Bounce rates on my bingo site

[Michael Martinez]

Yeah, but I think it's a core subset of the millions of possibilities. It's got to be looking at "how a webpage might behave" in how it presents content to visitors, how it references other sites, which sites it references, how often it references other sites, whether it allows users to contribute content, etc.

It might be more productive to think in terms of "number of page behaviors" than actual quantifiable elements and element relationships.

[Squizzel]

I see what your saying but my instinct says that a lot of those metrics, even when combined to score hundreds of metrics, it seems like they would be just as unreliable as user metrics. Sure there is easy stuff to catch but a lot of spam sites are formatted very similar to high quality sites and would be seemingly very hard to differentiate.

Interesting discussion anyway.

[Michael Martinez]

...a lot of spam sites are formatted very similar to high quality sites and would be seemingly very hard to differentiate.

And a lot of spam sites do seem to be getting past the Panda.

I don't know what metrics they are using. I can sit around and dream up possible metrics all day long but there is no way I can test them.

Anyway, maybe an analogy would suffice. Suppose you gathered 1,000 poker players into a room. You want to separate them into "Really Good" and "Everyone Else" groups. You get to pick any 20 players that you know personally well enough to consider them Really Good or just one of the crowd (no one else's opinion matters -- just yours).

You sort them into two groups. Now a computer gets to examine everything you know about those 20 people and, applying that knowledge, uses it to sort the other 980 into two groups. It begins by looking for the best mix of signals that would produce your division of 20. From there its starts to sort everyone else.

Occasionally you might grab someone from one of the computer sorted groups and place that person into one of your groups. The computer then looks at your knowledge of everyone again and once again tries to find the best mix of signals to produce your hand-selected groups.

What do you know about these poker players?

You know which ones try to go for flushes and which ones try to go for full houses (more often than not).

You know which ones drink whiskey and which ones drink coffee.

You know kinds of clothes they wear when they play.

You know all their nervous tics, their obsessive compulsive body language, the ways they conduct themselves when they are confident, when they are bluffing, when they are worried.

You know what they read for entertainment.

You know their favorite movies.

You know what cars they drive, what furniture they buy, what kind of women (or men) they like to date.

You know how many children they have, how many parties they attend every year, how many hands of poker they play per game, how many poker games they play, etc.

You know a LOT about these people. Intuitively you just sort them into "Really Good Poker Players" and "Everyone Else". But the computer takes everything you know about these people and looks for statistical correlations that -- when taken in the right mix collectively -- suggest how the people might be sorted (even if internally your brain is using entirely different criteria).

It's the correlative data models that make the difference. The computer can match them up to your choices faster than anyone else (and everyone else combined).

It could be that in your initial selection all the Really Good Poker Players chose to wear blue sweaters more often than the Everyone Else players, and that the Really Good Poker Players on average drink more beer than whiskey while the Everyone Else players on average drink more bourbon than martinis.

It's not obvious, glaring stuff. It's combinations of little things that statistically fall into place more often for one group than the other.


Another analogy: Let's say you have collected all the astronomical data that mankind has ever compiled. Out of that data you have identified what YOU believe are 10 really bright shiny stars and 10 black holes. You turn your data over to the Panda algorithm and say, "Find these 10 bright shiny stars and these 10 black holes". It sorts through every piece of data and picks out 10 things it says are your bright shiny stars and 10 things it says are your black holes.

You keep telling it to go back and do it again until it gets the right 10 stars and 10 black holes. Once it has done that, it has a formula for sorting other stuff into stars and black holes. So you turn it loose on your data to find as many stars and black holes as it can. As it starts sorting through the universe you notice it gets a few things wrong so every now and then you add a few stars to your group of hand-selected stars and you add a few black holes to your group of hand-selected black holes. Each time you enlarge your hand-selected groups the algorithm goes through all the data again until it can match your groups.

And so on and so on and so on. You don't have to know anything about the signals. You just handed the algorithm every astronomical fact on Earth. It figured out what the signals should be.

I suspect strongly that the Google engineers don't know much more about which signals matter most than the rest of us until the algorithm has figured out the latest combination.

[Squizzel]

Nice analogy, makes sense Leaves me with more questions than answers though. hah. I cant help but feel this pondering is just fruitless. Ill go back to building links tomorrow.