« A.O.I.R 6.0 | Main | State of Play III »

Oct 07, 2005



What kind of data should a virtual world expose in order to facilitate more/better research?


Well - when I got to PARC and they showed me the WoW data they were getting, I remember my first reaction was "that's cool, but how much analysis could we really do with 7 variables?"

- Name
- Level
- Class
- Race
- Guild Name
- Grouping Status (Y/N)
- Zone

Let's just say we've analyzing it for the whole summer and we've just gotten to the social networks.

So honestly, it doesn't take that much. The key thing for us was being able to systematically sample the servers via an automated script to generate the most important 8th variable - time.


Mike, to expand a bit on what Nick just said, I would like to add a few things related specificallly to social networks. A key factor in doing social network analysis is knowing exactly who interacts with whom. It might sound simple but in many virtual worlds you have to work with proxies, e.g. co-location in our WoW research. Back when we were working on SWG we tried to use gestures as a proxy but the results were not as satisfying.

So, simply knowing things such as the ones below get you a long way:
- who is talking to whom using any modality (e.g. for now, spatial chat is much more difficult to analyze than private messages: who is really "hearing" and reacting to what is said?);
- who is spending time with whom. For instance in WoW we have the "grouped" variable, but we don't know the membership of the group. While we can estimate it for guilded players with our proxy, that leaves a significant number of grouped players out of our analyses;
- any indicator of frienship or other social relation, e.g. friend's list, guild rosters, etc.

Another important point to mention is that social network analysis works well with anonymized data. Very few analyses require knowing someone's name since we are looking at structures and patterns. Exposing this data to the outside world can therefore be done in ways that maintain the players' privacy.

Now, if I had access to reliable data such as this, I would get greedy and ask for the chat logs too - there all sorts of interesting things you can do when sprinkling a little bit of natural language processing on top of socio-structural data :)


Ye gods, Nick, how many PhDs do you want?! This is excellent data, presented highly informatively, I'm just in awe. It points to answers to so many questions, yet raises so many more new questions to add to them. How do relationships change over time? What happens when a lynchpin in a social network quits? Can a configuration of a social network today predict a new configuration tomorrow (will mage 16-21 inevitably link to warrior 12-16, or only under certain very specific conditions?). You're producing new data faster than we can consume it.

This is fascinating stuff. Please don't get run over by a truck.



Next question: What questions are you really trying to ask of the worlds?

For example: I suspect that one of the "questions" you'll ultimately want to know is whether your findings can be translated from the the virtual world in the real world.

The real world doesn't have magic users, fighters, etc., so that information isn't terribly relevant. Therefore, you'd want to collect data that's more translatable to the real world, such as guestimates of a player's personality... Do they prefer to tackle problems using (a) brute force, (b) stealth, (c) intelligence, (d) social networks, etc. (These can be guessed from class/race, but I suspect they're weakly correlated.) Then, you can ask, "Do people who use intelligence" hang out with with other people that prefer to use intelligence, or are they attracted to opposites? Etc.

In essence, the virtual world is useful because (a) you can watch everything that people do, and (b) you can set up little experiments and call them games/quests.

Or am I wrong?


Richard said: "Please don't get run over by a truck."

Duly noted :)

Mike asked: "What questions are you really trying to ask of the worlds?"

I have several takes on this (and Nic probably does as well):

- Given that more and more people will interact in virtual worlds (via avatars, etc.), behavior in VWs is interesting in and of itself.
- How many of RL norms actually transfer into VWs? And what are the differences?
- And could we use VWs as large-scale experiments/simulations for economics/law/sociology?


For the purposes of comparative analysis, how would one both define and determine "RL norms"? And even if those behind the avatars were known such that a more thorough analysis was possible, does the mere fact they're "behind the wheel" affect their behavior (e.g. "road rage") sufficiently to render the results suspect?


csven asked: "For the purposes of comparative analysis, how would one both define and determine "RL norms"?"

The differences at the individual level (which you seem to be thinking of) would be more suited to a lab experiment.

I was thinking more at the aggregate level:

- In the real world, male-male dyads maintain less eye contact than female-female dyads. Female-female dyads stand closer together than male-male dyads. Do these gender norms transfer into virtual worlds even with all the gender-bending happening?
- Are social networks in virtual communities comparable to social networks in physical communities?
- Given the instrumental nature of MMOs, what is the ratio of instrumental chat messages compared with social chat messages? And is this ratio comparable to the typical ratio in a business meeting?


Very nice.

My wish list:

Graphs for horde vs alliance at particular level ranges.

Graphs for the "top" guilds from different servers.

Change over time graphs of particular guilds..

Graph of my own guild.. =)

It would be interesting to see how graphs of effective guilds compare with less effective ones. I'm also curious about the effects of alts on the data. I'm guessing they would tend to show up as a large number of relatively weakly associated nodes.


This is really interesting.

Can player gender and character gender be made the 9th and 10th variable? It would be interesting to see how these two variables affects the social network map.

Is the druid in the center lower-right map a female character or a male player? Is this an priest archetype? For the cluster for 4 hard-core players are they male "band of brothers"?


Ooh, I love this stuff! Mind if I ask some potentially stupid questions? Let's say for a moment that you were able to anonymize non-anonymous data for public consumption and include data from enrolled participants. How big of a data set would be needed? And would it be useful to compare declared networks to defacto networks? Finally, how difficult would it be to obtain and integrate additional variables that might add to the richness of the maps? (if, Nicolas, you could get those chat logs?) I'm with Mike and the others - I love the idea of looking at a VW this way, but I get greedy almost immediately, wanting to know what lies behind those maps in terms of relationships, behaviours, interactions, etc.

Although I'm now taking a wholly qualitative approach, I actually thought about doing some social network analysis using chat logs from my participants (my CoH logs include grouping info and chat modalities - including guild and friend channels - to help structure them), but worried that it just doesn't represent enough data points, plus it would only illustrate a player-centric view (though I thought about trying to enroll a whole guild, but still wondered if that would say anything useful). But could this be combined with the more global view somehow? Could a few blanks be filled in here and there through some explicit and possibly qualitative participation? (my qualitative tendencies attempt to mask my quantitative inadequacies, though I'm quickly becoming a convert!) The questions I have relate to how one knows if a group is functioning well... is there some commonality to the profile/behaviours of successful guild leaders, etc.? This is the stuff that I think might help move this data from really interesting to possibly actionable. My experience with commercial research suggests that the aggregate can always benefit from an occasional zoomed-in view -- the closer view is where the a-ha moments really come from, although meaningless beyond anecdote without the context of the aggregate.

And yes, I second Richard's comments... thanks so much for this contribution, Nick/Nicolas and others.


Even from a purely-player's perspective, this highlights so many elements and insights. As usual (tired of hearing it? ;) ) great work!

I do want to echo Richard's question though:

Richard Bartle wrote:

How do relationships change over time? What happens when a lynchpin in a social network quits?

How hard would it be to build an animated sequence showing the charts changing over, say, three months?

I recommend three months because six would seem to assume too much class consistency and retention in a game where non-hardcore tend to jump around. Conversely, one month would be a good estimate for the time it takes to level into next strata, but that doesn't seem like it'd be long enough to show lasting social impacts.

The comments to this entry are closed.