Facebook Fans Influence: Mining Time-Dependent Power using FP-Growth

TL;DR

In the world of social media marketing, a high Klout score isn't everything. This paper presents a framework to find Time-Dependent Influential Users—individuals who may not be general celebrities but hold significant power to trigger information cascades during specific intervals. By combining time-segmentation with Frequent Pattern Mining (FP-growth), the authors provide a roadmap for businesses to time their product promotions for maximum impact.

Background: The Limits of Static Influence

Most businesses today rely on aggregate metrics to identify influencers. However, social influence is rarely static. A user might be highly influential on Monday mornings regarding coffee but invisible on Friday nights. Previous methods fail because they ignore the temporal dimension, treating influence as a stationary attribute rather than a dynamic variable.

Methodology: From Clicks to Frequent Patterns

The proposed framework operates through a rigorous five-step pipeline:

1. Time Segmentation: Dividing a long duration (e.g., a month) into discrete segments $T=\{t_1,t_2,\dots ,t_n\}$ (e.g., days).
2. Influence Quantification: Calculating an Influence Score $S(n_j)$ for each user based on feedback loops: $S(n_j)=w_1\cdot extthumbs\_up(n_j)+w_2\cdot extshares(n_j)$
3. Discretization: Converting continuous scores into levels: Low (0-30), Medium (30-80), and High (80+).
4. Transaction Modeling: Each time segment is treated as a "transaction" containing participating fans and their influence levels.
5. FP-Growth Mining: Applying the FP-growth algorithm to find frequent influence patterns (L-frequent patterns) across the reduced dataset.

Figure 1: Conceptual visualization of successive time segments for influence tracking.

Why FP-Growth?

The choice of FP-growth over the classic Apriori algorithm is strategic. In social network data, where the number of "items" (users) can be in the thousands, Apriori's repeated database scans and candidate generation become a bottleneck. FP-growth uses a compressed tree structure (FP-tree), scanning the database only twice, which is significantly more efficient for high-dimensional social transactions.

Experimental Insights

Through a theoretical experiment across four weeks of data, the authors identified users who exhibit "bursty" influence. For example, while some users appeared frequently across all segments, a specific user, n8, was identified as a key time-dependent influencer who peaked only during specific segments like $t_1$ and $t_6$.

Table 1: Discretized influence transactions used for pattern mining.

Key Takeaways from the Data:

• Precision Timing: Businesses can now identify the exact time segments where specific users are likely to trigger a purchase cascade.
• Hidden Gems: Users with low global influence (Klout) but high "segment influence" can be recruited as cost-effective micro-influencers.

Critical Analysis & Conclusion

This work successfully bridges the gap between frequent itemset mining and temporal social analysis.

Strengths:

• Intuitive and computationally efficient.
• Directly applicable to real-world marketing automation.

Limitations:

• The study uses a simplified weight system ($w_1=w_2=1$); in reality, a "Share" is often exponentially more valuable than a "Like."
• The threshold for "Low" vs "High" influence is currently arbitrary and needs to be tuned per industry.

Future Outlook: The integration of sentiment analysis (judging if comments are positive or negative) and social graph topology (who follows whom) would further refine this model, turning it from a pattern miner into a predictive influence engine.