Snip from someone who is smarter then me at these things...
Assuming that the first measurement is x likes, the second is y likes,
Then the estimate of the natural logarithm of the growth is given by
log(y / x) with an error estimate of sqrt(1 / x + 1 / y)
But since you are interested in the conservative estimate of the growth, you should use something like ~ 5% confidence interval. So I would recommend ranking your dataset using the folllowing function. log(y / x) - 2 * sqrt(1 / x + 1 / y)
For example:
growth from 1 to 10 will get the score of 0.2
growth from 100 to 400 will get the score of 1.16
growth from 10000 to 15000 will get the score of 0.38
One of the important properties of this estimator will be that the growth from say 10000 to 100000 will be ranked higher than the grown from 1000 to 10000, which in turn will be ranked higher than the grown from 100 to 1000 etc...
Matt's teacher's statement on the lack of knowledge on Linear Algebra was: ‘How can you make cheese if you don’t know where milk comes from!? Its plain, common ordinary horse sense!’
That hit home hard :(
On another note, Googling for: log(y / x) - 2 * sqrt(1 / x + 1 / y)
throws up one heck of an awesome graph
I think your algorithm puts too much weight on StumbleUpon "recommendations" which are really StumbleUpon "views".
I was happy but surprised to find my startup http://taskarmy.com in the list, and having a closer look at the metrics taken in account, it seems that it is because of the StumbleUpon stats.
Fair point- I've tried not to make any arbitrary decisions on how the data is used but I'll take a look at what the results look like sans stumbleupon.
What are the metrics based off of? What is fast growth vs big growth? I couldn't find it on the site and I even downloaded the CSV data but it only had the aggregated numbers.
Not sure you can reliably measure the size of a business based on its social media engagement, nor Stumble Upon "views" for that matter. Some of the fastest growing companies (based on things like, you know, revenue, profit etc) are enterprise businesses which inherently have less social interaction than a recipe or deal sharing site.
Also, no offence intended to Guy (the creator), but is it a coincidence that his Lifx startup is #1, and that that is heavily due to the StumpleUpon views which have jumped massively in the last week? Looks a little too convenient to me ;)
"better then hype alone" was my philosophy. Yes I'm involved in the current #1 (and others on that page). I disclosed this on twitter, published the algorithm and made the metrics visible.
Is bugmenot and retailmenot Australian? I can find references to an American company whaleshark. The rest I find of dubious value as I am not in their target market. I find the Australian startup scene as mostly a poor copy of US projects from 6m-2y ago. New ideas don't stay down under for long.
I'd like to see a startup specializing in web fonts that aren't completely unreadable on Windows. Seriously, most of the letters on that page are literally in multiple pieces.
26 comments
[ 4.8 ms ] story [ 63.6 ms ] threadBut what is the logic behind the formula used for the fast_growth metric?
I see that the formula is: log(day14_sum_of_metrics/day1_sum_of_metrics) - 2*sqrt(1/day1_sum_of_metrics + 1/day_14_sum_of_metrics)
Is this a standard metric? Would be great if someone could share the theory behind the math..
Assuming that the first measurement is x likes, the second is y likes,
Then the estimate of the natural logarithm of the growth is given by
log(y / x) with an error estimate of sqrt(1 / x + 1 / y)
But since you are interested in the conservative estimate of the growth, you should use something like ~ 5% confidence interval. So I would recommend ranking your dataset using the folllowing function. log(y / x) - 2 * sqrt(1 / x + 1 / y)
For example:
growth from 1 to 10 will get the score of 0.2
growth from 100 to 400 will get the score of 1.16
growth from 10000 to 15000 will get the score of 0.38
One of the important properties of this estimator will be that the growth from say 10000 to 100000 will be ranked higher than the grown from 1000 to 10000, which in turn will be ranked higher than the grown from 100 to 1000 etc...
Coincidentally there is another post on HN's front page right now on Data Science resources: http://news.ycombinator.com/item?id=4930965
Matt's teacher's statement on the lack of knowledge on Linear Algebra was: ‘How can you make cheese if you don’t know where milk comes from!? Its plain, common ordinary horse sense!’
That hit home hard :(
On another note, Googling for: log(y / x) - 2 * sqrt(1 / x + 1 / y) throws up one heck of an awesome graph
I was happy but surprised to find my startup http://taskarmy.com in the list, and having a closer look at the metrics taken in account, it seems that it is because of the StumbleUpon stats.
Pretty much every club such as Rotary, Freemasons etc only admitted men until relatively recently.
Fast growth: Given the last 14 days, score = log(day_14_sum_of_metrics / day_1_sum_of_metrics) - 2 * sqrt(1 / day_1_sum_of_metrics + 1 / day_14_sum_of_metrics)
Big growth: Given the last 30 days, score = day_30_sum_of_metrics - day_1_sum_of_metrics
(You can keep Russell Crowe though...)
We've been trying to push Russell onto the yanks. Not much luck so far.
(I've removed xero from the db)
Also, no offence intended to Guy (the creator), but is it a coincidence that his Lifx startup is #1, and that that is heavily due to the StumpleUpon views which have jumped massively in the last week? Looks a little too convenient to me ;)