Starting the 2017 Hubway Data Challenge

Time for a New Project 

A couple weeks ago, Hubway, Boston’s bike share service, announced their 2017 Data Challenge. For the challenge, Hubway is providing trip data for previous years, station data, as well access to real time data. Those who enter the challenge will build a wide variety of visualizations and analysis.  I think I might participate, so I downloaded the detailed month-by-month data for 2015 and 2016, as well as the station data and started to experiment.  This post will outline some of my early work before I actually figure out what I will (might) do for the challenge.

For those interested, submissions are due on April 10, 2017.

Making the Data Usable

Hubway provided the data in the only format that matters, csv files.  Since I don’t do much with text files (I am a database person), I wrote a few PostgreSQL scripts to wrangle the data from csv files into PostgreSQL.

The first script to I wrote was a loading script –  Hubway2017_loading.sql. The script is pretty simple. and does the following:

  • Build the tables – the schema is pretty straight forward
  • Load data into staging tables
  • Check for ‘bad values’ in each column – values that don’t meet the data type – they used a ‘\N’ for null. Make sure you check for that.
  • Load data into the final tables – I have each year in seperate tables.
  • Build geometry values for geographic analysis and visualization

The second set of scripts I wrote are analysis scripts.  To start with the data analysis, I wrote three simple analysis scripts:

Feel free to check out my github page for this project and grab whatever code you like.  I anticipate I will be adding more to this project over the next couple weeks.

Starting the Visualization

At the end of the Hubway2017_loading.sql script, I loaded the station data into its own table.  With that data, I created a GeoJSON file of the stations with the reported capacity value with QGIS.  I am using the GeoJSON format for a couple reasons.  It works more seamlessly with CARTO, and it can properly store a date value (something shapefiles don’t do well).

I have uploaded the dataset to my hubway github here.

For anyone who knows Boston/Somerville/Cambridge/Brookline, this pattern of stations will make sense.  The stations with lots of capacity stand out near South Station, MIT, and Mass General.  There are 187 stations in this dataset, however, I need to double check to make sure the stations that appear in the map below were actually in use during 2015/2016, as stations aren’t necessarily permanent.

The next visualization I wanted to make was a time series map displaying the daily starts across the entire system for 2016.  The first step was to build a table with all the relevant data.  For those interested in the script, check out the OriginsByDay_Hubway2016.sql script. Once the script was run and the data created, I built a GeoJSON file in QGIS and uploaded it into CARTO. CARTO is a great online mapping service that is easy to use. If you are looking to make some maps for this challenge and don’t want to spend a lot of time leaning how to map or use mapping specific software, I encourage you to check out CARTO.

The following map steps through each day, visualizing the number of trip starts using CARTO’s Torque feature.  It is fun to watch as the trip starts ebb and flow with the seasons.  One can see stations come in and out of service across the city throughout the year, see major peaks and valleys in usage, and observe the strong relationship in trip starts between downtown Boston and the outlying stations.

Click here for the full size version (that works much better than the version limited by my wordpress CSS).

This simple visualization has given me a number of ideas on what to look into next including:

  • Quantify the relationships between usage and weather
  • The Giver and Taker stations – what is the net usage by station for each day
  • Is that station at MIT really that busy every single day?
  • Relationships between population density and usage
  • Usage in regards to major days in the city, i.e., Marathon Monday, MIT/Harvard/BU move in days, college graduation days, Boston Calling, bad T days (for those who ride the T, you know what I mean).

There are some real patterns in this dataset and it will be fun to look into them and share the results.

Busiest Days in 2016

The last script I put together was to find the busiest days in regards to trip starts.  The busiest day was August 9, 2016, with 6949 starts.  This was a Tuesday, which blows my mind. I am shocked that the busiest day wasn’t a weekend day.

The rest of the busiest days all had over 6k starts and all happened between the end of June and the end of September.  And again, all were on weekdays.  This is really weird to me, as I tend to think that Hubway is used by tourists, and presumably, on weekends (especially downtown). Seeing that the busiest days are weekdays, is actually a real positive for the system, as it can be seen as a viable alternative transit option.

As you can see, there is still a ton to do.  I need to get into this data some more and start to plan the story I want to tell.  Also, I need to do some more QA on the data, so I fully understand what I am dealing with. The biggest part of any data analysis project isn’t the generation of fancy interactive graphics (which no one uses)  or writing ground breaking algorithms; it is the dirty data work.  Without checking and double checking the inputs, the analysis could be wrong, and no one wants that.


