EVE General Discussion

Of harassment detection in text logs? That's easy: http://dl.acm.org/citation.cfm?id=2187995

It's not fundamentally different from the common examples I already gave and I imagine you're going to make a lame counter-argument with the exact experimental results (not understanding the sampling methods or trade-offs), but there's more supplemental knowledge than I can really hope to address.

An example perhaps more on your level of understand is the LoL tribunal. Lots of people have played with that data set (though not in an academic context afaik) and hit 90-95+% accuracy rates with naive methods.



You haven't; giving information to the human and not the algorithm is obviously not a fair comparison and invalidates your "thought experiment".



Even you must be familiar with Watson?

And this is where you fall down, because identical actions are performed whether the intent is to harass or make gains from it. I know you understand this by now. I know you don't want to admit you failed to see this when you made your original statement, but it's long past time you stop stating falsehoods as truths to try to support your claims.

OK, so here you've admitted you don't know the rules. This is at least a start, I guess. I'll offer a hint: CCP have said that certain actions can be against the rules - how is that determined? More importantly (really, much more importantly) why is it done in this way instead of stating a limit on the number of times someone can be bumped?
C'mon buddy, do it. You're nearly there! You're almost at the point of understanding why classification and case-by-case judgement are different concepts.

You're selectively quoting from a passage where I say computer analysis can't show intent to disguise the fact you can't demonstrate this.

If this is "easy" in the same way everything else you say is "easy" - then what you really mean is "yeah, there's kinda some research in the area, it's being improved on but is yet to reach statistical significance" ... then I think we've proven that when you say something is "easy" what you mean is "potentially possible in the future" and I think we can leave it there.
My issue was with your original post

Which is still demonstrably untrue, now even more so since you have proven several elements of any such data-mining task would be unable to achieve statistical significance. You can't very well use chat-logs to determine whether identical actions are harassment or not, if the machine is flipping a coin on it.

On a very basic level, the stated usage case of such an analysis technique (more developed than currently available) would be to raise flags for moderators to pass judgement on. They are not stating it can determine, for itself, whether the material qualifies - it is simply spitting out a number stating how close it fits the accuracy of its (not well refined) model.
In EvE, players raise petitions when something like this happens, so there is no need to have a massive data analysis tool running around and repeating that task.

"I imagine you're going to make a lame counter-argument with the exact experimental results"
Well yes, it might be quite boorish of me to insist that we discuss matters of provable fact when we're questioning fact, but dem's the breaks.

Yes, but this is just you (once again) seemingly clinging onto the dream of what might be and (for some reason) claiming it would be trivial to implement.

Just two years ago, one of the world's leading tech companies, with a massive budget, produced a highly specific algorithm, running on a supercomputer, for answering knowledge-based questions in a clear unbroken and expected format. The reason you know about it, the reason I know about it, is because it is/was a computing breakthrough.

You claimed that going WAY beyond the scope of Watson was possible 30-40 years ago.

Do you have any idea why you are being pointed and laughed at?

I already gave you my understanding, as well as how it's done in virtually every other context. If you disagree, perhaps you should put a citation where you mouth is.



Naw, man up mate. You said it was impossible despite there being a good decade+ of productive/useful research in the field. Read the paper; read some other stuff in the field; then we can have a two-way discussion.



Oh jeez, I hate teaching statistics. Do me a favor and either take my word for it(the authors wouldn't have published statistically insignificant results without saying so) or do your own reading?



If you'll recall, I made the disclaimer several times that this obviously wasn't something appropriate for the problem. I'm strictly indulging your tangent in the hopes that you'll learn something despite yourself.

I don't recall saying it would be trivial to implement. Data mining techniques can take quite a lot of manual effort to fit the data set.



It beat the most adept humans in the world at a fairly complex task (notably harder than what we're talking about mind you). It only required a supercomputer to meet the latency requirements of the show (and not a particularly high-end supercomputer); you can run the development version on an ordinary desktop. The system itself allows for very general application; some are pretty neat, I suggest you read up on them. Annnd, it wasn't a breakthrough, more of a highly publicized milestone.



Text analysis can be challenging to do properly, but what I had in mind were decision trees (which popped up in the 70's if memory serves). Lots of the techniques used for cutting edge stuff (Neural networks, ect.) were conceived ages ago and it's getting the necessary hardware/optimizing them for it that's hard.



By you? I don't really care to conjecture on issues of psychoanalysis. You do remind me a lot of college freshmen though.

