Paul Resnick's Occasional Musings

moved to presnick.wordpress.com

Sat, 17 Dec 2011 16:29:23 GMT

This blog has now moved to presnick.wordpress.com, where spam comments seem to be handled a little better.

I apologize to all of my past real commenters, whose comments are now hidden. The spam comments had overwhelmed your genuine contributions.

Personalized Filters Yes; Bubbles No

Sun, 17 Jul 2011 18:32:42 GMT

On Thursday, I gave the closing keynote for the UMAP conference (User Modeling and Personalization) in Girona, Spain (Catalunya). I had planned to talk about by work directed towards creating balanced news aggregators that people will prefer to use over unbalanced ones (see project site). But then Eli Pariser’s TED Talk and book on “Filter Bubbles” starting getting a lot of attention. He’s started a trend of a little parlor game where a group of friends all try the same search on Google and look in horror when they see that they get different results. So I decided to broaden my focus a little beyond news aggregators. I titled the talk, “Personalized Filters Yes; Bubbles No.”

As you can perhaps guess from my title, I agree with some of Pariser’s concerns about bubbles. But I think he’s on the wrong track in attributing those concerns to personalization. Most of his concerns, I argue, come from badly implemented personalization, not from personalization itself. I’ve posted a copy of my slides and notes. For anyone who wants a summary of my arguments, here goes.

His first concern I summarize as “Trapped in the Old You”. I argued that personalization systems that try to maximize your long-term clickthrough rates will naturally try to explore to see if you like things, not just give you more of the same. This is the whole point of the algorithmic work on multi-armed bandit models, for example. Moreover, once our personalization systems take into account that our interests and tastes may change over time, and that there is declining marginal utility, eventually, for more of the same (consider the 7,000^th episode of Star Trek; even I stopped watching). Personalization systems that are designed to optimize some measure of user satisfaction (such as click-throughs or purchases or dwell time or ratings) are going to be designed to give you serendipitous experiences, introducing you to things you like that you didn’t know you would like. Moreover, even today’s systems often do that pretty well, in part because when they optimize on matching on one dimension (say topic) they end up giving us some diversity in another dimension that matters to us (say, political ideology). From introspection, I think most people can recall times when automated personalization systems did introduce them to something new that became a favorite.

His second concern I summarize as, “Reinforcing Your Baser Instincts”. Here, too, good personalization systems should take into account the difference between short-term and long-term preferences (entertainment vs. education, for example). We will need delayed indicators of long-term value, such as measuring which words from articles we end up using in our own messages and blog posts, or explicit user feedback after some delay (the next day or week). It may also be helpful to offer features that people can opt in to that nudge them toward their better selves (their long-term preferences). Here, I gave examples of nudges toward balanced news reading that you can see if you look at the slides.

His third concern I summarize as, “Fragmenting Society”, but there are two somewhat separable sub-elements of this. One is the need for common reference points, so that we can have something to discuss with colleagues at the water cooler or strangers on the subway. Here, I think if individuals value having these common reference points, then it will get baked into the personalized information streams they consume in a natural way. That is, they’ll click on some popular things that aren’t inherently interesting to them, and the automated personalization algorithms will infer that they have some interest in those things. Perhaps better would be for personalization algorithms to try to learn a model that assumes individual utility is a combination of personal match and wanting what everyone else is getting, with the systems learning the right mix of the two for the individual, or the individual actually getting a slider bar to control the mix.

The second sub-concern is fragmenting of the global village into polarized tribes. Here it’s an open question whether personalization will lead to such polarization. It hinges on whether the network fractures into cliques with very little overlap or permeability. But the small-world properties of random graphs suggest that even a few brokers, or a little brokering by a lot of people, may be enough to keep average shortest path short. Individual preferences would have to be strongly in favor of insularity within groups in order to get an outcome of real fragmentation. It turns out that people’s preferences with respect to political information, as concluded by my former doctoral student Kelly Garrett, is that they like confirmatory information but have at best a mild aversion to challenge. Moreover, some people prefer a mix of challenging and confirmatory information and everyone wants challenging information sometimes (like when they know they’re going to have to defend their position at an upcoming family gathering.) Thus, it’s not clear that personalization is going to lead us to political fragmentation, or any other kind. Other forces in society may or may not be doing that, but probably not personalization. Despite that, I do think that it’s a good idea to include perspective-taking features in our personalization interfaces, features that make it easy to see what other people are seeing. My slides include a nice example of this from the ConsiderIt work of Travis Kriplean, a PhD student at the University of Washington.

