Social Ranking: Beyond PageRank for the Spoken Web

TL;DR

In the developing world, the "Spoken Web" allows illiterate or low-literacy users to create and browse audio content via basic mobile phones. Since there are no "hyperlinks" between audio files, traditional PageRank fails. This paper proposes a Social Ranking algorithm that uses real-world relationships and community trust—mined from browsing logs—to ensure that search results are socially relevant to the listener.

Background: The Web of Speech

For billions of people in rural, developing regions, the Internet isn't a screen; it's a voice on a phone. The Spoken Web enables users to host "VoiceSites." However, searching audio is a nightmare. Because audio is sequential, a user can't "skim" 10 results in seconds. If the first two results aren't perfect, the user hangs up.

The authors argue that in these close-knit communities, who said something is often more important than what was said.

The Problem: Why Links Don't Work

The World Wide Web's success was built on in-links (PageRank). But in a VoiceSite ecosystem:

1. No Link Structure: Audio files don't naturally link to one another.
2. Local Relevance: A farmer in Gujarat cares more about a neighbor's agricultural tip than a generic expert from a city.
3. Sequential Constraint: The "Top 10" UI of Google is a "Top 1" or "Top 2" UI in speech.

Methodology: Quantifying Human Relationships

The authors break down "Social Rank" (S) into two key components: Relationship ($R$) and Trust ($T$).

1. The Relationship Function ($R_{q,c}$)

This measures how close the searcher (query-creator) is to the person who recorded the audio (content-creator). It combines:

• Following Score ($f_f$): Does the searcher frequently listen to this specific creator?
• Similarity Score ($f_s$): Do the searcher and the creator listen to the same types of content?

2. The Community Trust Factor ($T_c$)

Independently of the searcher, how much does the whole community listen to this creator? This identifies "local celebrities" or "trusted experts."

Figure 1: Comparison between traditional PageRank (link-based) and Social Rank (relationship-based).

The Final Formula

The ranking is a weighted sum: $S=\gamma R_{q,c}+(1-\gamma )T_c$ This allows the system to balance "Personalized Interest" (what I like) with "General Authority" (what everyone trusts).

Experimental Validation: Field Study in Gujarat

The researchers tested their algorithm on 4,827 audio documents from a live VoiceSite used by 250 grain farmers in India. They interviewed 15 farmers to map their "real-world" social networks and then checked if the algorithm could recreate that map using only log data.

Key Findings:

• Accuracy: In most cases, the algorithm’s "Top 5" relationships matched the people the farmers specifically named in person.
• Authority Identification: The trust score ($T$) successfully identified the regional experts that the community relied upon.

Figure 2: The R-score (lines) effectively captures the actual people (dots) farmers identified as important.

Future Outlook and Limitations

While the social ranking shows immense potential for localized, illiterate communities, the authors note a few hurdles:

• Cold Start: If a user is new and hasn't listened to anything, $R_{q,c}$ is hard to calculate.
• Scalability: While it works for a community of 250, how does social trust scale to millions?
• Evaluation: The authors admit they still need to compare this against a traditional baseline in a controlled A/B search task.

Conclusion

This work highlights a critical shift in Information Retrieval (IR): as we move toward "Web 2.0" and "Spoken Web" environments, Social Context becomes the new PageRank. For developers building for the "Next Billion Users," the lesson is clear: mine the social interactions, not just the keywords.