Data Mining: Text Mining, Visualization and Social Media

Twitter Trending Terms – Could Do Better

Twitter has oodles of data – millions of tweets a day. They have smart people working on this data, and they make all the right noises about social search. However, I’m looking at the trending topics on the site just now and I see these: MJ’s, Rip MJ, RIP Michael Jackson, Farrah Fawcett, #iranelection, Pop, Thriller, MTV, Iran, #michaeljackson.

I don’t get it. There are a number of problems here:

1. These aren’t topics, they are words or phrases. There are only 4 topics present here (Michael Jackson’s death, Farrah Fawcett’s death, the Iranian Election and MTV).
2. The ‘phrases’ present in the terms are pretty lame: ‘MJ’s’? Earlier today the phrase ‘Did Michael Jackson’ was a trending topic.
3. There is no attempt at normalization (RIP MJ == RIP Michael Jackson)
4. They actually are not at all interesting – anyone out there not know about Michael Jackson, Farrah Fawcett or Iran? Perhaps the MTV thing is a little more obscure.
5. They aren’t trending – ok, the RIP stories are, but the Iranian election? that’s been top of mind for many days now.

Twitter’s featuring of these topics and the immaturity of the technology in spite of the promise of the data does not paint a good picture for their prospects.

Be Trendy at your Peril

There was an interesting comment today from Rashmi Sinha of SlideShare during our panel at ICWSM. SlideShare (long story) has got onto the top trending terms on Twitter. It seems that when that happens, bots take over and start posting using the trending term and the true trend is buried under noise.

List of Twitter Search Engines

1. Twitter Search: a good base line. Note that it doesn’t provide search over URLs (they are all hidden behind the shortening services). I’m also suspicious of seeing duplicates in the results, but haven’t really looked to hard at this issue.
2. BackTweets: this is where I like to go for URL search over twitter.
3. IceRocket: I was surprised to see these guys still going and that they were doing something with Twitter. I’m not sure how they are getting data, but they seem to have quite a bit of it.
4. Scoopler: the latest thing on the radar (thanks Eric).
5. TwazzUp: not too dissimilar to Twitter Search
6. TweeFind: again, more like a clone of Twitter Search
7. Tinker: this is certainly more my kind of thing – trend graphs over events people are talking about on Twitter!

To me, the most interesting of these is BackTweets, Scoopler and Tinker.

For a broader list (including things a little orthogonal to search) check out Phil Bradley’s post Twitter Search which includes 20 services.

If you know of any other credible engines, please leave a comment (there are plenty of things out there that look like Twitter search engines, but aren’t).

About those twitter numbers : part 2

Briefly, I posted earlier about abandonment rates in Twitter. Nielsen has gone further and released a more detailed report.

Twitter Cascades

I saw a tweet from danah about a talk she had given. I replied to here indicating that I appreciated it. However, that reply – though a public act – has no context at all. You can’t determine which of danah’s tweets I’m referring to, and you can’t determine what it is I’m talking about.

Twitter, due to its character limits, does not do a good job of quotation. One can re-tweet (as about 2% of tweets currently are), but that is not particularly satisfying as one can’t add to the conversation.

Perhaps this is the first axiom of Twitter discourse structure: there are no explicit cascades.

Update – well, I really showed my ignorance there. I don’t use the txt version of Twitter, but was reading tweets as if I were. I would guess, however, that in the txt world, there is no ‘in reply to’ link – is that correct?

Note also that in search results, the ‘in reply to’ information is lost (I’ve always been confused as to why Twitter search doesn’t produce Twitter XML formatted results). This tweet is in reply to this one, but the search results don’t preserve that relationship. I see the ‘show conversation’ links in search results but, as Andrew Baron points out, it is only a feature of search. I’m guessing that there are still system divisions between search and the main Twitter CMS and that integration is going to be like changes the wheels on an aeroplane meloning down the runway.

@pete I’d love to see what you are doing re Twitter cascades. Leskovec et al et all have a paper on blog cascades which may make a useful comparison point: Finding Patterns in Blog Shapes and Blog Evolution.

About those Twitter Numbers

Every time Technorati published its state of the blogosphere posts it would come under fire regarding the number of blogs: 100 million you say? but how many are active, huh? At Blogpulse, we also suffered with this issue and side-stepped it by referring to the number of blogs that we had identified.

While everyone is crazy about those insane Twitter numbers, how many are active? What do we mean by active? Let’s say – if you haven’t tweeted in the past 5 days you are inactive.

Well, how on earth are we going to measure that? I came up with the cunning idea of doing a search way back in time on Twitter search for some term and then seeing how many of those that tweeted in the search results were still tweeting. Unfortunately, Twitter search only appears to provide results from about a month ago at best. So I went back to March 23rd – the oldest date I could get results for – and searched for ‘first tweet’.

Of the accounts for the first 15 results, 6 hadn’t updated on or after April 16th (40%); 9 hadn’t updated in the past 24 hours (60%).

That was a pretty unscientific experiment.

I then looked at Obama’s followers and found that 10 of the first 20 (50%) had no updates. However, I don’t think this means anything as I believe plenty of people sign up and follow Obama not knowing that he doesn’t Twitter. Perhaps followers are listed in order of most recently attached.

Obama has 904, 954 followers, and with 20 per page, that is 45, 248 pages. So if I go back to page 45, 000 I should be back with the older crowd. 5 of these 20 hadn’t updated in the past 5 days (25%), the next 20 also showed 5 inactive accounts, and the next 10 (50%).