The final point I’d like to bring up is that personalization broadens the set of things that are seen by *someone*. That means that more things have a chance to get spread virally, and eventually reach a broader audience than would be possible if everyone saw the same set of things. Instead of being horrified by the parlor game showing that we get different search results than our friends do, we should delight in the possibility that our friends will be able to tell us something different.

Overall, we should be pushing for better personalization, and transparent personalization, not concluding that personalization per se is a bad thing.

At the conference banquet the night before my talk, attendees from different countries were invited to find their compatriots and choose a song to sing for everyone else. (The five Americans sang, “This Land is Your Land”). Inspired by that, I decided to compose a song we could all sing together to close the talk and the conference, and which would reinforce some themes of my talk. The conference venue was a converted church, and the acoustics were great. Many people sang along. The melody is “Twinkle, Twinkle, Little Star”.

(Update: someone captured it on video: )

The Better Personalization Anthem

User models set me free
as you build the Daily Me

Yes exploit, but please explore
could just be that I’ll want more

Broaden what my models know
UMAP scholars make it so

Words: Paul Resnick and Joe Konstan
Melody: Wolfgang Amadeus Mozart

Yelp gets more reviews per reviewer than CitySearch or Yahoo Local

Wed, 29 Jun 2011 19:04:42 GMT

Author attributes it to the fact that reviewers are anonymous at CitySearch and Yahoo local, but build up reputations on Yelp. Of course, there are also other differences between the sites.

Zhongmin Wang (2010) “Anonymity, Social Image, and the Competition for Volunteers: A Case
Study of the Online Market for Reviews,” The B.E. Journal of Economic Analysis & Policy: Vol.
10: Iss. 1 (Contributions), Article 44.
Available at: http://www.bepress.com/bejeap/vol10/iss1/art44

Abstract:
This paper takes a ﬁrst step toward understanding the working of the online market for re-
views. Most online review ﬁrms rely on unpaid volunteers to write reviews. Can a for-proﬁt
online review ﬁrm attract productive volunteer reviewers, limit the number of ranting or raving
reviewers, and marginalize fake reviewers? This paper sheds light on this issue by studying re-
viewer productivity and restaurant ratings at Yelp, where reviewers are encouraged to establish a
social image, and two competing websites, where reviewers are completely anonymous. Using a
dataset of nearly half a million reviewer accounts, we ﬁnd that the number (proportion) of proliﬁc
reviewers on Yelp is an order of magnitude larger than that on either competing site, more produc-
tive reviewers on all three websites are less likely to give an extreme rating, and restaurant ratings
on Yelp tend to be much less extreme than those on either competing site.

Need recommender systems contest ideas

Mon, 13 Sep 2010 21:32:44 GMT

Do you have an idea or plan for a future challenge/contest that you think could move the field of Recommender Systems forward? I’d love to hear about your idea or plan, even if only in sketch form, and even if you’re not in a position to carry it out yourself. At this year’s RecSys conference in Barcelona, I’ll be moderating a panel titled, “Contests: Way Forward or Detour?” As part of that panel, I’d like to present brief sketches of several contest ideas for the panelists to respond to.

Please send me your ideas!

----------------------Abstract of the Session
Panelists:
Joseph A. Konstan, University of Minnesota, USA
Andreas Hotho, University of Würzburg, Germany
Jesus Pindado, Strands, Inc., USA

Contests and challenges have energized researchers and focused attention in many fields recently, including recommender systems. At the 2008 RecSys conference, winners were announced for a contest proposing new startup companies. The 2009 conference featured a panel reflecting on the then recently completed Netflix challenge.

Would additional contests help move the field of recommender systems forward? Or would they just draw attention from the most important problems to problems that are most easily formulated as contests? If contests would be useful, what should the tasks be and how should performance be evaluated? The panel will begin with short presentations by the panelists. Following that, the panelists will respond to brief sketches of possible new contests. In addition to prediction and ranking tasks, tasks might include making creative use of the outputs of a fixed recommender engine, or eliciting inputs for a recommender engine.

Gerhard Fischer paper at C&T

Fri, 26 Jun 2009 14:57:01 GMT

"Towards an Analytic Framework for Understanding and Supporting Peer-Support Communities in Using and Evolving Software Products" at C&T

Participation in SAP's online community.

