PeopleRank: What are some existing “PeopleRank” algorithms?

PeopleRank: What are some existing “PeopleRank” algorithms? – Quora.



What are some existing “PeopleRank” algorithms?…

Users on sites like Quora can be represented as collections of text messages (Q&A, comments, status updates, profile info). What are some of the existing text quality ranking/clustering techniques, e.g in spam blocking? Which algorithms take into account the user’s editing trail? 

In addition social endorsements (upvotes/downvotes/number of followers), reputation of user’s ‘nearest neighbors’ and social interaction patterns can be used as validating factors. What are some of the existing ranking algorithms used in social networks / wikis / newsgroups / e-commerce sites that account for these factors? What else constitutes ‘user quality’ or in better words user contribution quality, online reputation and influence? How do you build a user behavior profile?

11 Answers

Scott CountsResearcher, Microsoft Research

Aditya Pal from U Minn and I are presenting a paper at the WSDM 2011 conference on topical authority finding in Twitter. 
Paper here:…
This is conceptually a PeopleRank for topics in Twitter, though it doesn’t work statistically like pagerank. Instead, we probabilisticly cluster people over a set of metrics and then heuristically pick the cluster with the most authoritative users (and then rank within that cluster). This is much less expensive to compute and can avoid your results being dominated by users with heavy network properties (i.e., celebrities who have orders of magnitude more followers, but aren’t necessarily experts)..
The dozen or so features we compute for each user are in the paper, and inlcude some things like ‘do those around you also tweet on this topic’, do you tend to tweet before or after others on this topic’, ‘do you always tweet on this topic or was this a one-off’.Some of these features are really helpful for weeding out spammers – like if your ‘self-similarity’ score is crazy high, you are probably a spammer (and also wouldn’t get clustered with legitimate authors).Overall it seems to work pretty well. There are example results in the paper.
For Twitter, there’s also a paper that is a more straightforward application of pagerank (here:
Scott Counts

Seb PaquetB. Sc. in Physics, Ph.D. in Computer …

The WikiTrust system features a reputation system for wiki users.…

The reputation system is content-driven, as it measures the 
reputation of users on the basis of how their contributions fare, rather
than on the basis of user-to-user feedback.  The main idea is as 

  • Authors whose contributions are preserved in full or in part gain reputation.

    In particular, users gain reputation when their reputations are

    preserved unchanged, but also (in slightly lower measure) when they are

    adapted, reworded, or re-formatted.

  • Authors whose contributions are reverted, or almost entirely undone, lose reputation.

Seb Paquet

F. Randall FarmerCoauthor: Building Web Reputation Sys…

[a nod to the papers already cited here - wikitrust is particularly interesting, but limited in it's application to high-traffic wiki contributions...]

Being a personal score with social and commercial aspects, web Karma (reputation about users) is a complicated issue. Especially if the score is publicly displayed in any form:… which draws from the book I co-authored with Bryce Glass: Building Web Reputation Systems… - draft in wiki form here:…

An example of using (corporate internal) karma to drastically reduce abusive content is covered in chapter 10, which covers work done on Yahoo! Answers:…
F. Randall Farmer

Daniel Gayo-AvelloAssociate professor @ the University …

In Twitter context TunkRank by Daniel Tunkelang (…) is rather well-known.

It has been implemented by Jason Adams (
Daniel Gayo-Avello

There are good but older papers on the topic are by Kleinberg and Adamic, which look at expertise ranking in networks. Two of Adamic’s are listed below.

In the latter they tests various approaches including Kleinberg Hits and PageRank.

FIDE, the international chess ranking system, uses ELO ratings to build reputation ranks. For that, a sufficient number of pairwise comparisons is needed.

For social platforms (Twitter, Facebook, etc.) PeerIndex, my company runs a topic-specific PageRank-like calculation based on sharing activity and some other features. For relevance, we do it on a topic-by-topic basis, extracted by topic-modelling. We refer to this as resonance and it essentially shows the trustworthiness or resonance of a user on a given topic.

As an example visit… 
The column marked “Res.” shows the resonance of people listed in the topic of venture capital. The scores run 1 to 100 and maps a fixed distribution of the PageRank of nodes in the venture-capital network. 

Every user, e.g.…, is associated to his or her resonance on a list of Topics (in little green boxes). We calculate those for several thousands of topics. Where the network is too small, we don’t.

