LorenzoC wrote:Arteekay wrote:There's not a greater chance of false positives, it's assured, at least initially and never, ever without any. That's the nature of the deal, it may approach perfection, but it will never reach it.
Hold on.
Imagine you start the day 0 with your bayesian AdBlock and say "mark this as AD". The software will recognize some elements of the object and add those elements to its internal database. The day 1 you meet another object that has some of the elements of the marked one and the software, using its rules, estimate the probability as "high" and removes the object from the page. Now, how can you tell the software "no, this is not an AD" since it was already removed and you can't see it any more? You need a list of "removed objects" like the "spam mail" so you can review what was marked as "AD" and correct the false positives.
Yep, I believe we're saying exactly the same thing, or at least that was the intended point of my post if I munged it up. On day 0 you'd get 0% false positives and 100% false negatives (shouldn't be blocked). If this page had one ad and one context object, you'd block one and still see the content (hopefully) and the stats would shift to 100% false positives and 0% false negatives. Then you'd go on to a page with two ads and two context objects, and hope you still had 100/0, it you get 50/50, you'd then block the second ad and be hopefully be back at 100/0. This of course is very unrealistic and I'd assume it would probably come by default with something better than a completely "new born" detection and false positive training level.
LorenzoC wrote:Arteekay wrote:Until it's at least close to perfection, it will definitely require user interaction [...]
Actually after some instructions the filter should be able to block 99% of the ADS, also because more or less you visit the same web sites. As i said before it depends on what elements the software would evaluate. In any case the interaction with the user would be only "yes" and "no", since you don't need to build a blacklist of elements or worse to write regular expressions.
Here's where we strongly disagree, at least at this point and without any knowledge of how Wlad might implement. I would think it would be exponentially more difficult to get from 98 to 99%, and even getting into the 80% range would prove to be difficult. Even speaking only about sites on the "regularly visited" list, any change in content, based on whatever properties the filter works on, would throw the status way out of the 90% range.
Unless of course we're speaking only of perceived and/or historical stats, in which case the above unreal example would let you get to 100% in a single move. In my opinion, any filter that hits 70% or better with 0.0000001% false positives is an incredibly excellent filter.
Getting to that 0.0000001% falsely blocked element would require two things:
1) the user would have to recognize that a false positive occurred.
2) some method of seeing what was blocked, and of course some way to unblock it. A "yes" or "no" scenario might entail asking the user after each page element loaded (not before, since we wouldn't know what exactly it is until we can see it).
Once again, this is speaking from the viewpoint instilled from spam filtering. Another thought, imagine the trouble a page of joke banner ads might cause.
). Try to read this again without alcohlic influence and more IQ: No one said the ADP 0.5 will be developed further. But it will be maintained.