Before/After point system introduced in SAP:
Mean response time decreased (51 min vs. 34 min.)
Mean helper count increased (1.89 vs. 2.02)
Percentage answered (12% vs. 30%)

Some evidence of gaming of the system—people just ask questions to gain points.

Karim Lakhani at C&T

Fri, 26 Jun 2009 13:40:10 GMT

Karim Lakhani is giving a great keynote at C&T about tracking innovation. He has worked with MatLab programming contests that have a fascinating format. There's clear performance outcome; source code of all entries is available to other people; leaders are tracked. Researchers can track which lines of code get reused.

What leads to displacing the currently leading entry?
--novel code
--novel combos of others' code
--NOT borrowed code
--complexity
--NOT conformance

What leads code to get reused in future (leading) entries?
--novel code
--novel combos of others' code
--borrowed code
--complexity
--conformance

Also did experiments with TopCoder.
One experiment with computational biology contest problem.
Three conditions (random assignment?):
Fully collaborative vs. fully competitive vs. mixed (competitive first week, then all code shared)
Fully collaborative got the best performance
Best performing entries did better than state-of-the-art in computational biology

Universally Utility-Maximizing Privacy Mechanisms

Sat, 13 Jun 2009 09:51:17 GMT

Interesting paper presentation by Tim Roughgarden.

He gave a nice introduction to the recent literature on provably privacy-preserving mechanisms for publishing statistical summaries such as counts of rows from databases satisfying some property (e.g., income > 100,000). Suppose a mechanism computes the actual count, and then reports something possibly different (e.g., by adding noise). There is a definition of a p-privacy if, for every possible output (count), for any person (row) the ratio of the probability of that output with the row present to the probability of that output with the row omitted is always in the range [p, 1/p]. Intuitively, whatever the actual count, there's not much revealed about whether any particular person has high income.

One technique that works for counts, LaPlace-p, is to add to the correct count +/- z, where probability of z is 1/2(-lnp)e^^(z/lnp). For any reported count, there's some confidence interval around it, and the size of that confidence interval is independent of the count. Thus, for reported count 1, you can't really tell whether the correct count is 1 or 0, and thus you can't really tell whether a particular person has high income, *even if you have great side information about everyone else in the database*. On the other hand, if the reported count is 27,000, you still can't tell much about any one person, but you can be pretty sure that the correct count is somewhere around 27,000.

Roughgarden's paper is about how much value you can get from the best count function (in terms of some loss function comparing true result to reported result) while still preserving the p-privacy requirement. It turns out that a mechanism very much like LaPlace-p, but discretized, works to minimize the expected loss no matter what the user of the count's priors are about the distribution of correct counts. It is in this sense that it is universal. This requires a little user-specific post-processing of the algorithm's output, based on the user's priors about the correct counts. For example, if the reported count is -1, we know that's not the correct count; it must really be 0 or something positive, and you can back out from the report and the user's prior beliefs to infer a belief distribution over correct counts.

Babaioff; Characterizing Truthful Multi-Armed Bandit Mechanisms

Fri, 12 Jun 2009 14:46:26 GMT

At Economics of Search conference.

Moshe Babaioff presented an interesting paper.

Suppose that you're conducting an auction for adwords, where you want to rank the bidders based on expected revenue in order to allocate slots and determine prices for slots based on bids. But suppose you don't know what the clickthrough rate will be for the items.

In a multi-armed bandit model, there are multiple bandit slot machines and you have to decide which arms to pull. There is an explore/exploit tradeoff-- you need to explore (experiment) to estimate the clickthrough rates, including some experimentation with those you have low estimates for, in case that estimate is wrong. But over time you switch to more exploitation, where you pull the arm of the highest expected value.

The new twist in this paper is that you want advertisers to truthfully reveal their valuation for a click. If clickthrough rates are known, you can set price essentially using a second-price mechanism based on bid*clickthrough. But if you're using a multi-armed bandit algorithm to determined clickthrough rates, the correct prices would depend on estimated clickthrough rates that you don't necessarily see because you don't test them.

It's a theory paper. They prove that, with the requirement that the mechanism induce trruthful bidding, there's always a pure exploration phase, where the selection of the winners can depend on previous clickthroughs but *not* on the bids; and then a pure exploitation phase, where the clickthroughs no longer affect allocation of slots in the next round. The best multi-armed bandit algorithms without the truth-telling requirement don't have that separation of phases. And, it turns out that the best algorithms without the truth-telling requirement have less "regret" relative to the best you could do if you magically knew the clickthrough rates at the beginning.