Taking some of these inactive accounts at random, I found 3 of 8 (37.5%) followers inactive; 5 of 10 (50%); 8 of 20 (40%).

Update – a sample from Ashton Kutcher’s followers (aplusk) at page 50, 000, found that 13 of the 20 were inactive (65%).

Update – a sample from Britney Spear’s followers (britneyspears) at page 50, 000, found that 15 of the 20 were inactive (75%).

Conclusions – there’s no science here, so everything is with a pinch of salt. That being said, it seems you can’t move too far along the Twitter tree before you find some dead wood. Of course, I chose the 5 day limit because, with the first experiment, that is when something interesting happened. It’d be better to get a full distribution of inactivity for a good sample of accounts.

Twitter – the tide raises all boats

I carried out a simple experiment: grab a good chunk of Twitter data; pull out all the URLs, rank the domains and then track the trends in Compete.com. Below is the head of the list.

Rank	Domain	Monthly Growth	People
1	tinyurl.com	21.6%	4.5M
2	bit.ly	48.8%	1.2M
3	twitpic.com	54.6%	2.4M
4	is.gd	21.8%	0.7M
5	ff.im	25.5%	0.2M
6	twurl.nl	29.4%	0.3M
7	ow.ly	85.3%	0.6M
8	blip.fm	18.3%	0.3M
9	tumblr.com	14.8%	2.3M
10	plurk.com	19.9%	0.3M
11	tr.im	63.0%	0.4M
12	cli.gs	38.7%	0.3M

The data seems conclusive: cli.gs is definitely the  URL shortening service with the weirdest name.

Comparing Twitter and Blogosphere Latency

With all the bzzz around Twitter, and its apparent ability to walk on water right up to the brass ring of search, I’m still looking for the definitive study comparing discovery times in Twitter with those in other forms of social media. The study is pretty easy to perform:

1. Grab some Twitter data and extract some URLs
2. Use Twitter search to find the first mention of those URLs
3. Use a blog search engine to find the first mention in the blogosphere
4. Vice versa

This news article Defiant N Korea launches rocket first appeared on Twitter here 5 hours ago, but the earliest I can find it in the blogosphere is here 4 hours ago.

Endangered right whales appear to be on the rebound appeared a full day in the blogosphere before Twitter.

Neither the blogosphere nor Twitter cares about Ice bridge ruptures in Antarctic, and it seems neck-a-neck for Russia to unveil spaceship plans.

Of course, one would have to do a bit more work to really see what is going on. Twitter search doesn’t dereference URLs, so you can’t really search for the first mention of something, never mind the whole can of worms called URLs (the BBC usually has 2 URLs per story…).

And which stories do you want to check for anyway? Those that snowball into big piles of links may be less important (some signal will reach the user anyway).

Twitter search’s trends currently include ‘North Korea’, but then again, Google’s blogsearch has a top cluster for the same topic (and using the title of a post ‘North Korea launches rocket’ is more informative). Twitter also includes ‘DSI’, but Google has ‘Console Review: Nintendo DSi’.

Twitter in Web Search

It is interesting to see how twitter users are ranked in search engines. By searching for 'site:twitter.com' on live search, the top 5 accounts are:

1. Barack Obama 562k  followers
2. Jonathan Ross (Wossy) 169k
3. Biz Stone 248k
4. Pete Cashmore (mashable) 300k
5. ev 339k

Google has quite a different story. A search for 'site:twitter.com' produces the following accounts ranked from the top:

1. Andy Tarczon 213
2. Nadia W. 51
3. Ryan Stewart 4, 500
4. Zna Trainer 5, 281
5. Matt Haughey 3, 491

There appears to be a clear difference in ranking criteria. Googles' top ranked Twitter users are 2 orders of magnitude less 'popular' in terms of followers.

It is also interesting to look at the latency in the crawl. We can do this by looking at the cache pages for the accounts. Jonathan Ross' cached page on Live Search is from Feb 17th, Google's cache of Andy Tarczon is from March 2nd.

Web search engines are going to have to figure out what to do with Twitter. Certainly it is of huge value as a feature for ranking other content. Is it useful to index individual tweets? or accounts?

How to Improve Twitter Search

I’m starting to get Twitter as a publication mechanism, and I’ve always understood the value of mining the aggregate of tweets for any number of reasons. However, I don’t totally get Twitter search. In trying to, I’ve been playing around with it a little more. Here’s what I though:

• Why no link search? Most urls on Twitter go through TinyUrl or something similar, and yet these are not dereferenced. 23 % of tweets contain a url and so there is a big space of search targets for url search. I did a search for links to a blog post that I’d shortened and no results came up.
• Language classification doesn’t work. Of course, language classification is harder the shorter the text is, so Twitter is either asking for trouble by having it at all, or has some very cutting edge thoughts on how to do it. Plenty of languages can be bucketed by character encoding analysis (e.g. Korean, Japanese, Thai). However, do a search for something obscure and you’ll hit the language classification problem. E.g. a search for Elie Fife, my home town, produces no results (in the default English category), but a search over all language produces only English language results.
• Why no search counts? When results come back, we don’t know how many there were. For any type of rudimentary tracking, this count is important – at least for relative analysis.
• Add in time series. Time series in the results would be awesome. Some sites (like Twist from Flaptor) are attempting to provide this, but without the full data set, it is not clear how good or accurate these results are.
• Authorial search. Wouldn’t it be fine to search limited by features of the author? Certainly location is a no brainer.

I’m pretty sure that the lack of these features has nothing to do with the team’s ability to deliver them. It is likely a set of pragmatic issues to do with the value of these features in the context of their yet-to-be-decloaked business model.