Time-Aware Indexing: Predicting the Future of Social Links via Temporal Decay

TL;DR

Researchers have developed Time Score (TS), a new feature for link prediction that doesn't just look at who you know, but when you last interacted with them. By applying a temporal damping factor and measuring the synchronicity of shared connections, this method outperforms standard structural benchmarks in Facebook and Coauthorship networks, with Precision boosts up to 14%.

The Problem: Social Networks are Not Static

Most classical link prediction algorithms treat social graphs as static snapshots. Features like Adamic/Adar or Common Neighbors assume that a collaboration from 1990 carries the same weight as one from 2024.

The Reality: Links have a "shelf life." In professional networks, researchers change fields; in Facebook, friendships fade as users move or graduate. Previous SOTA methods ignored the "gap" between the current time and the last interaction, leading to a "zombie link" problem where the model predicts connections based on obsolete social circles.

Methodology: The Anatomy of Time Score (TS)

The authors propose that the strength of a potential link is a function of Recency and Proximity. Instead of simply counting common neighbors, the Time Score (TS) calculates a weighted value for each shared node:

1. Temporal Damping (${\beta }^k$): Uses a decay factor $\beta$ raised to the power of the time gap $k$. If you haven't interacted recently, the score drops exponentially.
2. Interaction Proximity ($|t_1-t_2|+1$): This unique insight suggests that if node A and node B both interacted with common neighbor C at roughly the same time, they are much more likely to link up than if their interactions were years apart.
3. Harmonic Mean ($H_m$): Captures the intensity/frequency of those interactions.

Mathematically:

$TS={\sum }_n\frac{H_m\cdot {\beta }^k}{|t_1-t_2|+1}$

Figure 1: Conceptual illustration of common neighbor interactions and their respective time stamps.

Experiments: Proving the Temporal Advantage

The authors tested their index using a supervised learning approach (J48 Decision Trees) on two distinct datasets:

• Facebook New Orleans: Tracking wall posts and friendship establishments.
• Cond-mat Coauthorship: Tracking scientific collaborations over 13 years.

Performance Breakthroughs

The "Time Score" was integrated into a feature vector alongside standard metrics like Jaccard's Coefficient and Preferential Attachment.

Figure 2: Performance metrics comparing models with and without the Time Score index.

• Facebook Findings: The TS index was particularly effective during periods of "rapid growth" (June 2008 – Jan 2009). When the network is highly active, temporal recency becomes the dominant predictor of new friendships.
• Coauthorship Findings: The improvement was even more stark here. Scientists are highly sensitive to current trends; hence, collaborations are driven by the most recent shared contacts rather than distant historical ones.

Critical Analysis & Takeaways

The brilliance of Time Score lies in its simplicity—it bridges the gap between graph topology and temporal dynamics without requiring heavy Recurrent Neural Networks (RNNs).

Limitations:

• Beta ($\beta$) Sensitivity: The damping factor $\beta$ must be tuned manually for different networks (0.85 for Facebook vs. 0.5 for Physics papers).
• Data Sparsity: Like most link prediction tasks, the "class imbalance" problem (too many potential links, too few actual ones) remains a challenge, requiring SMOTE oversampling to achieve stability.

Future Outlook: This work paves the way for "Time-Aware" recommendation engines. By focusing on the velocity and timing of social interactions, platforms can move beyond static "People You May Know" lists to dynamic, contextually relevant suggestions that reflect a user's current social trajectory.