Tag Archives: experiments

What Are The 10 Most Cited Websites On Twitter When Tweeting About Hot Trends?

Lately I wrote a post on how to build a relevant real time search engine prototype in few hundreds lines of code.  Using a tailored ranking algorithm based on link popularity in twitter,  I showed that the prototype was able to return very relevant answers in response to very hot queries like the ones that can be found in the hourly updated list of google hot trends.

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):

The Most Cited Websites When Tweeting About Hot Trends. 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.

Website Sample of 5 covered google hot trends of this past week
www.cnn.com 2011 budget
ipad tablet
concorde crash
federer murray
sports.espn.go.com federer tsonga australian open
aaron miles
tom brookshier
jackson jeffcoat
paul pierce
wpparty.com jackson jeffcoat
leon russell
wanamaker mile
buffalo exchange
recalled toyotas
www.huffingtonpost.com governor of virginia
obama republican retreat
obama gop
apple tablet announcement
groundhog prediction
twitpic.com miss america 2010 winner
what celeb do i look like
footprints in the sand
apple itablet
www.youtube.com i pad
grammy awards 2010
bob kellar
lakers celtics
ipad a disappointment
www.facebook.com general beauregard lee
roberta flack
action express racing
slightly stoopid
rolex 24 hours daytona
www.actionnewsblast.com codswallop meaning
jonathan antin
fred baron
codswallop definition
stevie nicks
www.netnewsticker.com arc energy
ego ferguson
kim burrell
ivan mccartney
mashable.com national lady gaga day
ipad tablet
ipad thoughts
doppelganger week facebook
tebow super bowl ad

Few remarks:

  • 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.

Collage of the most cited twitpic links in twitter for a week of google hot trends (Click 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 :).

miss america 2010 winner, what celeb do i look like, miss america 2010, roberta flack, lady gaga and elton john, addicted to love, jim florentine, apple itablet, lost season 6 premiere, candy crowley, to make you feel my love, swagger crew, footprints in the sand, gasparilla, miss virginia, duke georgetown, celebrity look alike, katherine putnam, itablet, andrea bocelli, monster diesel, peta ad.

How To Build A Relevant Real Time Search Engine Prototype In Few Hundreds Lines Of Code

gootterBy 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.

Comparing the results on Nexus One. Click 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).

byron de la beckwith
First ranked result by the prototype for the query "byron de la beckwith". Click to follow the 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.

The first link given by the prototype on "nexus one" about one day before its official presentation by Google. Click to follow the article

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 :).

Few weeks after the release of the Avatar movie, first links are a fast changing list of parodies

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).

You’ll find the source code here prototype_src.zip.

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.

Flexible Collaborative Filtering In JAVA With Mahout Taste

Mahout-logo-164x200 I recently had to build quickly a prototype of recommendation engine for a promising start-up company. I wanted to first test state of the art collaborative filtering algorithms before to build a customized solution (potentially on top of those algorithms). Most importantly, I wanted to be able to compare quickly all the different algorithm configurations with which I would come up with. Mahout Taste (previously a sourceforge project but recently promoted to the Apache Mahout project) was simply answering all those needs in one place.

I describe below how in few easy steps, I was set up to express my creativity without having to reinvent the wheel. This tutorial is based on the 0.2 release of Mahout.

Step 1: Set up your environment with mahout taste

I usually use Eclipse with Maven to simply add a dependency but the mahout pom had some repository issues by the time I tried, so I worked around it by adding the required libraries in eclipse manually (all the libraries found in the directory lucene/mahout/trunk/core/target/dependency of their latest release).

Step 2: Familiarize yourself by building a simple recommendation engine based on the movie lens data

To see a recommender engine in action, you can for instance download one of the movie Lens ratings data sets (I choose the one with one million ratings). Unzip the archive somewhere. The file that will interest you is ratings.dat. Its format is as follows:


The basic mahout taste FileDataModel only accept the simple following format:


There are many ways to transform your original file in that format, I used the simple following perl command:

perl -F"::" -alne 'print "@F[0],@F[1],@F[2]"' ratings.dat > ratingsForMahout.dat

You think: “what about the timestamp information???”. Yes, you right, it is a pretty crucial information given that it is based on temporal dynamics that the winning team of the Netflix prize made the difference (BTW, if you’re interested in the subject, you must see this video of Yehuda Koren’s lecture at KDD).

So, don’t worry, you can customize later your own DataModel class that parse any information you want, you’ll just have to implement the DataModel interface (you can also extends the FileDataModel class).

