IRIS: Bridging the Gap in Direct Trust Computation for Social Networks

TL;DR

While many algorithms handle how trust spreads through a network, few address how to measure the initial trust between two friends. IRIS (Interactions, Relationship types, and Interest Similarity) is a novel framework that uses semantic web data (FOAF) to automatically calculate these direct trust values. It achieves a 1%–15% improvement in Fscore over traditional baselines like TidalTrust.

The Missing Link in Trust Modeling

In the world of Online Social Networks (OSN), trust is the currency of interaction. Most research assumes that if Alice knows Bob, a "trust score" already exists. Researchers then focus on whether Alice should trust Dave (who knows Bob).

However, the authors of this paper argue that Direct Trust is rarely discussed. Most models use random values or binary "trusted/untrusted" labels. This lack of granularity makes security policies and recommendation systems unreliable. The motivation behind IRIS is to replace these assumptions with a data-driven computation.

Methodology: The Three Pillars of IRIS

The core of the IRIS method lies in its multi-dimensional approach to evaluating a relationship. It doesn't just look at who you are, but how you act and what you like.

1. The Direct Trust Formula

The Direct Trust ($DT$) between user $v$ and $v^{\prime }$ is a weighted sum: $D{T}_{v\to v^{\prime }}=\alpha f{t}_{v\to v^{\prime }}+\beta i{t}_{v\to v^{\prime }}^n+\gamma s{t}_{v\to v^{\prime }}$ Where $\alpha ,\beta ,\gamma$ are weights (set to $1/3$ in this study).

2. The Relationship Hierarchy (Friendship Trust)

Using the RELATIONSHIP vocabulary, IRIS categorizes links into five levels, assigning trust values accordingly:

• Close Relationships (e.g., ParentOf, SpouseOf): $ft=1.0$
• Friendships (e.g., ColleagueOf, FriendOf): $ft=0.75$
• Acquaintance: $ft=0.5$
• Superficial/Bad Acquaintance: $ft=0.25$ to $0$

Figure 1: A directed labeled graph representing relationship types between users.

3. Behavioral and Interest Synergy

• Interactions ($it$): This accounts for history. If Alice has had 10 interactions with Bob and 9 were positive, Bob gains a high interaction score.
• Similarity ($st$): Based on the Homophily principle ("Birds of a feather flock together"), the model compares interest tags (e.g., "Semantic Web", "Comics") between two users.

Experimental Results

The authors validated IRIS using the foafPub dataset, containing over 200,000 RDF triples. They compared IRIS against several standard trust composition methods (Max-Min, Max-Mean, TidalTrust).

Performance Highlights:

• Fscore Leadership: IRIS reached an 82.0% Fscore, the highest among all tested methods.
• Error Reduction: The Mean Absolute Error for IRIS was consistently lower (~0.098) than competitors, even when the decision threshold shifted.

Figure 4: Fscore performance showing that a threshold of 0.5 is optimal for balancing trust/distrust states.

Table III: IRIS vs. SOTA baselines. IRIS provides a slight edge over Max-Weight and a significant lead over TidalTrust.

Critical Insight: Beyond Binary Trust

The paper's most interesting contribution is the concept of Controversiality Percentage. By analyzing the trust values received by a user, IRIS can categorize users:

• Benevolent: Universally high trust.
• Malicious: Universally low trust (distrust).
• Controversial: Split opinions (some trust, some distrust).

This distinction is vital for modern social platforms plagued by bots and polarizing figures. Identifying a "controversial" user is often more useful than a simple average trust score.

Limitations and Future Work

The study relies on semantic web data (FOAF), which is less prevalent in modern centralized platforms like X (Twitter) or Meta. Future iterations would benefit from:

1. Heterogeneous Data Sources: Integrating real-time API data from modern OSNs.
2. Dynamic Weighting: Allowing $\alpha ,\beta ,\gamma$ to adapt based on the community's context (e.g., interest similarity might matter more in a hobbyist forum than in a family network).

Conclusion

IRIS provides a robust, mathematically grounded way to calculate the "first hop" of trust. By combining who people are (Interests), how they are linked (Relationships), and how they behave (Interactions), it offers a more nuanced and accurate "Trusted Social Network" than previous heuristics.