So now I'm curious what the best algorithms are without the truth-telling requirement. My guess is that they put more exploration into things that the best estimate so far has higher value for. We actually need to use an algorithm like this for the next version of our "converation double pivots" work on drupal.org, where we're going to dynamically change the set of recommended items based on a combination of prior generated from recommender algorithms and actual observed clicks. But we don't have any truthful revelation requirement, so we should be able to use the standard algorithms.

Rewards program for social networking activity

Tue, 27 May 2008 18:38:27 GMT

As part of the CommunityLab project, for the past five years I've been doing research related to incentives for participation in online communities. Now one of my colleagues, Yan Chen, is working with a startup company, urTurn, that has created a cross-platform rewards program. That is, you accumulate points for posting photos or making friend links in social network sites like Facebook and MySpace. Then you turn in the points for prizes.

I'm not quite sure what their business model will be (what do they get from having people accumulate points on their site)? But it will be interesting to see how motivating the points are for people, and how they will prevent various attempts to game the system.

So, sign up, help Yan with her research (she has no financial stake in the company), and win valuable prizes!

How newsgroups refer to NetScan data

Fri, 29 Jun 2007 18:53:29 GMT

Reflections and Reactions to Social Accounting Meta-Data. Eric Gleave (U of Washington) and Marc Smith (Microsoft Research). At C&T.

In 18 months, there were about 5000 messages that explicitly referred to "netscan.research". Analyzed/coded 952 messages.

Basic findings:

Half discuss groups. 80% of those linking to the Netscan report card for the group, 17% explicitly discuss the group's "health".

22% discuss the message's author, such as saying that the author is #1 in the group.

31% discuss others, including their stats; 5% of these are "troll checks"

48% discuss the Netscan system itself

Some discussion points:

Helpful for comparisons between competing groups on similar topics

Reduces costs of monitoring and sanctioning

Facilitates construction and maintenance of status

Identifies people who are trolls

Rhythms of social interaction at Facebook

Fri, 29 Jun 2007 18:18:37 GMT

Rhythms of social interaction: messaging within a massive online network. Scott A. Golder, Dennis M. Wilkinson and Bernardo A. Huberman (HP labs).

Scott Golder presenting at C&T. Log analysis of Facebook messaging patterns, from 496 North American universities.

The college weekend goes Friday noon to Sunday noon. Message traffic follows the same pattern Mon-Thurs. Friday morning is same as Mon-Thurs. morning. Sunday afternoon/evening is same as Mon-Thurs. Saturday all day, plus Friday PM and Sunday AM, have much lower traffic.

45% of messages and pokes went to people at different schools. However, this percentage was much lower in the late night/early morning hours.

Perhaps the most surprising result is the seasonal variation in the percentage of messages that are within versus between schools. During vacations, the percentage of within-school messages increases! The authors give the plausible explanation that the messaging is substituting for in-person communication between the same people that would occur when school is in session. This seems surprising to me, however, as I would have thought that the complementarity effect would be stronger-- you send a poke or message to someone that you saw earlier today or expect to see later today. It would be interesting to see some future research that explores more directly the complementarity/substitution effects of various communication modalities with f2f meetings in everyday use.

Group Formation in Large Social Networks

Thu, 28 Jun 2007 21:01:11 GMT

L. Backstrom, D. Huttenlocher, J. Kleinberg and X. Lan. "Group Formation in Large Social Networks: Membership, Growth, and Evolution", Proceedings of KDD 2006.

Datasets on membership in LiveJournal groups and explicit "friend" relationships; and on publishing in conferences and explicit citations between authors.

Question 1: How does the probability of joining a group depend on the friends who are already in it?
A: 'The data suggest a “law of diminishing returns” at work, where having additional friends in a group has successively smaller effect but nonetheless continues to increase the chance of joining...' But if a greater percentage of the friends are linked to each other, the probability of joining is even higher. They suggest that a "strength of weak ties" argument would suggest the opposite of this finding (you're more likely to find out new info from weak ties who don't know each other). But I think decisions about joining require much more than just finding out about the community. (See next blog entry on what makes people commit to/stay in a community.)

