The MapD GPU-Powered Interactive Tweetmap: Lightning Fast Social Media Analytics
Download OmniSci Free, a full-featured version available for use at no cost.
While we love datasets of all shapes and sizes at MapD, Twitter holds a special place in our hearts. This is perhaps because we find Twitter data to be almost peerless among public datasets in its ability to provide a glimpse into the human experience - revealing what people are saying when and where. Twitter is powerful in that it provides insight into a wide variety of social phenomena both at the level of individual tweets as well as rolled up by user, geography or topic/hashtag.
Twitter is also meaningful to me personally in that my graduate work involved using Twitter to examine the Arab Spring in Egypt, and it was the inability of existing tools to allow interactive analysis of hundreds of millions of tweets that spurred me to start building the GPU database that would become MapD. Waiting for hours for my batch jobs to finish, I dreamed of a tool that would allow for real-time hypothesis generation, testing and iteration. The tool would need to allow both roll-ups over billions of data points and drill-downs to individual records - and would need to accomplish this over high-throughput streaming data. The only way to achieve these goals was to harness the massive parallelism and memory bandwidth of graphics processors, or GPUs - essentially bringing supercomputing to data analytics.
So today we are proud to announce the unveil MapD's new Tweetmap demo. This demo leverages the power of MapD’s visual analytics platform in the following ways:
- All charts are driven by MapD’s lightning fast GPU-powered SQL engine that delivers response times measured in milliseconds across multi-billion row datasets and hence immersive interactivity.
- While most of the charts are rendered on the frontend (from the results of SQL queries processed on the backend), the map itself is rendered directly by the backend GPUs. It would be essentially impossible to transfer the data needed to render millions of points between the backend server and the client web-browser, but by executing the requisite SQL query and rendering the result all on the backend GPUs, all that needs to be sent to the client is a compressed PNG.
- The charts themselves are generated via our frontend charting API. This means that a customer could create their own presentation-quality interactive dashboard in a few hours of programming time (we also provide a drag-and-drop interface, MapD Immerse, that will be covered in another post).
Now on to the demo itself.
First: the dataset. You are looking at approximately 60 million geocoded tweets from November 2014 to February 2015. Geocoded tweets make up approximately 1-2% of the entire Twitter feed, and are typically sent from smartphones with the geolocation turned on. The geocoded tweets shown in the demo are themselves from a 10% sample of the entire Twitter stream (or “firehose”) – so this demo showcases less than 0.2% of all tweets sent during the time period. This should give you an idea of how big the Twitter stream is.
European tweets colored by language.
While MapD can scale to much bigger datasets (think many billions of records on a single server) – we wanted to start off with a relatively small public demo and work up. MapD can run on up to 8 Nvidia K80 GPUs per server (which actually comprises 16 GPUs, with 192GB combined VRAM), but the demo here is running on a single GPU. Soon we plan to increase the size of the dataset many-fold and even ingest the Twitter feed in real-time.
The first thing you might notice when looking at the demo is the map covered with millions of multi-colored points. Each point symbolizes a tweet, with the color mapping to the language the tweet was sent in (see the legend at the bottom of the screen). Note that you can click on one or more of the legend items to only see the tweets from those languages. You can also change the coloring scheme to color by the phone OS or the program that the tweet was sent from by clicking on the Language menu item and selecting “Source”. You can have a lot of fun exploring the geographical popularity for each OS/platform. For example, here are maps showing tweets sent from Ios vs. Android, Windows Phone vs Blackberry, and Foursquare vs. Instagram. We invite you to click on any of the images to be taken to the demo with the same filters applied to explore the data for yourself.
Tweets sent from Android (green) vs ios (blue). Note the relative popularity of ios in the US, France, Britain and Japan.
Tweets sent from Blackberry (purple) vs Windows phones (green). Note the continued dominance of Blackberry in places like Central America, Venezuela, Nigeria, South Africa, Saudi Arabia and Indonesia.
Tweets sent from Fourquare (pink) vs Instagram (orange). Notice the worldwide popularity of Instagram and the relative popularity of Foursquare in Eastern Europe, Turkey and Japan.
