Slashdot Mirror
"Long Tail Effect" Doesn't Work As Advertised, Say Wharton Researchers
Death Metal writes "In a working paper titled, 'Is Tom Cruise Threatened? Using Netflix Prize Data to Examine the Long Tail of Electronic Commerce,' Wharton Operations and Information Management professor Serguei Netessine and doctoral student Tom F. Tan pull information from the movie rental company Netflix to explore consumer demand for smash hits and lesser-known films. Netflix made its data available as part of a $1 million prize competition to encourage the development of new ways that will improve its ability to introduce customers to lesser-known titles they might find appealing." In short, the researchers say that the Long Tail effect described by Chris Anderson is much less important in the real world than popularly held. Says the article: "The key difference between the opinion of [Anderson's] book and the study by Wharton researchers is how they define 'hits' and 'niches.' In the book, Anderson focuses on the definition of hits in absolute terms such as the top 10 or top 1,000 products, while Netessine and Tan argue that, to take growing product variety into account, one has to define popularity in relative terms, such as the top 1% or top 10% of products, to properly assess the presence or absence of the Long Tail."
That's precisely the point. If the shape is such that a top movie gets only 1% of the market, top movies won't make enough profit to justify hiring Tom Cruise and it's a problem for him.
OK, I am not a mathematician, but this paper makes me deeply skeptical.
If the input data is indeed heavy tail (non-existing higher moments) or quasi-heavy tail (existing, but extremely large higher moments) how on earth they can use variance, R^2 and other measures? They may not even exist! And if the input is quasi-heavy tail, then of course they exist, but the convergence time could be arbitrarily long!
I had the unpleasure to work with quasi-heavy-tailed data, and it is really enlightening. You watch the evolution of some metric (e.g.: avg) as the function of incoming data, and you see of course convergence. At least for a while. And then in sudden an extreme outlier comes in, and the avg takes a huge jump! Now if your input is heavy tailed enough, you can be never sure that your measure finally came to rest (converged), or the next jump is just over the corner!
I hope a more educated person clarifies this, I am just an engineer.