Question 2: Which communities will grow over time?
A: Here the characteristics provide a little less predictive power. One obvious one, given the result above, is if there are a lot of people who have a lot of friends in the group, then the group will have larger growth in the next time period. Somewhat more puzzling is that the more three-person cliques in the group, the less the group grows. This could reflect that stagnant groups eventually develop more links among members and hence more cliques.

Question 3: "given a set of overlapping communities, do topics tend to follow people, or do people tend to follow topics?"
A: More frequently, people active in a conference where a topic is hot start going to other conferences where the topic is already hot, rather than the transplantation of people causing the topic to become hot.

Kraut: Developing Commitment Through Conversation

Thu, 28 Jun 2007 19:19:05 GMT

Today I'm at the Communities and Technologies conference, at the workshop on studying interaction in online communities.

Bob Kraut is discussing some of the data analysis issues in his study in Usenet newsgroups of what independent variables predict whether a message would get responded to.

They first did some machine learning techniques to identify the signature of messages that have a "self-introduction". Then they used that as a regressor, along with some directly measurable variables like using first-person pronouns.

He and Moira Burke have a paper tomorrow where they did a controlled experiment. They found that the key ingredient is saying that you're part of the community, not that you share the interest/condition around which the group has formed.

Collusion-resistant, Incentive-compatible

Sun, 17 Jun 2007 13:17:23 GMT

At EC-07, Radu Jurca presented a paper extending work on eliciting honest ratings to consider situations where a set of players may collude to increase their payments for ratings. The setting is the same as that of my paper with Nolan Miller and Richard Zeckhauser, on the "The Peer Prediction Method". That is, a set of raters are scored based on comparing their reports to the reports of other raters-- there is no ultimate ground truth of whether the item is "good" that can be used to evaluate the raters.

Our paper showed that it is possible to construct payments that make honest reporting a Nash Equilibrium (i.e., best thing to do if others are doing it) while creating an expected reward large enough to encourage effort required for the raters to acquire a quality signal about the item. The technique is based on proper scoring rules, applied to the posterior distribution for a reference rater, computed from the prior distribution and the rater's report. Jurca and Faltings considers whether it's possible to make such incentive payments resistant to collusion (e.g., all raters agree to report that the item is good).

Interestingly, the authors find that it is useful to make incentive payments based on the ratings of more than one reference rater. Instead of just adding up the payments determined independently by each of their reports, which I assumed would be the most effective way to do it, the payments are tied to a count of the number of reference raters who report that the item is good. Consider, for example, if the implied probability distribution for each of the reference raters is that each will report "good" with probability 0.6. Then, the number who will report good follows a binomial distribution. By carefully choosing the points a rater gets for reporting "good" or "bad" when n other people report "good", it is possible to rule out some forms of collusion. For example, with 10 raters and a prior probability distribution that each will report "good" with probability 0.5, it is easy to see that we can make the payoff be 0 when either none or all report "good", yet make the payoff for 6 total "goods" when you report good be high enough that you will want to report "good" whenever you see it, if you think others will report honestly. Nolan Miller, Richard Zeckhauser and I had the basic intuition that we could punish all the raters if there was "more than the expected amount of agreement". This fleshes out that intuition with a concrete way of setting the incentive payments.

The most interesting result in this paper comes in section 7, which considers "sybil attacks". One person controls several raters, which I'll refer to as sybils (split identities of the person). They each acquire a real signal. The person is trying to maximize the sum of the expected payoffs of the raters. The authors find that, depending on the particular prior distribution, if one or just a few reference raters is assumed to act honestly, the incentive payoffs can be constructed so that even if the rest of the raters are sybils controlled by a single entity, they cannot do better than to report the same number of "good" ratings as they actually perceived. The technique is a brute force approach (automated mechanism design) that just writes down each of the incentive compatibility constraints (for each possible number of good ratings perceived, the expected payoff given the distribution of ratings from the honest raters, is higher for honest reporting than for any false report) and then solves the linear programming problem to find the smallest expected payment subject to those constraints. It would be nice to get some stronger intuitions about what kind of payments will be selected by the brute force approach. That is, how is it leveraging the small number of honest raters to drive the colluding raters toward honest reporting? Still, I laud the authors for fine work in demonstrating that it generally is possible to resist such collusion, so long as they expect there to be a few honest raters around.

Radu Jurca and Boi Faltings, "Collusion-resistant, Incentive-compatible Feedback Payments", Proceedings of ACM EC'07, P.200-209.

