# “Interestingness” function for StackExchange questions

I am trying to put together a data-mining package for StackExchange sites and in particular, I am stuck in trying to determine the "most interesting" questions. I would like to use the question score, but remove the bias due to the number of views, but I don't know how to approach this rigorously.

In the ideal world, I could sort the questions by calculating $\frac{v}{n}$, where $v$ is the votes total and $n$ is the number of views. After all it would measure the percentage of people that upvote the question, minus the percentage of people that downvote the question.

Unfortunately, the voting pattern is much more complicated. Votes tend to "plateau" to a certain level and this has the effect of drastically underestimating wildly popular questions. In practice, a question with 1 view and 1 upvote would certainly score and be sorted higher than any other question with 10,000 views, but less than 10,000 votes.

I am currently using $\frac{v}{\log{n}+1}$ as an empirical formula, but I would like to be precise. How can I approach this problem with mathematical rigorousness?

In order to address some of the comments, I'll try to restate the problem in a better way:

Let's say I have a question with $v_0$ votes total and $n_0$ views. I would like to be able to estimate what votes total $v_1$ is most likely when the views reach $n_1$.

In this way I could simply choose a nominal value for $n_1$ and order all the question according to the expected $v_1$ total.

I've created two queries on the SO datadump to show better the effect I am talking about:

Average Views by Score

Result:

Average Score by Views (100-views buckets)

Result:

Results, not sure if straighter is better: ($\frac{v}{n}$ in blue, $\frac{v}{log{n}+1}$ in red)

-
This certainly is an interesting question, but I think you might be better off asking this on stats.SE. –  Theo Buehler May 3 '11 at 22:05
@Theo You may be right, actually. I'll flag for the mods to migrate if they think it's best. –  Sklivvz May 3 '11 at 22:08
Why would views not contribute to interesting-ness? (but worse, why would they contribute negatively?) More interesting things tend to be viewed more often... The fundamental problem here is what does interesting even mean? Does it means questions of general interest or questions that are of interest to a more specific higher level audience? For someone to answer this question with "mathematical rigourousness", it needs to be posed rigorously first. –  Eric Naslund May 3 '11 at 22:11
Views bias the questions because one question might, say, be link by a good site and receive tons of views--if you look at the top rated questions they are all high view questions; by interesting I mean the questions that have more value as perceived by the users of the site. In any case, the question still stands: what is the correct way of combining views and votes to get the best predictor of quality? –  Sklivvz May 3 '11 at 22:14
The math people asked good questions. This question's logic seems circular: it appears to ask us for a formula to measure the "quality" of an SE question but it doesn't stipulate what "quality" means except to give non-operational synonyms like "value as perceived by the users of the site." You can't get something for nothing! –  whuber May 3 '11 at 22:29

## migrated from math.stackexchange.comMay 3 '11 at 22:15

This question came from our site for people studying math at any level and professionals in related fields.

One might define an interesting question as one that has received comparatively many votes given the number of views. To this end, you can create a baseline curve that reflects the expected number of votes given the views. Curves that attracted a lot more votes than the baseline were considered particularly interesting.

To construct the baseline, you may want to calculate the median number of votes per 100-view bin. In addition, you could calculate the median absolute deviation (MAD) as a robust measure for the standard deviation per bin. Then, "interestingness" can be calculated as

interestingness(votes,views) = (votes-baselineVotes(views))/baselineMAD(views)

-