To obtain your first recommendations in few lines of code, you can use

import java.io.File;
import java.io.FileNotFoundException;
import java.util.List;

import org.apache.mahout.cf.taste.common.TasteException;
import org.apache.mahout.cf.taste.impl.model.file.FileDataModel;
import org.apache.mahout.cf.taste.impl.recommender.CachingRecommender;
import org.apache.mahout.cf.taste.impl.recommender.slopeone.SlopeOneRecommender;
import org.apache.mahout.cf.taste.model.DataModel;
import org.apache.mahout.cf.taste.recommender.RecommendedItem;

public class MahoutPlaying {
	public static void main(String[] args) throws FileNotFoundException, TasteException {
		DataModel model;
		model = new FileDataModel(new File("/home/padjiman/data/movieLens/mahout/ratingsForMahout.dat"));
		CachingRecommender cachingRecommender = new CachingRecommender(new SlopeOneRecommender(model));

		List recommendations = cachingRecommender.recommend(1, 10);
		for (RecommendedItem recommendedItem : recommendations) {


which creates in few lines of code a slope one recommendation engine and print the 10 first recommendations for user 1. You’ll see there only movieIds so you’ll have to check the file movies.dat to see the actual movie title (you can also write a simple method or script that shows you directly the movie title if you want to play with several users or to create your own user).

You can replace the slope one recommender with whatever other recommendation engine provided in the package. For instance, let’s say you want to use a classic user based recommender algorithm using the Pearson correlation similarity with a nearest 3 users neighborhood, simply replace the line that build the recommender in the above code by the code below:

UserSimilarity userSimilarity = new PearsonCorrelationSimilarity(model);
UserNeighborhood neighborhood = new NearestNUserNeighborhood(3, userSimilarity, model);
Recommender recommender = new GenericUserBasedRecommender(model, neighborhood, userSimilarity);
Recommender cachingRecommender = new CachingRecommender(recommender);

Few issues you might have during step 2:
– OutOfMemoryError: the slope recommender is pretty greedy and on the 1 Million Movie Lens Dataset, you may have to set the -Xmx VM option to 1024m (in eclipse, just add -Xmx1024m to the VM arguments in the run configuration options).
– Some errors during the FileDataModel initialization: pay attention that the directory containing your file to parse does not contains other files starting with the same name; for some reasons it disturbs the DataModel initialization in some cases.

Step 3: Test the relevance of the algorithms

In my opinion the most valuable part of the whole process. To feel immediately if your intuition of choosing a particular algorithm is a good one, or to see the good or bad impact of your own customized algorithm, you need a way to evaluate and compare them on the data.

You can easily do that with mahout RecommenderEvaluator interface. Two different implementations of that interface are given: AverageAbsoluteDifferenceRecommenderEvaluator and RMSRecommenderEvaluator. The first one is the average absolute difference between predicted and actual ratings for users and the second one is the classic RMSE (a.k.a. RMSD).

Since I’m playing with a movie dataset and that Netflix evaluation process was based on RMSE, I put here an example of use of the RMSRecommenderEvaluator:

import java.io.File;
import java.io.IOException;

import org.apache.commons.cli2.OptionException;
import org.apache.mahout.cf.taste.common.TasteException;
import org.apache.mahout.cf.taste.eval.RecommenderBuilder;
import org.apache.mahout.cf.taste.eval.RecommenderEvaluator;
import org.apache.mahout.cf.taste.impl.eval.RMSRecommenderEvaluator;
import org.apache.mahout.cf.taste.impl.model.file.FileDataModel;
import org.apache.mahout.cf.taste.impl.recommender.CachingRecommender;
import org.apache.mahout.cf.taste.impl.recommender.slopeone.SlopeOneRecommender;
import org.apache.mahout.cf.taste.model.DataModel;
import org.apache.mahout.cf.taste.recommender.Recommender;

public final class EvaluationExample{
	public static void main(String... args) throws IOException, TasteException, OptionException {

		RecommenderBuilder builder = new RecommenderBuilder() {
			public Recommender buildRecommender(DataModel model) throws TasteException{
				//build here whatever existing or customized recommendation algorithm
				return new CachingRecommender(new SlopeOneRecommender(model));

		RecommenderEvaluator evaluator = new RMSRecommenderEvaluator();
		DataModel model = new FileDataModel(new File("/home/padjiman/data/movieLens/mahout/ratingsForMahout.dat"));
		double score = evaluator.evaluate(builder,


Note that the evaluator need a RecommenderBuilder provided here as an inline implementation of the interface.
For a detailed description of the parameter of the evaluator, look at the javadoc in the sourcecode (as of today, the one that you’ll found on the web is outdated since it concern mahout release 0.1). But basically:
– 0.9 here represents the percentage of each user’s preferences to use to produce recommendations, the rest are compared to estimated preference values to evaluate.
– 1 represent the percentage of users to use in evaluation (so here all users).