Recommenders and Sales Diversity

Thu, 14 Jun 2007 18:14:44 GMT

At the EC '07 conference, Kartik Hosanagar presented a paper modeling the impact of recommender systems on sales diversity. Do they contribute to a long tail, where lots of products get a few sales, or do they reinforce blockbusters. The paper suggests the latter.

There are actually two effects that we should expect from recommenders. One is discovery-- once one person discovers an item, some other people with similar tastes who would not have found that item do find it. The other is reinforcement-- an item that many people have sampled will be more likely to get recommended.

The paper provides a simple two-item, two-player, two-urn model in section 4. Unfortunately, it begins with an assumption that both players have the same probabilities of choosing the two items, in the absence of a recommender. Without diversity in what people who choose without the recommender, it doesn't seem to capture the discovery effect for recommenders.

Section 5 seems to provide a more promising simulation framework. Consumers have different "ideal points" in the space, and thus are likely to select some distribution of items in absence of a recommender. The recommender that increases the salience of some items to people that are little farther from their ideal point. Even here, however, it doesn't quite seem to capture the phenomenon that the recommender makes salient an item that is in fact closer to the consumer's ideal than what the consumer would have found. It seems to me that you'd need a variant of the Hotelling model where there's a separate model of item salience that is not completely determined by the distance from the customer's ideal. Things that are already blockbusters would be more likely to be noticed and chosen, even if farther from the customer's ideal. That's kind of how the recommender is modeled, but I think it needs to be applied to the base choice model, not just the effect of the recommender system.

D. Fleder, K. Hosanagar "Recommender Systems and Their Impact on Sales Diversity", Proceedings of ACM EC '07, pp.192-199.

Peer Prediction Method with Reduced Payments

Wed, 14 Jun 2006 18:34:05 GMT

ACM EC 06. Jurca and Faltings

Work extends my "Peer Prediction" paper, written with Nolan Miller and Richard Zeckhauser, on eliciting honest reports, by comparing reports between people.

Automatically selects a scoring rule, with lower expected payments but still incentive compatible.

Has some mechanism for probabilitistically filtering out unusual ratings. I'll have to look at the paper to see the details of this.

Claims that the honest reporting equilibrium is evolutionarily stable, meaning that small coalitions can't attack it. Again, I'll have to take a look at this.

Collaborative Filtering with Privacy

Wed, 14 Jun 2006 18:21:11 GMT

ACM EC '06. Presentation on privacy-preserving collaborative filtering.

Previous approaches:

secure multi-party computation to compute eignevectors (Canny).

add noise to each rating

This paper shows that adding noise may not preserve as much privacy as you' d like. If the noise for each rating is a random draw from the same distribution, and if there is a finite set of possible ratings, then you can make a pretty good backward inference about what the original ratings were. The basic idea is...

The solution in this paper is to have users add a variable amount of noise to their ratings, not the same draw for each item.

I haven't had a chance to read the paper in detail yet, but it seems quite elegant. I hope I'll be able to use it in my recommender systems course this fall, though the math may be too advanced.

ACM EC 06: Fudenberg invited lecture

Tue, 13 Jun 2006 16:24:48 GMT

I'm at the ACM EC conference for the next couple days. Computer Science theory/algorithms/AI people looking at economic incentive issues.

This talk: "Stable Superstitions and Rational Steady State Learning", given by Drew Fudenberg (joint work with Levine)

(These are scattered notes taken during the actual talk. If it seems to the reader that it's getting at something interesting, you can probably get better intuitions about it, and more accurate characterization of results, from a paper, or a set of slides posted by Levine.)

Context: Learning in games. Anonymous random matching. Some history of previous papers that went too fast to capture.

"Self-confirming equilibrium"; less restrictive than Nash. No one can do better with "rational experimentation." Nash requires people to know what would happen if you deviate.

Agents off equlibirum path play infrequently, so have much less incentive to experiment. Wrong steps one step off equilibrium can't be stable, but wrong steps two off equilibrium can.

Illustration: Hmmurabi's second law. Accused person is thrown in river. If lives, accuser is killed. if dies, accuser gets their property.
Superstition: guilty are more likely to drown than innocent. This supersition is stable, because accusers rarely get to find out, because if they believe it, they won't accuse the innocent, and they don't get to find out.
Alternative supersitition: guilty will be struck by lightning. This superstition is not stable. Kids try petty crime and discover they're not struck by lightning.

