Hacker News Reader

Comment by gus_massa | Hacker News Reader

Correlation does not imply causation.

Alternative theory (without any data :)):

Perhaps most of the photos with 30 tags are dew to semiautomatic fill-all-the-keywords strategies in SEO attempts. So most of them have naturally more exposure than photo from individuals that only use a few tags (like “dog” and “sleeping”).

Then they get more likes because they have more traffic, not because they have more tags

Proposed experiment:

* Pick 70 new photos, and separate them in 7 groups of 10 photos at random.

* Post them with 0; 5; 10; ... ; 30 tags and wait ...

* Count the likes for each group

* Kill outliers? 10 is too small? Find someone with more knowledge of statistics to check the experiment.

View on HN

But where did they get more traffic from to begin with? You're implying they got traffic from "fill-all-the-keywords strategies", which is more-or-less what the analysis is advocating. That may not be true, but your alternative hypothesis isn't actually an alternative. At least not as I understand it.

That said, I do think controlling for follower counts would be informative. It would help tell you if there's actually something else at play (e.g. once you control for follower count, there's no more correlation). I suspect it could also tell you the # of hashtags is even more important than currently shown because it also helps build your base of followers, by reaching more new visitors.

E.g. If increased hashtags increase my number of favorites, they also increase my number of followers by some percentage. Then each successive post has a larger direct reach of followers, in addition to the new visitor reach from hashtag discovery. A single hashtag could actually have a compounding value.

My alternative theory might explain it: Bad actors who stuff their photos with 30 keywords also use shill accounts to like them.

My comment was not very clear. I’ll try again.

For simplicity let’s suppose that there are only two types of users:

* Some users have an account in every social network, have a very popular blog, perhaps pay for some advertisement, fill in all the keywords / tags fields, and use all the SEO they can.

* Some users take the photo, and just copy the nouns of the title to the tags field.

If this model is correct, most of the photos with lots of tags are from users with lots of followers and that is the cause of the additional likes.

The original (implicit) idea of the article is that the photos with more tags appear in more search results, and this is the cause of the additional likes. [I hope this simplified version is an accurate enough.]

Real life is usually more complicated, there are more than two types of users, and perhaps both effects are present. I like the analysis of the data, but I’m not sure if this can be transformed directly in a strategy to gain followers. Perhaps this is a consequence of cargo cult SEO.

How can we setup an experiment to distinguish between the two effects?

Just keep in mind an R^2 of 15 isn't that high, so these variables are only accounting for a small portion of the variance in the model.

I think there's an important control here: number of followers per user. It'd be interesting to see how followers would affect the model, especially because you could assume that "knowing to add more hashtags" likely correlates with "use of the platform," which they likely correlates with "increased # followers," who can then produce more likes on each photo.

EDIT: Another thought: if you're looking at # likes, you might want to go for a negative binomial model instead of an OLS regression, because it'll account for the dependent variable being a "count" measure.

Well, what is your expected outcome? If you're just measuring to see what happens, you're not really going to know unless you do a larger experiment. If you get more specific it's easier to tell what exactly you need to do in order to get an accurate result.

If somebody repeats that cliche in person to me any time soon, I will stick a fork in their head. And I thought the "curse of dimensionality" and "embarassingly parellizable" were annoying.

edit: Although, yes, certainly a valid point.

Reply on news.ycombinator.com