RMSE = 0.8988.
To give you a point of comparison, the Netflix baseline predictor (called Cinematch) had an RMSE of 0.9514 and the Grand Prize was for the team providing an improvement of 10% (not that this tutorial is not based on netflix data but on Movie Lens data).

The number not really matters here: the important thing is that it provide you an easy way to compare different algorithms or the same algorithm with different settings (thresholds or other parameters).

Step 4: Now start the real work…

You guessed that you won’t win any Prize using the recommenders given by Mahout as-is :).
Depending on your data and on your needs, you may have either to simply customize an existing algorithm or to plug any specific similarity measure or to create your very own recommender from scratch. All of those are pretty easy to do in Mahout.

Let’s say for instance that you want to exploit the category of the movies to build a specific user similarity that includes this information.
What you will have to do is first to be able to capture the new information about categories.

To do so you can for instance extends the FileDataModel class to another class that also parses the movies.dat file and build relevant data structures to store the data about categories. I found more convenient to build my own Statistics object. Then you will have to build a new User similarity. It is as simple as that:

import org.apache.mahout.cf.taste.common.Refreshable;
import org.apache.mahout.cf.taste.common.TasteException;
import org.apache.mahout.cf.taste.model.DataModel;
import org.apache.mahout.cf.taste.similarity.PreferenceInferrer;
import org.apache.mahout.cf.taste.similarity.UserSimilarity;

import com.padjiman.algo.Statistics;

public class ProfileSimilarity implements UserSimilarity {

	private final Statistics stats;
	private final DataModel dataModel;

        public ProfileSimilarity(Statistics stats, DataModel dataModel) {
		if (stats == null) {
			throw new IllegalArgumentException("stats is null");
		if (dataModel == null) {
		      throw new IllegalArgumentException("dataModel is null");
		this.dataModel = dataModel;
		this.stats = stats;

	public double userSimilarity(long userID1, long userID2)
	throws TasteException {
		//build your similarity function here
		//exploiting the stats and dataModel object as you wish

	public void refresh(Collection alreadyRefreshed) {
		// TODO Auto-generated method stub

	public void setPreferenceInferrer(PreferenceInferrer inferrer) {
		// TODO Auto-generated method stub

Complete the method userSimilarity with you own secret sauce. Et voila: you can now plug your new user similarity measure in a GenericUserBasedRecommender, for instance instead of the Pearson correlation similarity measure (showed in step 2) and simply compare which one is the best using your evaluator.

You’re not satisfied with the GenericUserBasedRecommender or any other recommender provided by Mahout? No problem, try to implement your own. You’ll just have to start with a class declaration of this kind:

public class MostPopularItemUserBasedCombinedRecommender extends AbstractRecommender implements Recommender {
       //override the necessary methods

Here again, you can use as member of the class any customized object containing any statistics that you would judge relevant to build a better recommender. Then, again, plug your new recommender in the evaluator and compare, experiment, improve.


Mahout Taste is a very flexible platform to experiment collaborative filtering algorithms. It certainly won’t provide you an immediate solution to your recommendation problem, but you’ll be easily able to either tune the existing algorithms or plug your own creative ones into the mahout taste interfaces set.

By doing so, you’ll immediately get the benefit of a platform allowing you to compare, tune and improve iteratively the results of your different algorithm configurations. Last but not least, Mahout taste provide an external server which exposes recommendation logic to your application via web services and HTTP.

Other ressources:

  • After reading this quick start guide, I recommend you to have a look at the official mahout taste documentation. As of today it is not updated with the release 0.2 so you might find some old method signatures there but you’ll find useful and complementary information about the big picture of Mahout Taste design.
  • A nice article on mahout in general (not only the taste part). I felt that Taste was not enough detailed there, in particular on the testing part, this is why I wrote this tutorial.

Drawing A Zipf Law Using Gnuplot, Java and Moby-Dick

whaleThere are many tools out there to build more or less quickly any kind of graphs. Depending on your needs a tool may be more suited than another. When it comes to draw graphs from a set of generated coordinates, I love the simplicity of gnuplot.

Let’s see together a simple example that explain how to draw a zipf law observed on a long english text.
If you’re not familiar with zipf law, simply put it states that the product of the rank (R) of a word and its frequency (F) is roughly constant. This law is also know under the name “principle of the least effort” because people tends to use the same words often and rarely use new or different words.