Posted in GIS, Mash Ups, Open Source GIS, Spatial Analysis | Comments Off on Starting the 2017 Hubway Data Challenge

Racing Myself – Using Torque in CARTO with runBENrun

As most runners do, I run a lot of the same routes over, and over again. During a run yesterday, I had the idea that I could pull all my runs on my three mile loop and race them against each other using CARTO‘s Torque feature.  It took a little bit of data prep to get my GPS data into a format to “race itself”, but I will spare the technical details for later.

Here are 25 separate runs I ran from 2016 on my Somerville 3 mile loop.  Each point is the lead GPS point of an individual run, with time steps synced, visualized by meters per second speed. To see the full size, click here (it is much better in full size).


Couple points about the race

  • Winner – 10/11/2016 – 3.13 miles, 19:10 time,6:07 pace
  • Loser – 12/12/2016 – 3.13 miles, 24:26, 7:48 pace
  • There are a few deviations on the route, especially at the end.  This is because of number of factors, either because I made a different turn or had to run a little longer to get the required distance due to GPS errors earlier in the run.
  • I am able to race myself because the data I generate with runBENrun project uses elapsed time, so I am able to compare run against run.
  • I used a Nike+ watch, and scraped the data into my own environment using Smashruntapiriik, and my own code.
  • The very last point to leave the map is a run where I didn’t turn off my watch at the end and walked into my house!

Here is How I Created the Race

Warning – Technical Details Ahead! Ahh Yeah!

In 2016, I ran my Somerville loop 25 times.  It’s a pretty flat and fast course, that has a good long straightaway down the bike path, but it does have a couple tight turns and pauses waiting for traffic to cross Broadway.

I run this loop in many different phases of training periods.  Sometimes I try to run fast on this loop, but other times I am using this course for a recovery run. As I was preparing the data I thought my pace and times would be all over the place.

First step was to run a query against all of my 2016 runs to find all three mile runs, that where not classified as interval runs (github here). The script returns any run that rounds to three miles. So I had to do some post processing.

The query returned 42 three mile runs in 2016.  The next step was to pull all of the shapefiles I generated for these datasets a while back (code here!) and check the routes using QGIS. I removed a number of races I ran, and a few three mile runs that weren’t along this route. Once the set was cleaned, I ended up with following 25 runs.

You will notice that the routes don’t all follow the same path.  In fact, I often end at different places on different streets.  This is for a couple reasons: I may have to run a little extra at an end of a run due to pauses in my GPS, or I took a turn a little early toward the end of the run down and I had to make up the distance at the end. Overall, the 25 runs represent a pretty consistent route.

Querying my runBENrun database, I can get my stats for the 25 runs, and checkout how consistent, or inconsistent, I am on this route (github here). The spread of times isn’t too bad, so it should show a decent race.

From here, I wrote a script to create a postgreSQL table with all the relevant runs from the master GPS point table for 2016 (github here).  I made sure to cast the finaltimecounter column as time so that I could use it in CARTO later on.

The output table contains over 29k points, as seen below.  This dataset is what I need to use in CARTO for the animation using Torque.  Using QGIS, I exported the dataset  as a GEOJSON.  Why GEOJSON? Because I had a time field and shapefiles don’t play nice with time data.

I imported the GEOJSON dataset into CARTO and then used the following settings in the Torque Cat wizard.  I found the following settings gave the best view of the “race.” CARTO is super easy to use, and the Torque Cat tool provided a lot of options to make the map look really sharp.

In the end, I got a nice map showing me race myself.  I have a few ideas on how to improve the map and data, but that will be for another time.

Thanks for reading.

Posted in GIS, Mash Ups, Open Source GIS | Comments Off on Racing Myself – Using Torque in CARTO with runBENrun

The Best Way to Map a Run

From my runBENrun project I have generated a lot of data; over 1.2 million data points in 2.75 years.  It is easy enough to write SQL scripts to analyze the data and gain insight into the runs, however, trying to build meaningful maps that help me interpret my runs isn’t as easy.  I have made plenty of maps of my running data over the past year, some good, some bad. In this post, I will explore a few different methods on how to best visualize a single 5k race dataset from my runBENrun project.

The Problem

With most GPS running apps and fitness trackers, you are often generating lots and lots of data.  My old Nike+ watch collects a point every ~0.97 seconds.  That means if you run a six minute mile in a 5k you can log over a 1000 points during the run.  The GPS data collected by my Nike+ watch is great, and I can generate lots of additional derivative attributes from it, but is all that data necessary when trying to spatially understand the ebbs and flows of the run?


