I wrote a small program on top of this prototype to run an experiment: each hour, the program crawl the new list of hot queries from google hot trends, then it runs the prototype on each of those queries and keep the hottest link found in twitter for the corresponding hot query. I wanted to see which websites were mostly cited in those tweets talking about hot trends.
So I let ran the program for a week, collected the links (more than a thousand), expanded all those into their long URLs version (using an improved version of my java universal URL expander), extracted the domain names and compiled the whole into a top 10 list of the most cited websites. Here it is (click to enlarge):
I was surprised to see some websites that I’ve never heard about before (like wpparty.com or actionnewsblast.com).
To have a better idea for which kind of hot queries/topics those websites are most cited in twitter, find below, for each of those top website, a sample of 5 google hot trends query they covered last week.
Sample of 5 covered google hot trends of this past week
All the links spotted by my prototype and that appear in the table are coming from real tweets around those google hot trends queries.
You’ll notice that apple iPad announcement is a theme that was covered by 4 of those top 10 websites!
I recommend you to have a look on the youtube video in the table around the google hot trend “ipad a disappointment” :).
I also recommend you to have a look at the haiti 360 view covered by cnn.
For twitpic, it is only pics, so what you’ll find there is a sample of “trendy pics” (see below for more on that…)
Sometimes the hot query seems to be not connected with the related article at first view (like with fred baron). But when you take a closer look, there is always a connection! This is not for nothing that people tweet about a link with the text of the hot query in the tweet…
To finish, find below a picasa collage that I built using the most cited twitpic pictures in twitter for this past week of hot trends (not only the 5 cited in the table). You’ll identify easily some sarcastic pictures before the iPad announcement or pics around the election of Miss USA. Click the picture to enlarge.
If you’re curious to map some pictures with its related hot topic, click the collage to enlarge it and try to guess which pics correspond to which google hot query below :).
By the end of the post you’ll find the code along with a small command line JAVA program to play with, but let me first describe the specifications of the real time search engine prototype that I’m targeting here.
Basically it should take as input a search query and return as output a ranked set of URLs that would correspond to the latest hot news around that search query.
In some way it is similar to what you would expect to find on google news or in one of the dozens real time search engine that were released last year (let’s cite oneriot, crowdeye and collecta).
The goal of my prototype is to demonstrate how to leverage twitter and a simple ranking algorithm to obtain most of the time relevant URLs in response of hot queries, without having to crawl a single web page! As my primary target is relevancy, I won’t invest any effort on performance or scalability of the prototype (retrieved results will be build at query time).
High level description of the prototype
Basically what I did is to use the twitter API through a java library called twitter4j to retrieve all the latest tweets containing the input query and that contains a link. For very hot queries, you’re likely to get a lot of those (I put a limit of the last 150 but you’ll be able to change it). Once I got my “link farm”, what I do is to build a basic ranking algorithm that would rank first the URLs that are the most referenced.
As most of the URLs in tweets are shortened URLs, the trick is to spot the same URLs that were shortened by different shortening services. For instance both of the following shortened URLs points to a same page of my blog: http://tinyurl.com/yajkgeg and http://bit.ly/SmHw6. It can sounds as a corner case but it actually happens all the time on hot queries. So the idea is to convert all the short URLs in their expanded version. To see how to write an universal URL expander in JAVA that would work for the 90 + existing URL shortening services check the post that is referenced by the two short URLs above.
Note that you can improve the ranking algorithm in tons of way, by exploiting the text in the tweets or who actually wrote the tweet (reputation) or using other sources like digg and much more, but as we’ll see, even in its simplest form, the ranking algorithm presented above works pretty well.
Playing with some hot queries
To find some hot queries to play with, you can for instance take one of the google hot trends queries (unfortunately down from 100 to 40 to 20). Let’s try with a very hot topic while I’m writing this post: the google Nexus One phone that was about to be presented to the press two days after I started to wrote this post.
Below I have compiled the results obtained respectively by Google News, OneRiot and my toy prototype on the query “nexus one”. Click the picture to enlarge.
I hope you enjoyed my killer UI :). But let’s focus on the three URLs corresponding of the first result of each one:
Given the fact that at the time I issued the query, the Nexus one was not yet released, I would say that the article that the prototype found is the best one since it is the only one that present an exclusive video demonstrating the not yet released phone. This is also why so much people were twitting about this link: because it was the best at that precise time! We’ll see even more in the next section.
Before, let’s try with another hot query today (in the top 20 hottest queries of google hot trends): “byron de la beckwith”.
That time, it is not clear what is the story/news hidden behind that hot query but running it on the prototype gives as the first link the article below (click on the picture if you want to see the full article).
