I agree with Gregory that it is very useful to quantify the usefulness of
trust information on text -- otherwise, all comparison are very subjective.
In our WikiSym 08 paper, we measure various parameters of the "trust"
coloring we compute, including:

   - Recall of deletions.  Only 3.4% of text is in the lower half of trust
   values, yet this is 66% of the text that is deleted in the very next
   revision.
   - Precision of deletions.  Text is the bottom half of trust values has
   probability 33% of being deleted in the next revision, agaist a probability
   of 1.9% for general text.  The deletion probability raises to 62% for text
   in the bottom 20% of trust values.
   - We study the correlation between the trust of a word, sampled at random
   in all revisions, and the future lifespan of a word (correcting for the
   finite horizon effect due to the finite number of revisions in each
   article), showing positive correlation.

Some aspects are not captured by the above measures:

   - We ensured that every "tampering" (including cut-and-paste) are
   reflected in the trust coloring, so it is hard to subvert the algorithm
   (does "age" provide this?).
   - We ensured the whole scheme is robust wrt attacks (see the various
   papers if you are interested).

I fully believe that it should not be hard to improve on our system re. the
above measurements.  And I fully agree that the "reputation" we compute is
essentially an internal parameter of the system, and does not really
constitute a good summary of a person's overall Wikipedia contribution; for
this and other reasons we do not display it.

Luca

A simply objective challenge for any predictive coloring system would
> be to use them in the following experimental procedure:
>
> * Take a dump of Wikipedia up a year old, use this as the underlying
> knowledge for the systems.
> * Make several random selections of articles and include the newer
> revisions not included in the initial set up to 6 months old. Call
> these the test sets.
> * The predictive coloring system should then take each revision in a
> test set in time order and predict if it will be reverted (Within X
> time?).
> * The actual edits up to now should be analyzed to determined which
> changes actually were reverted and when.
>
> The final score will be the false positive and false negative rates.
> So long as e assume that the existing editing practices are not too
> bad we should find that the best predictive coloring system would
> generally tend to minimize these rates.
> _______________________________________________
> Wikipedia-l mailing list
> [email protected]
> https://lists.wikimedia.org/mailman/listinfo/wikipedia-l
>
_______________________________________________
Wikipedia-l mailing list
[email protected]
https://lists.wikimedia.org/mailman/listinfo/wikipedia-l

Reply via email to