I will be using PostgreSQL/PostGIS, QGIS and CARTO in this project. In my maps, I am using Stamen’s Toner Light basemap.

The Data

For this post, I am using a single 5k race I ran in November 2016, in Wakefield, Massachusetts.  The race course loops around Lake Quannapowitt, and is flat and fast with several good long straightaways, and some gentle curves.  I’ve run a couple races on this course, and I recommend it to anyone looking for a good course to try to PR on.  I am also selecting this dataset because the course is a loop, not an out-and-back.  Out-and-back running datasets are a lot harder to visualize since the data often interferes with itself.  I plan on doing a post about visualizing out-and-back runs sometime in the near future.

In case anyone is interested, I have exported a point shapefile and a multiline shapefile of this data, which can be found on my github account.

Before We Starting Mapping…

What’s spatially important to know about this run?  Beyond mile markers, speed is what I am most interested in.  More importantly, how consistent is my speed throughout the run.  I will add mile markers and the Start/Finish to the maps to give some perspective. I will also provide histograms from QGIS of the value and classification breakdowns to help give context to the map.

Let’s Make a Lot of Maps of One Run 

Mapping all 1,117 Points – Let’s start with a simple map. When only visualizing the points I get a map of where I was when I ran. Taking a point about every second, the GPS data isn’t very clear at this scale.
AllPoints Is this a good running map?  No.

Mapping Meters Per Second Bins using Point Data

Points on a map don’t tell us much, especially when the goal here is understand speed throughout the race. The next step in this project is to visualize the range of values in the Meters per Second (MpS) field.  This is a value I calculate in my runBENrun scripts.  The next set of maps will take a couple different approaches to mapping this point data, including visualizing the MpS data by quantile, natural, and user defined breaks.

Quantile Breaks

The first MpS map uses quantile breaks to classify the data.  Since there is a tight distribution of values, quantile breaks will work (there are no major outliers in the dataset). In the following histogram from QGIS we see the distribution of values coded to the five classes.  In all of the maps green equals faster speeds while red values are slower.


The map displays the points classified as such.  What’s important to note from the point based map, is that since there are so many points in such a tight space, that seeing some type of meaningful pattern is tough.  To the naked eye there are many “ups and downs” in the data.  There are clear sections of the race where I am faster than others, but in other parts of the race a “slow” point is adjacent to a “fast” point.  This pattern will show up in the next maps as well.  I am looking into this noise and will hopefully have a post about understanding this type of variation in the GPS data.


Is this a good running map? Not really.  The data is busy; there are too many points to get a real perspective on how consistent the speed was.

Natural Breaks

The next map uses natural breaks classification scheme.  When comparing the histogram using quantile breaks to the natural breaks, one will see that natural breaks algorithm puts fewer values into the lowest (or slowest) bin.

The difference in binning is apparent in the map.  Overall, the reader is given the impression that this is a better run, since there is more non-red colors on the map. Without a MpS legend one wouldn’t know one run was faster than the other. Overall, the general speed patterns are better represented here, as I believe there is a better bins transition between the bins.


Is this a good running map? It’s better.  The natural breaks classification works better than quantile breaks with this dataset, but there is still too much noise in the dataset. That noise won’t be eliminated until the dataset is smoothed.

 Self Defined Classification – Ben Breaks

In this example, I wanted to set my own classification scheme, to create more friendly bins to the “faster” times.  I call this classification scheme the “Confidence Booster.”


One can see that I have larger bins for the faster speeds, and really minimize the red, or slower bins.  The resulting map has a smoother feel, but again, there is too much noise between the MpS values from point to point.


Is this a good running map?  It’s not bad, but as with all the point maps, there is a lot of data to communicate, and at this scale it doesn’t work as well as I would have liked.

Overall, the point data, using every point in the dataset isn’t a good approach for mapping the run.

Mapping Multiline Data

Using my runBENrun scripts, I generated not only point geometries, but also multiline geometries (single line calculated between each sequential point).  At the scale we are viewing these maps, their isn’t much visual difference between the point and line maps, which is understandable.  The multiline datasets are much better utilized when one wants to zoom into a specific area or see the actual details of the route.

I generated the same set of maps using the multiline based data as I did with the points, so I won’t repeat the maps here. However, I will share a map of the multiline data loaded into CARTO, symbolizing the MpS value with the multiline data using a natural breaks classification.