Step 1 : Install gnuplot

For mac, check this.
For linux, depending on your distrib it should be as simple as an apt-get install (for ubuntu you can check this howto).
For windows you can either go the “hard” way with cygwin + X11 (see Part 1,4 and 5 of those instructions) or the easy way by clicking on pgnuplot.exe located in the gpXXXwin32.zip located here (this last solution may be also easier if you want to have copy/paste between the gnuplot terminal and other windows).

Step 2: Generate the Zipf Law data using Java and Moby Dick!

As I told you above, gnuplot is particularly simple for drawing a set of generated coordinates. All you have to do is to generated a file containing on each line a couple of coordinates.

For the sake of the example, I will use the full raw text of Moby Dick to generate the points. The goal is to generate a list of points of the form x y where x represents the rank of the word (the more frequent the word is, the higher its rank) and y represents its number of occurrences.

Find below the java code I used to do that. If you want to execute it, you will need lucene and the google collections (soon to become part of guava) libraries.

import java.io.File;
import java.io.FileReader;
import java.io.IOException;
import java.util.ArrayList;
import java.util.Collections;
import java.util.Comparator;
import java.util.List;

import org.apache.lucene.analysis.Token;
import org.apache.lucene.analysis.TokenStream;
import org.apache.lucene.analysis.standard.StandardAnalyzer;

import com.google.common.collect.HashMultiset;
import com.google.common.collect.Multiset;
import com.google.common.collect.Multiset.Entry;

public class ZipfLawOnMobyDick {
	public static void main(String[] args) throws IOException {

		//Multiset for storing word occurrences
		Multiset multiset = HashMultiset.create();

		//Creating a standard analyzer with no stop words (we need them to observe the zipf law)
		String[] STOP_WORDS = {};
		StandardAnalyzer analyzer = new StandardAnalyzer(STOP_WORDS);

		//Initializing the multiset by parsing the whole content of Moby Dick
		TokenStream stream = analyzer.tokenStream("content", new FileReader(new File("C:moby_dick.txt")));
		Token token = new Token();
		while ((token = stream.next(token)) != null){

		//Sorting the multiset by number of occurrences using a comparator on the Entries of the multiset
		List> l = new ArrayList>(multiset.entrySet());
		Comparator> occurence_comparator = new Comparator>() {
			public int compare(Multiset.Entry e1, Multiset.Entry e2) {
				return e2.getCount() - e1.getCount() ;

		int rank = 1;
		for( Multiset.Entry e : l ){

This will generate the following output (the set of coordinates) that you can put in a file called moby_dick.gp. If you’re curious about what are the 100 hottest keywords of the whole text you can check them here.

Step 3: Drawing using gnuplot

What you can do first is simply to type the following command in the gnuplot console (you have to be on the same directory as the moby_dick.gp file):

plot [0:500][0:16000] "moby_dick.gp"

It simply draws the points and rescale the range of x and y respectively to [0:500] and [0:16000] so we can see something.
Play with the ranges to see the differences.
If you want the dots to be connected, just type:

plot [0:500][0:16000] "moby_dick.gp" with lines

If you want to add some legends, you can put some labels and arrows.
Here is an example of a gnuplot script that will set some information on the graph (you can simply copy/paste it in the gnuplot console):

set xlabel "word rank"
set ylabel "# of occurrences"
set label 1 "the word ranked #14 occurs 1753 times" at 70,4000
set arrow 1 from 65,3750 to 15,1800
plot [0:500][0:16000] "moby_dick.gp

As you can see it is pretty straightforward. You can play with the coordinates to adjust where to put the labels and arrow.
You will obtain this graph (click to enlarge):


To export it as a png file just type:

set terminal png
set output "moby_dick.png"
plot [0:500][0:16000] "moby_dick.gp"

You also might want to try a log scale on the vertical axis as to not waste the majority of the graph’s scale (thanks Bob for the remark).
To do so, you can simply type in the gnuplot console:

set logscale y

by plotting within the range [1:3000][5:10000], you’ll obtain:


Finally, you might want to use a log-log scale that are traditionally used to observe such power laws. Just set the logscale for x as you did for y and you’ll obtain:


You can of course add as much eye candies as you want (the demo page of the gnuplot website gives tons of example).

Also, there are probably dozens of ways to draw the same thing, I just loved the fun and simplicity of that one.

Google Hot Trends Clustering: The 100 Hottest Queries Tell You About 67.76 Stories In Average

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.

As an example of output, see the file generated for the 100 keywords studied in this post.

What It Is Useful For?

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 :).

Can You Guess What Is The Hottest Trend Of Google Hot Trends ?

screenshot019Either 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.