Rational Steady-State Learning
Agent's decision problem: each agent in role i expects to play T times. Agent observes only terminal node each time. Agent believes faces time-invariant distribution of opponents' strategies. (This is wrong, but hopefully a reasonable model of how people would actually be thinking.) Steady states are where people play strategies that are optimal given the information they have from the previous rounds.

Results focus on characterizing steady states as T tends to infinity-- most players have lots of observations of play (but only rational experimentation in those rounds of play), and htere are few novices in the game in any round.

Asymptotic result for Hammurabi caes: there will be no crimes (in the limit of arbitrarily long lifetimes). With long but finite T, some crimes are committed, some false accusations take place, and people making false accusations learn that they work. But if there are few opportunities for being a witness, then there's no rational interest in experimenting with false accusation, because you won't get to do it very often even if you find out that the false accusation works.

Model highlights the role of experimentation in determining when a superstition is likely to survice.

David Parkes: question about applications to Sponsord Search design-- implications for encouraging experimentation or sharing information learned from experimentation.

NetSquared Human Rights Session

Tue, 30 May 2006 22:47:47 GMT

Patrick Ball, Benetech
Small organizations on the ground don't want to share their data-- it's their ticket of entry to policy discussions. They do need crypto and communication so they can get their data to a secure place even if their laptops are impounded.

Make it serve the local need of the person entering the data, and by the way have it do the stuff that's good for the organization and the long haul.

Has been doing statistical analysis to estimate prevalence of Human Rights violations, based on counts and overlaps between sources.

Dan McQuillen, Amnesty International
Mashups are a great publicity/marketing opportunity for human rights organization.
The big human rights battles are about to be fought out on the Internet-- things like

Bryan Nunez, Witness
Trains human rights activists/defenders on use of video (cameras, editing, distribution). Help them use the video as part an action plan.
-------------------------------
Patrick is very concerned about Internet filtering. (Years ago he challenged me about PICS at a CFP conference. Now he's concerned about Google's community tagging and how it might be used by ISPs for filtering. Had an interesting conversation with him at lunch about this.)

NetSquared: state of Open Source Software for Nonprofits

Tue, 30 May 2006 22:16:05 GMT

Some audience questions before the start of the session:

Is Open Source relevant? Or are open APIs all that matters?

Are there underlying values for NPOs choosing tech, or is it just a question of picking what works best

David Geilhufe's arguments for open source for the non-profit sector: avoid duplication of effort; encourage innovation.

OpenBRR (open business readiness rating)-- more appropriate criteria for making decisions on open source than if you use the usual criteria that have been articulated for commercial products.

Zack Rosen on the CivicSpace ecology.
Communities:

Drupal

CivicSpace Foundation

OpenNGO-- the CRM portion

Vendors/Service Providers

CivicSpace, Inc.

CivicActions

Echo Ditto

...+20 more

In the CRM space, biggest three vendors are Kintera, Convio, GetActive. Then there's a long tail with the little vendors. But if you aggregate all the vendors, the CiviCRM community is number two, and much more profitable. Tools are advancing exponentially faster. Vendors in the OpenSource space are bidding 10-50% of commercial market leaders. One and two person shops are bidding against market leaders and winning.

The Mambo/Joomla fork. Major developers didn't like what the people in charge of Mambo did, so they left on masse, and were able to take the source code with them.

NetSquared, CitizenJournalism session

Tue, 30 May 2006 19:12:54 GMT

Dan Gillmor. Citizen Journalism is becoming the norm. Eyewitness reports from disasters are just the beginning. Digg is the darling example now because it has ratings of news stories, though he also mentions Slashdot for rating the commentary. (Look for the new interface reading comments on Slashdot, coming soon, that I've been working on with students Youn-ah Kang and Nathan Oostendorp!)

The OhmyNews story. Korea. Extremely successful; has become one of the most influential publications in Korea. 43,000 citizen reporters==>screening by news Guerilla Desk. Mostly reviews, commentary. Also 65 staff reporters, mostly hard news, analysis. But there's a lot of blending between them. Now trying an international version, and a partnership with a prestigious newspaper in Japan. 86 countries with 1000 citizen reporters so far on international version. Doubling about every 3 months.

Ethan Zuckerman, Global Voices. Story of Hao Wu, blogger detained without charge in China. Effort to publicize his case got much easier once Hao Wu's sister started blogging about the case. Lesson: "Don't speak. Point." Don't try to speak on behalf of others-- just point to those who are speaking on their own behalf.