Is this a good running map?  Yes and No. The line data symbolized with natural, quantile, or self-defined breaks works better in an interactive setting where the user can pan and zoom around the dataset. However, the static versions of these maps have the same issues the point data maps do.

Mapping Multiline Data Aggregated to Tenth and Quarter Mile Segments

For this dataset (and almost all running datasets), visualizing every point in the dataset, or every line between every point in the dataset isn’t a good idea.  How about we try a few methods to look at the data differently.  The first approach is to smooth and aggregate the data into quarter mile and tenth of a mile segments.

Using PostGIS, I simply aggregated the geometry based the distance data in the table, and then found the average MpS for that span.  I wrote the output to a table and visualized in QGIS.

Quarter Mile Segments – Quantile Breaks

Since there is less data to visualize, we get a much cleaner, albeit, dumbed down version of the race.  There are clear patterns where I was faster than where I was slower (green=fast, red = slow, relatively speaking).  The consumer of the map isn’t wondering why there was so much variation.  I made this map with both natural breaks and my self-defined breaks, but the quantile classification gave the best view of the race.


Is this a good running map?  Yes, if you just want to know the general trends of how your race went, then this map will let you know that. My second mile, as always, was my worst mile. I traditionally struggle in mile two.

One Tenth Mile Segements – Quantile Breaks

How about comparing different aggregation approaches?  Let’s look at the race broken into tenth of a mile segments using a quantile classification scheme.  In this approach, there is more detail in MpS differences during the race than the quarter mile map. The color for the middle bin does get washed out in the map, so I should probably go back and fix that.


Is this a good running map?  Yes.  The general message – where was I fast and where was I slow – is answered and the data isn’t distracting, like it is in the point maps.  A way to improve this visualization would be to add the actual breaks between tenth mile segments, and maybe a table with the time splits.

Using Standard Deviation and Average Bins

The last set of maps will visualize the race using some basic statistical measures – standard deviation and average.

Standard Deviation

The distribution of values are fairly compact.  The resulting maps using the standard deviation bins reflect that.

With the point dataset, MpS values classified using standard deviation, you actually get a pretty decent looking map.  Since there are so few very fast or very slow MpS values, you don’t get many points in those bins extreme bins. This means that the color ranges fall more in the middle of the range. This map won’t tell you have fast or slow you were really going, but it gives you an idea of how well your run was relative to the rest of the race.  For what I plan to do in a race, I would hope to see a majority of values in the +1 or -1 standard deviation bins.  This would mean that I was pretty consistent in my MpS.  Ideally, I would also see values in the higher plus standard deviation bins towards the end of the race, as I really try to pick up the pace.


Is this a good running map? If you know what you are looking at, then this map can tell you a lot about your run.  However, if you aren’t familiar with what a standard deviation is, or how it is mapped, then this might not be a good approach.

Average Values

The last map for this post is simply mapping those points that are above, at, or below the average MpS for the race.  In this race, my average MpS was 4.52 (For reference, Mo Farah won the 2016 Olympic 5k in 13:03, or 6.39 MpS!).  I created three classes – green – points with an above average MpS, yellow – points that were average, and red – points that had a below average MpS. The view of the run isn’t that bad with this approach.  The user get a fairly clear indication of relative speed during the race, without all the noise from previous attempts to classify the data.  Using the average value here though only works because the range of values is fairly tight.  If there was a wider swing in values, this approach might not work.


Is this a good running map? Yeah, it’s not that bad. The colors are a little harsh. In this case it works, but depending on the range of values, mapping compared to the average may not work. Another test would be to compare values against the median.

What map was the best approach?

In the end, what map was the best approach to visualizing the data from the race with the goal of best understanding my MpS?  I had two maps that I think met the requirements:

  • Quarter Mile Segement Quantile Breaks – smooth transitions between classes, easy to view, and informed readers of the general race speed trends
  • Standard Deviation – good approach if you know what a standard deviation is, and if your data is compact (don’t have huge swings in value).  This approach gives the reader a clear indication of how they were doing relative to the rest of the run, without worrying about the individual MpS values.

There is value in all the maps, and with a little work, they could be improved as well. However, these two maps were my picks.

What’s Next?

I actually made another 10 or so maps when working on this blog, including maps using proportional symbols, incorporating more data smoothing, and some ideas about flow maps.  The next steps will include exploring those visualization methods with the goal of getting them into the blog.

Have any other suggestions? Send me a note on twitter @GISDoctor!

Posted in GIS, Open Source GIS, Spatial Analysis | Tagged , , , | Comments Off on The Best Way to Map a Run