Again this is a very relevant result (oneRiot and Google News gave the same one at that time).
The temporal aspect of hot queries
What is interesting with hot queries is that you expect the result to change even within a short amount of time. Indeed, any story or breaking news generally evolve as new elements comes in. As promised let’s follow our “nexus one” query.
In the previous section, the prototype’s first result was a very relevant article from engadget. I relaunched the same query, but after 12 hours. The first ranked result returned by my prototype gives me now a different result: still another article from engadget (see picture below), but that time with a much more in depth review of the phone with more videos including a very funny comparison between the android, iphone and nexus one.
Then I waited for Google doing its press conference one day later. I issued the query again. Can you guess what was the first link given by my prototype? You got it, the official Google Nexus One website.
Again this is not a corner case. This temporal aspect happens all the time, for any type of breaking news or events. As a last example of that phenomenon, let’s take the movie avatar. The first days before and after that the movie were released, all you got is links to see the trailer or even the movie. Now, few weeks after, what you get is a very fast changing list of links around fun pictures of parodies of the movie with title like “Do you want to date my avatar” (picture below) or a letter attempting to prove that avatar is actually Pocahontas in 3d :).
Playing by yourself with the prototype
If you just want to run the prototype through the command line
You must have java 6 installed (you can check by opening a console and type java -version). On recent mac, see those instructions for having java 6 ready to use in a snap.
Then just download this zip archive: jarsDependencies.zip.
Save it and extract it somewhere in your computer. It will create a directory named prototypeJars.
Open a command prompt. Go inside the directory prototypeJars.
If you are on windows, just type:
java -cp "*;" com.philippeadjiman.rtseproto.RealTimeSEPrototype "nexus one" 150 OFF
If you are on Linux or Mac just type:
java -cp "*" com.philippeadjiman.rtseproto.RealTimeSEPrototype "nexus one" 150 OFF
You’ll notice the three last arguments (all are mandatory):
“nexus one”: is the query. Type whatever you want here but keep the quotes.
150: is the maximum number of tweets to retrieve from the timeline. Put whatever number between 1 and 1000 but 150 is good enough.
OFF: whether or not you want the prototype to expand the short URLs. If you put ON, you should be patient, it may take a while. Even if duplicate short URLs happen all the time, going with OFF gives a good approximation of which are the leading results. Unless a problem with Twitter, putting OFF should provide you the results within few seconds.
Only the top 20 first results will be printed.
If you want to play with the code
As the title suggests, that just few hundreds lines of (JAVA) code. As it is a toy project and to keep things simple I voluntarily didn’t use any DI framework like spring or guice and tried to use as less external libraries as possible unless necessary (even no log4j!). I did wrote a minimal amount of unit tests since I cannot code without it and I did use the google-collections library for the same reason :).
Also I tried to wrote at least a minimal amount of comments, in particular where I think the code should be improved a lot for better performance but remember: the prototype is of course not scalable as it does not rely on any indexing strategy (it computes the results at query time). Building a real a real search engine would at first involve building an index offline (using lucene for instance).
If you are using maven and eclipse (or other popular IDE), you should be ready to go in less than a minute by unpacking the zip, typing “mvn eclipse:eclipse” and importing the existing project.
Some final remarks
What I wanted to prove here is mainly that without crawling a single webpage, you can answer to “hot queries” with a relevancy comparable to what you can find on google news or any “real time search engine”. This is made possible by judiciously using the tremendous power that twitter provide with its open API.
Of course building a real “real time search engine” would require much more than few hundred lines of code 🙂 and hundreds of features could be added to that prototype, but I would keep two core principles:
real time search results should be links and not micro blogging text like tweets. The text of some tweets can be relevant but as a secondary level of information.
let the “real time crowd” do the ranking for you. If a link is related in some way with your query and was highly and recently tweeted or digged (you name it), then there is a good chance that it will be a relevant “real time” result.
In that sense, among the dozens of real time search engines I have tested, my favorite one remains oneriot.
This is for the “pull” side of the things (when the user knows what to search for). I did not talk about the “push” side of the real time web here, probably in another post…
If you have issues running the prototype or any other question/remark, please shoot a comment.
Did you noticed that among the 100 (hourly updated) Google Hot Trends, there are always several hot queries that are related one to the other?
Let’s take a look at the Hot Trends of the current hour by the time I’m writing this post: Hot Trends of September 24 at 11PM PST Time (clicking on the keywords won’t work, it is just a local copy of the file at that time). In few seconds, we can spot some similar queries, for instance Hot Trend #5 “sean salisbury” is clearly related to Hot Trend #45 “sean salisbury internet postings” and also to Hot Trend #57 “sean salisbury cell phone incident” (click the picture to enlarge).