* PVP flags CAN be created and further extended after log-off even if the owner did not have a PVP flag at the time of disconnect.. If Char A logs off in space (with or without a PVP flag), and then char B attacks A, then A will get a PVP flag. Char A's ship will then remain in space for as long as that PVP flag exists.
These changes should ensure that unavoidable disconnects (eg caused by network problems) aren't massively penalising, whilst ensuring that manually killing the client to avoid PVP is never a viable strategy.

https://forums.eveonline.com/default.aspx?g=posts&m=2079573#post2079573

I doubt it, the devs don't have a history of catering to special snowflakes.

speacial or not corn/snowflake or not i think that wat i witnessed was not intended gameplay , hence why i started this thread .

The problem with this system is it isn't "emergent gameplay". It is absurd abuse of game mechanics on the same level of OGB. The fact that there is no counter to it that is feasible is really off putting in a so called space sim. Fun game mechanics are when you have a viable counter to a tactic that is used against you. Where is the counter in this? Basically it is sanctioned harassment designed by the inept and perpetrated by the small with no chance of recourse. If bumping and suiciding are allowed to keep a person locked down for over an hour, then why was POS bowling patched out? They are both viable tactics right? Yet one was deemed an exploit and one was not. If you want a person to stay there and not be able to warp off, you should have to aggress and use a scram/point. Not some chickenshit tactic of bumping someone for over an hour completely safe from repercussions due to flagging mechanics.

Well .. it's been linked many times, but ok

CCP considers the act of bumping a normal game mechanic, and does not class the bumping of another player’s ship as an exploit. However, persistent targeting of a player with bumping by following them around after they have made an effort to move on to another location can be classified as harassment, and this will be judged on a case by case basis.

Now we know two things based on this and other CCP GM and DEV replies:

- It can be considered harassment, which will be judged on a case-by-case basis
- CCP do not stick to any particular criteria for classifying it. They do this to prevent someone gaming the system. They try to determine the players intent from their actions, and if they deem it so, they class it as rule breaking.
You can see this in their rookie-system rules where GM Homonia specifically said they judge the intent, not what happens or where:

"It is impossible to define [...] in a way that is comprehensible, to the point and without loop holes, in addition to our players able to apply these rules to their fellow players around them. This means that we will not provide a hard definition to our player base"

CCP said they have internal measures that ensures rulings are consistent, but that these are not defined by any facts or has a numerical basis, since players will reverse engineer them to try to exploit them.

CCP have been very forth coming on the fact they judge each case on it's own individual merits, and this (needfully) requires considering things outside of what simply happened on the server in a loggable way.

It's really the classic definition of pornography - "I know it when I see it." -- that statement is oft-repeated because it's a simple idea that encapsulates the concept that something can be one thing or another, even if every possible description of it is identical to another. The human system of perception is above where computers are with simple data classification.

I mention this, because CCP GMs specifically mentioned this concept with regards to how they determine whether the intent of the player was to harass, or whether they were using game mechanics normally.

Er, I did. The paper from a scientific measurement POV, was simply "we can kinda analyse this a little bit, here is one possible analysis technique that does a little better. Yep, it did better! Still can't tell us harassment from non-harassment in a manner which is statistically significant, though"
Are you going to continue this bizarre tradition of literally stating something is different than the published facts?
The object of the research was to move closer to being able to ID harassment accurately, so yes, a paper that shows they are getting closer to this (but have not achieved it) is not unusual in any way at all. Non-significant results are interesting unto themselves.

Well, sorry friend but my key points are still the same, and unchanged. There doesn't exist a computer model which can judge a players intent. CCP make judgements based on (what they believe is) the players intent.

What I actually wanted to talk about with the initial example of the credit card / fraud data (until you ran off with the need to insult me) was data-size. The biggest problem of ever trying something like this will be data-size; the disparity between population data and sample data.
You probably can look at every instance of cases ruled as harassment and analyse it against every instance of non-harassment and find elements in either that have a non-neutral effect on the groups. The problem will be your sample for the instance in question will, in effect, be a small sample and be over-ruled by the individual biases that would otherwise be smoothed out in a larger dataset. In short, you're going to have large standard deviations even if you can pick apart the means, meaning it is inordinately hard to tell where data lies by comparing it to such.
For instance, you might find the biggest indicators of harassment is where the infringing player is trying their best to make contact with the victim (to collect tears and validate the harassment) - well, this might hold statistically true for the dataset as a whole but can never be better than a mild statistical weighting in an individual case.
If it's a provably statistically significant difference, but the means between groups are 4.5 communication attempts and 4.7 - with a standard deviation of 3, what weight can you possibly put on your small sample of data (one infraction) where the player made 3, or 6?

Mathematically classifying things isn't hard, even when you are looking at human subjects. What is hard to classify is singular data points, wherein you can only make a judgement call and "know it when you see it".