Trending Topics: The Good and the Bad

January 4, 2010

Trending topics are becoming a big deal, but they aren’t without their problems. It seems like exactly the kind of tool that should help people solve their issues with information overload, but so far, current implementations are not up to this task. This has led some to prematurely declare them useless. I think they show much promise and that it’s too early to judge.

It’s probably best to start with the problem statement. Trending topics are a distillation of the popular things in some data set, usually during a certain time frame. For example, Google publishes their Year-End Zeitgeist reports with the topics that were the fastest rising or falling through an entire year. On the other end of the spectrum, my company Collecta and sites like Twitter try and distill trends on much shorter timescales.

The Importance of Time

Clive Thompson lamented that there were no insights to be had in Twitter’s top trends of 2009, which included topics like Tiger Woods and Michael Jackson:

These subjects are all screamingly obvious, each having been long ago chewed into a tasteless cud by the 24-hour news cycle.

This is, of course, absolutely true. This is also the way it is supposed to work. When you ask the question “what were the most important trends of 2009?”, you are going to get an answer that includes mostly stuff that you remember. Once something is big enough to be a trend in the context of an entire year on a large service, it probably received lots of coverage in mainstream sources and probably showed up in pop culture references on TV.

The problem is that this statement is not true when you look at these trends in the context of time. When Michael Jackson died, social media sites received millions of updates and comments as people expressed their feelings. This immediately made it a trending topic. This was not blindingly obvious before it happened.

There are now lots of stories that break on social media sites like Twitter before the mainstream press covers them. Some of these events trend before the general public is aware of them, and in these cases, the trending topics fulfill some of their potential.

Without the context of time, everyone should know the trending topics; that’s why they trended in the first place.

Geo-centrism

It’s no secret that mainstream news in the US is very domestically focused. A news story about a celebrity is likely to displace real news from elsewhere in the world. The same situation exists in most countries, I imagine.

Fortunately, the Internet is a global community, and this can be seen in trend data. For example, Iran was trending on social media sites before it got much coverage in mainstream press.

For an Iranian, or someone with Iranian friends and family, news of Iran is likely to be extremely important. These kind of trends show up very quickly on the Internet, which has participants from the far corners of the Earth.

If you look at the trend data for a site like Twitter or Google and compare that data to US magazines and newspapers, I think the difference will be one of geography. Many topics will surely overlap, as they are of general interest, but where the topics are disjoint will reveal local bias.

Select Data Sets

I think trending shines most brightly when it can be arbitrarily applied to various data sets. The fact that Iran is a trending topic is interesting, but if I really want to understand what’s going on, I want to know what is trending within that topic.

Similarly, I may want to see what is trending among my friends or among people within my city. Or to use an example close to my own work, you might want to know what is trending related to arbitrary keywords.

This eliminates a lot of the triviality of current trending. I wouldn’t expect to see a trending topics list for 2009 among my friends to look very much like the global one. It’s likely to contain a lot of technical topics and be less focused on entertainment.

Applying this to arbitrary keywords will yield some interesting surprises. What were the trending topics for your company or your product? Even if you’ve read every post people have written about the subject, you may fail to see the general trend.

Naive Algorithms

The main flaw with current trending topics implementations are that they are fairly simply algorithms that are easily gamed. These algorithms are the low hanging fruit; they tend to do a reasonable job if you finesse them, but they aren’t sophisticated.

Bradford Cross recently wrote an awesome piece on the current state of various AI techniques. One of the parts that resonated with me was “Pandora is not the state of the art.” It may seem like magic, but behind the scenes the learning algorithms are shallow and not scalable.

We are not up to the “state of the art” in trending yet, either. Luckily, unlike Pandora, most people think the current trending implementations are not great. This means companies will be pressured by users to investing in the research and design necessary to improve these systems.

I think it’s much too early to write off trending topics as useless. There are some real promising features that are on offer including faster detection, removing local bias, and applying trends to arbitrary data sets. However, it’s going to take some real work to get where people expect it to be.