Now, a small quizz: is there a link between Hot Trend #48 “julia grovenburg” and Hot Trend #8 “superfetation”, and what the hell is “superfetation”??.
So first, yes, there is a link between those two queries, and you can discover it if you click on “superfetation” which will give you its related searches:
So if you had time to loose, you would be able to click on the 100 queries and use this method to eventually build this cluster of 8 queries:
The words in the cluster can give more insights of what this story is all about: Julia Grovenburg was pregnant and was pregnant again (apparently during the same pregnancy) which is a phenomenon called superfetation. You can verify it on a news article of the same day:
Looking at the cluster, you can also think that the baby after birth was a “19 pound baby” but actually this a completely different breaking news, not linked at all with the previous one. This misleading link shows that related searches is a great feature but not an exact science and sometimes (not often however) some errors can arise in related searches:
I have some intuitions about how those related searches are detected and how those errors happens. It’s beyond the scope of this post but if you are interested about it, shoot me an email.
So I implemented a link-based clustering algorithm that knows how to plug to google hot trends data ant that build all that stuff automatically. Two queries are in the same cluster if one of the 3 following conditions is true:
the queries themselves are similar
one of the query is similar to one of the related searches of the other
one of the query related searches is similar to one of the related searches of the other
I used a similarity measure that works well for short text like queries, along with a black list of words to not disturb the similarity with words like “the” or “a”, etc… . I also empirically determined different thresholds for the three different cases described above. If you have more questions about that stuff, feel free to shoot a comment or to contact me.
So How Many Clusters Can I Build Out Of The 100 Google Hot Trends Queries?
You got it from this post title: 67.76 clusters in average (based on crawled data that represents few months of hot trends). Each cluster is supposed to represent a same “story” or breaking news. Note that this number is also dependent of my thresholds and that other algorithms and/or thresholds (more or less strict) can obtain slightly different numbers.
Of course, some errors can also arise, either because of some misleading related searches (like showed above) or because is some cases two queries look very similar but in reality they are speaking about two different things.
First of all it is fun :). Second, in information retrieval, order is always better than the opposite. But much more than that: if you are a breaking news website or blog, you’d better use in your article all the keywords of the same cluster since they represent the hottest searched queries of that particular story represented in its cluster! From an SEO point of view, I think the interest is pretty clear.
If you read the post up to here, I’d like to offer you a small bonus :). It is the HUGEST cluster that I was able to observe running my program on the last few years of google hot trends data. I think you already guessed to which breaking news it is related. Check it out!
Update: Coincidence, the day after I wrote this post the hot trends list was reduced from 100 to 40, so the screenshots and data above are in souvenir of the older version :).
Either if you are working in SEO, or if you are a “trends hacker”, or if you love like me doing useless comparisons like hanukkah vs passover, you obviously know the fantastic google trends tool.
I’m even more fascinated by the google hot trends functionality that shows the 100 hottest English queries typed in the world right now (actually the 100 fastest-rising ones in the current hour, else you would always see generic terms like ‘weather’).
I asked myself a simple question: is there some queries that always appearing over and over in this top 100 list? Can we discover patterns of queries? To answer it, I write for fun a simple crawler to crawl the daily list since the service exists (May 15, 2007) and I generated a list of the hottest phrases (meaning the hottest n-grams of words, not queries).
Can you guess if there is a clear winner?
Actually there is one. The phrase “lyrics”. As of today (August 31 2009), it always appears to be the most frequent hottest keyword in different settings:
759 occurrences if you consider the whole daily top 100 list. Think about it: since May 15, 2007, it’s been 809 days (thanks Jeffrey). Even if it appears sometimes several times in a single day, it means that almost everyday, the word lyrics appears in the 100 hottest English queries in the world!!!
207 occurrences if you consider only the daily top 10 list.
124 occurrences if you consider only the daily top 5 list.
34 occurrences if you consider only the daily hottest keyword.
But again, ‘lyrics’ is always the top ranked phrase of all the lists I generated. Seems however like a decreasing trend.
What about other phrases? Here are few other examples of the top phrases appearing over and over in all day top world queries. Note that you don’t necessarily want to build a business around one of those hot topics since all of them are in general already overcrowded niches.
What about patterns? If you perform some entity extraction you can observe some recurring patterns like ‘XXX death‘ or ‘XXX divorce‘ where XXX is the name of a celebrity. I also noticed that users are much more interested in celebrities divorces than marriages :).
In summary, Google hot trends is fun. In the new real time web buzz, this service is not really meant to be a competitor, but it is still my favorite way of feeling the pulse of the web.