Experiments & Results

The model was tested using Sina Weibo data spanning 184 days.

Key Findings:

• Optimal Fusion: The best results occurred when Social Relations were weighted at 0.4 and Temporal Topics at 0.6, suggesting that shared interests are slightly more predictive of new follows than demographics alone.
• Window Size: A 14-day time window ($\delta =14$) was the "sweet spot" for capturing interest shifts without introducing too much noise.
• Comparison: The hybrid SIF-FR significantly outperformed CplusL (a static content model) and ICACF-FR (which suffers from data sparsity).

Fig 7: SIF-FR displays a clear advantage in Mean Average Precision (MAP) across all k-values compared to single-module approaches.

Deep Insight & Conclusion

Takeaway

TempoRec proves that "Recency" is a feature, not just a timestamp. By integrating a decay function directly into the topic-modeling pipeline, the system can "forget" outdated versions of a user and focus on their evolving persona.

Limitations

• Complexity: The time complexity of constructing large-scale social graphs and running LDA across multiple time slices is high.
• Data Sparsity: As shown in Figure 7, MAP drops as the recommendation list grows, mostly because the underlying "friend density" in the dataset is only 3%.

Future Outlook

The authors suggest that future iterations should incorporate Sentiment Analysis (how a user feels about a topic, not just that they mentioned it) and Geographic Data to further refine the "context" of a recommendation.