You can zoom into the map at any point and the rest of the charts will update according to the filter applied by the map's bounding box. For example, if you zoom into Manilla in the Phillipines you will now see the top hashtags for tweets from Manilla, the current total number of tweets around Manilla and the number of tweets per 6 hour period on the time chart. Everything is “cross-filtered”, in that applying a filter to one chart will apply that filter to all other charts. Crossfiltering is a powerful paradigm for data exploration – allowing users to easily drill down on anomalies as well as explore correlations between data attributes.
Tweets colored by OS/Platform from Manilla, Phillipines.
We’ll use this crossfiltering ability to now check out activity around a given hashtag. If we look on the right we’ll see the top hashtags for tweets from Manilla. Let’s click on #popefrancisph. You should see everything update – the map now showing tweets with this hashtag and the timechart showing a spike in tweets with the hashtag around January 15, 2015. If we scroll down the list of tweets on the right or hover over any of the points on the map it is clear that the tweets were sent around Pope Francis’ visitto the Phillipines.
Tweets with hashtag #popefrancisph.
We can also use the tweetmap to search for arbitrary terms. First, zoom to New York City by typing “nyc” in the “Go to Location” box on the left. Then type “parade” in the search box at the top of the map. You’ll then see a spike of tweets around the Macy’s parade held at Thanksgiving every year.
Tweets colored by OS/Platform from NYC.
Tweets mentioning "parade" from NYC.
Looking at individual tweets is quite illuminating but it is also powerful to roll up by various geographical levels. To do this, click on the “choropleth” button to the left of the map. You’ll see a map of the countries of the world colored by the number of tweets from each country, with darker blue signifying a greater number of tweets.
Countries colored by number of tweets - dark blue is more.
If we search for anything the map will change to color by the percentage of tweets from each country mentioning that term. We find that coloring by the percentage of tweets rather than the raw number is much more powerful for spotting geographical trends. For example, if you type in “kfc” you get the following map suggesting that Kentucky Fried Chicken is quite popular in Asia and Southeastern Africa. If you change to coloring by just the number of tweets containing kfc (by clicking on “% Tweets” in the lower left), the result is much less informative since the most populated countries tend to dominate merely due to the sheer number of tweets being sent from those countries.
Percent of tweets containing "KFC" by country.
Number of tweets containing "KFC" by country - note this map is less informative than when we look at percentages.
We can compare KFC to McDonalds, and we get the following map:
Percent of tweets containing "McDonalds" by country. Note the chain's seeming popularity in Eastern Europe.
What’s the spike around Feburary 1st? If we narrow the time filter to focus on the spike, we can see it relates to an ad campaign McDonalds ran around the Superbowl (“#paywithlovin”). Looking at the tweets – people seemed to enjoy the ad!
Spike in tweets containing "McDonalds" around Feb 1st, 2015.
People were lovin McDonald's #paywithlovin campaign.
We can click on any country to zoom in on that country and see the data rolled up at the state/province level. For example, zooming into Turkey, we see the number of tweets by province:
Number of tweets by Turkish province.
Focusing closer at home, it can be interesting to examine regional trends in the US. MapD Tweetmap is great for examining how various slang terms are popular in different areas of the country. For example, let’s look compare the following: yall, hella and wicked.
Percent tweets containing "yall" by state. Very Southern.
Percent tweets containing "hella" by state. Solidly western.
Percent tweets containing "wicked" by state. Seems wicked is both popular in New England as well as parts of the Western US.
Here are a few additional interesting regional trends we found.
Percent tweets containing "church" by state. Note the concentration in the southern US and Utah as well as the weekly spikes on Sundays, as well as the larger spike for Christmas.
Percent tweets containing "Patriots" by state. If we drill in on the spike we see that the top hashtag is #sb49, showing that the associated event is the Patriots winning the Superbowl. We see an obvious concentrations in tweets about the Patriots in New England as well as in Arizona where the game was actually held. The MapD Tweetmap is very good for exploring where the fan bases for various teams are located.
These are just a few of the many interesting phenomena you can uncover using MapD Tweetmap, all made possible by the incredible parallelism and rendering abilities of the GPUs we run on. If you find interesting trends of your own, we invite you to post them below in the comments section. Or learn more about MapD and our GPU database and analytics product at http://www.mapd.com. Happy exploring!