Some notes from NetSquared, Session 1

Tue, 30 May 2006 19:12:29 GMT

At NetSquared

Howard Rheingold: "Still need a residue of hierarchy, but it can be a pretty small one"

Paul Saffo: "The power of the whisper"

My summary of the morning: when the analysis gets complicated, just remember, "It's the participation, stupid."

From the floor: "a just society means 'not just my society' "

On to Calaveras for WineCamp

Tue, 30 May 2006 16:47:09 GMT

At the dinner after Online Community Camp, Greg Beuthin from ComputMentor told me about WineCamp, where geeks and non-profits were camping out for the weekend. Some of the people from CivicSpace were going to be there, and a major goal of my trip to NetSquared (starting Tuesday, today) was to connect with them. So off I went on Friday afternoon, after a day talking about my research at Yahoo!

Unlike the Online Community Camp, which had borrowed some un-conference ideas, this was the real deal. Saturday morning people introduced themselves, gave a few "tags" to describe themselves, and said what they hoped to get out of the conference. Tibba Phillips, founder of Output for Hope, which helps people find missing persons who are "off the grid", said she was looking to upgrade their website to include a more easily searchable database, so that the project could scale up. Zack Rosen said his goal for the weekend was to build Libba's database. On my turn I piped in that I wanted to watch/help him do it. It became a big barnraising activity, with about 10 people involved by Sunday.

It actually turned out to be an informative dry run for the course I'm planning for winter semester, where teams of students will develop custom sites, using the drupal CMS platform, for non-profit organization clients. Saturday, when we had no power or connectivity, we did requirements analysis. On Sunday, indoors at a winery, we implemented. We only had about 3.5 hours. Zack, Tim Bonneman, and I trasnferred much of the content of the existing site. Several hackers from CivicCRM put together the database part, by using their tools to add custom fields to their basic person-data entity. WineCamp organizer Chris Messina made a new theme so that suddenly, two hours into the work, our generic drupal-themed site transformed to have the look and feel of the existing Output for Hope site that we were copying. I worked on adding help material to the site so that their web volunteer would be able to maintain it. We didn't quite get to a site they can roll out, but we got pretty close and there's a good chance that their web volunteer will be able to take it the rest of the way. Here's the work in progress.

I also connected with Laney from The Princess Project, which is trying to scale up or franchise or do a chapter model of some kind. In a quick brainstorm with Laney and David Geilhufe, we hatched the idea of an online kit that would allow people to self-organize in a new city, and have their progress tracked in various ways so that the national organization could provide appropriate resources at different points, and there could be peer to peer support among chapters. It's basically a franchising/chapter model of scalable organizing, but with some new twists made possible by technology and the peer-to-peer sharing ethos.

I think this peer-to-peer chapter organizing, coordinated by a central toolkit, could actually be the big idea about how IT can help rebuild social capital that I was supposed to come up with for the Saguaro Seminar, but never did. On the ride home, Zack pointed out that this new chapter/franchising model was pretty much what they had tried to do in the Dean campaign. It's also related to what Meetup has been trying to do. And it's sort of what BarCamp is already putting into practice.

It was also a truly wonderful experience for the senses. Wine from Ferriere vineyards, swimming through a cavern, sleeping under the stars, amazing vistas, yoga in the woods.

See photos from the Flickr feed (WineCampCalaveras):

Online Community Camp

Thu, 25 May 2006 18:08:46 GMT

I'm at an Online Community Camp.
"Camp" is the new word for conferences that are only loosely organized-- people propose topics at the beginning of the day and people go to whatever seems interesting.

Who's here:

Vendors

Consultants

Community managers, web producers at companies, non-profits, and media outfits

one student from Stanford, and me, reprenting academia

----------------
Topics they're interested in:

how to change platforms; how to select platforms

How to quantify ROI, to justify and get resources

Some inteest in reputation

Using online communities for market research

Media wants audience to talk with each other, how to facilitate that

Multiple communities, how to not require multiple destinations, cross-site integration

Practical applications of Web 2.0-- what's hype vs. useful

How to apply social networking/myspace

Extracting/summarizing from online discussions
--integration with corporate

Online/offline connection

Best practices across the board

Combining data from other sources about people with data from online communities.

Blogs and RSS vs. conventional discussion boards