Datamining Facebook 'Likes'

A recent publication in the Proceedings of National Acadamy of Science demonstrated the power of information analysis (data mining) to formulate predictive models of users using their likes from Facebook. The article is interesting because it sheds light on the methodologies and problems with the these methods.  In contrast, the popular media take on the story was ZOMG, Facebook can tell you're a dude, see USA Today and NPR. The money quote from the USA Today article comes from Rebecca Lieb, a digital media analyst at the Altimeter Group, a consulting firm in New York City, 'Advertising and marketing focus on this, but it's important not to isolate this as only an online issue or a social network issue,' she says. 'Data is being collected at every stage of our lives. If you're using a credit card, you're opening yourself up to as much data collection as if you're using Facebook or searching online and getting cookies collected in your browser.' So don't worry its not a big deal.

The use of likes is interesting from a data analysis standpoint because it is only capable of a binary indication and the negative value is a confounding of don't care, don't like, and unexposed to the tag. This makes the data set fairly weak compared with other categories of data that are easily accessible online. On the upside, the number of tags is a limited analytical space which facilitates regression. Principle Component Analysis (PCA) would likely produce the same kinds of outputs with 20 or 30 variables rather than the 700+ used here. Also problematic was that they include profiles with only a single like which nessisarily hurts their predictive accuracy.  It would have been interesting to look at the model with an increasing minimum number of likes to see when enough data has accumulated to form a useful model.

The article also demonstrates the problem of outliers. The correlation of liking Science with high intelligence probably reflects a general correlation, while the correlation of 'curly fries' with high intelligence is likely an outlier resulting from a few data points. One problem with data mining is that if you fail to move the p-value of correlations up significantly then you will detect more noise events in your data, while if you set it too high then you will overlook useful information. With 700 inputs, use of a p of 0.05 will result in 35 'significant' correlations by mere chance. The article uses p values of <0.001 but this still doesn't prevent the problem of spurious correlation.

More interesting is the predictive power when dealing with non-binary characteristics (e.g. age). Significant correlations were found with Intelligence, Extroversion, Openness, Emotional Stability, Agreeableness, etc. The problem is that given the very large number of measurements used, the threshold for significance is trivially low such that when compared with the natural variation is irrelevant. In this instance, given that n > 50,000 then a Pearson's correlation of 0.01 has a p value of less than 0.013. This indicates a problem of what significance means in a large data set.

Still it is an excellent insight into the type of data available from a moderate sized data set and we are unlikely to get publication of many sample sets such as this.