This isn’t exactly the published implementation of PageRank, but it’s close.

Azeem Azhar

Megan BerryMarketing Manager at Klout

Klout measures influence across the social web and we are often compared to PageRank for people. 

In an analogous way to PageRank we are looking at who is connecting to you, engaging with you etc and looking at their influence level. We give users a Klout Score on a normalized scale of 1 to 100, with 100 being the most influential. 

Right now we measure based on Facebook and Twitter, with many more social networks to come. You can learn more about our scoring on
Megan Berry

Mike SellersDesigned and developed the Dynemotion…

Unlike the other PeopleRank-esque algorithms that have been discussed here, we have a unique take on this that models how people create, consider, and transmit reputation, and applies this to the social graph. 

One key insight that we apply is that reputation is not a single number; it’s not something that exists “on” an individual. Rather, everyone else who knows about you has an opinion of you. If you don’t know me, your opinion of me will be a weighted aggregate of all the people you know who do know me. So if you don’t know me but your best friend thinks I’m great, from your POV I have a positive reputation. At the same time, there might be someone else who thinks I’m an idiot, and so for people who trust them I won’t have a good reputation. An easy example is Romeo Montegue: is he a great guy (as other Montegues would say), or a scoundrel (as many Capulets would maintain)? Both are true. His reputation depends on your point of view. This is a reality that’s lacking in virtually all other people-ranking systems.

The great thing about this is that this system allows for both global ratings (“who has the best reputation overall?”) and also much more useful local ratings. You may not care what movie critics think are the best movies, but you likely do want to know what your social network (your friends, theirs, and theirs) thinks. It also allows for opinion-makers: the more people who trust your opinion, the greater reach your opinion has. This all allows each person an individual view of the world, based on the people they trust — and everyone contributes to everyone else’s view, just as in real/offline human social systems. 

This work came out of our overall AI architecture — it started as a way for agents to be able to exchange opinions — and so works for both human-human and human-agent connections. You can read a paper about this from AIIDE 2008 at…
Mike Sellers

Wisdio’s WAR (Wisdio Authority Rating) serves to rank users based upon their activity on Wisdio. Essentially WAR is in place to distinguish experts from people who just claim to be experts. Here’s how the system works:

  1. A user answers a question on Wisdio. The corresponding topic of the question is then listed in a user’s knowledge profile. (List of all areas of knowledge the user has interacted in.)
  2. Based upon user votes, the quality of answer affects how a user ranks amongst other users who have shared knowledge in similar categories. Votes from users who are highly ranked in the corresponding category have more weight than those with lower ranks.
  3. The ranking/scores from all answers are then aggregated to assign users a Wisdio Authority Rating score between 1-100. A perfect 100 is near impossible to receive since no person is truly knowledgeable about everything.

This system is fairly new, but I think it definitely has a great direction. Instead of showing influence of knowledge shared like Klout, Wisdio wants to show the reliability of knowledge shared.

Scott Dzierzanowski

I recently read about an algorithm called SPEAR (…) that was tested over Delicious, to understand which people fell into specific categories, such as:

  • Discoverers
  • Followers
  • Spammers
  • Trolls

The main point of the algorithm was to be spam resistant and to favor quality over quantity.

Dragos Ilinca

Toma BedollaFounder of Veracious Entropy and the …

1 vote by Jared Polivka
The HowTru? project is aimed at solving this problem on a WWW scale; we’ve developed a set of algorithms that consider concepts like the contextual accuracy of previous posts, comments and publications, the referential value of each contribution to the community (locally and globally) and the awareness of users in regard to networks of content. The product of this system is two-fold; a VeracityScore for any piece of content and a series of contextual credibility scores for participants in the system, analogous to the financial credit score that no one applies for but everyone is held accountable to.

This project was first made public in May of last year: and we’ll be launching an alpha test in the coming weeks and releasing a white paper that describes our work before the WICOW conference in March.
Toma Bedolla

Dare ObasanjoWorked on Social Graph and News Feed …

Facebook’s EdgeRank (see…) is probably the world’s most used people ranking algorithm that most people don’t know about.
Dare Obasanjo

Leave a Reply

Your email address will not be published. Required fields are marked *

You may use these HTML tags and attributes: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <strike> <strong>