[adnzzzzZ]

>Then there are some speculation what data could it be and how would it be used. From that, it seems that the difference between other "Better audience targeting" tools used for typical commercial advertising campaigns would be mainly scale and accuracy?

Some of it is explained on this 5 minutes video (ignore the clickbait title). https://www.youtube.com/watch?v=yo3gOoOSdhY According to this professor one of the things they used for better targetting is a person's Big 5 personality profile that they extract from Facebook likes.

[NPMaxwell]

Great video. Lots to think about. Like personality tests in general, test of big 5 personality traits correlate with behavior at about .2 to .3. For example, they have a correlation of .2 to .3 with GPA (http://www.gwern.net/docs/conscientiousness/2007-noftle.pdf). This is a typical correlation level with behavior for all personality tests. To give you a sense of this, simulate a personality test result with 90 zeros and 10 ones. Then add a column of 90 zeros and 10 ones for the behavior you are predicting. Arrange your behavior ones so that 7 line up with personality zeros (were not predicted) and 3 line up with ones (predicted). That's a correlation of .24. If you picked 10% randomly, you would have found (on average) 1 of the people showing the behavior and you would have mistakenly flagged 9 who would not produce the behavior. With the test, you find 3 and mistakenly flag 7. If the test and behavior are spread 50/50, you find 32, when random selection of 50% would have found 25. Combine that correlation with the problem that the Facebook data is not perfectly correlated with how people would fill out the Big 5 personality instrument. If you insisted that your modelers create a model of the the Big 5 and then insisted that they use the Big 5 scores to predict behavior, chances are that you would have correlations near zero. The opportunity here is in creating models directly from Facebook characteristics to the behavior of interest. The Big 5 theory may help guide which FB characteristics to test, but then you need to set it aside to create working models.

[FeatureRush]

Oh, so it's also a better underlaying model. Looking now at the company website they are using less popular name OCEAN Personality Model for Big 5.

As far as I know Big5 is not the final, definitive model. There is this famous quote about performance of NLP machine learning system going up after firing a linguist, I wonder if we will someday hear similar comment about psychologists?

[bamboozled]

Thanks for sharing that, he raises some really good points!!!