Tag: data science

Oct 28 2014

Exploring the ONS

This post was first published at ScraperWiki.

The Office for National Statistics (ONS) is the United Kingdom statistical body charged by the government with the task of collecting and publishing  statistics related to the economy, population and society of England and Wales at national, regional and local levels. The data is typically published in the form of Excel spreadsheets.

The ONS is working on opening up their data, and making it more accessible to users. We’ve been doing a bit of work to help with that. This is typical of a number of jobs we have done. A customer has a website containing content which they want to move/process/republish elsewhere. The current website might have been built by aggregation over a number of years, and the underlying structure of the Content Management System may not be available to them. In these circumstances making a survey of the pre-existing content is an obvious first step.

The index for the ONS reference tables and datasets can be found here. Each dataset has a title, a release date, and the type of dataset. There is also a URL to the dataset inside the title field there is an indication of the size of the file. We wrote a simple scraper to collect these pieces of information.

First up, we’ll looking at the topics of the data released. There are a couple of routes into discovering these, one is to read the titles, this is OK as an approach but the titles are quite wordy and sometimes it isn’t clear what they refer to. An alternative, in this case, is to look at the URLs to the documents.They look something like this:

http://www.ons.gov.uk/ons/rel/lms/labour-market-statistics/november-2013/table-unem03.xls

This can be quite revealing since even if the website is not explicit about its structure the URLs can reveal the structure the builder used. We process the URLs by splitting them at the backslashes. The first part http://www.ons.gov.uk/ons/rel is common to all the URLs. Subsequent parts we can use to define a hierarchy. In this case we will focus on the fourth part of the hierarchy – “labour-market-statistics” in this instance, this gives us a human readable description of a topic. There are approximately 400 topics as defined by this metric as opposed to 90 or so defined by the third level of the hierarchy. Using the fourth level of the hierarchy key areas of the website by numbers of documents are:

  • Labour market statistics
  • National population projections
  • Family spending
  • Subnational labour market statistics
  • Census
  • Annual survey of hours and earnings.

We can visualise this as a treemap, here I am simply showing the top 20 areas by number of documents:

ONS-treemap-(simple)-v2

These 20 topics cover approximately the two thirds of the total number of documents.

We can identify file types using the file extension in the URL, this approach needs to be used a little cautiously since sometimes the extension doesn’t match the file type. Most of the files are Excel spreadsheets although there are a few CSV and zip files, the zip files containing Excel spreadsheets. CSV appears to have been used for some of the older datasets. Most of the files are pretty small, less than 290kb but there are a few up to much larger sizes.

Finally we can look at the release dates for the datasets. There are datasets from as far back as 1988, in fact the data set released in 1988 actually refers to data from 1984. There are some data released regularly from about 2001 but from 2011 a wider range of data has been released on a regular basis. We can see the monthly pattern of data releases in this timeline for 2014 which is restricted to the top 20 topics identified above:

Timeline-(detail)-v2

This shows the big releases of labour market statistics, both national and regional, on the third Wednesday of each month. Other monthly releases include retail sales and producer prices data. And every week provisional figures on the registration of deaths in England and Wales are reported.

You can explore these data yourself using the Tableau workbook here.

The actual content of these spreadsheets is another story.

This survey approach to a website is handy for a range of applications, and the techniques used are quite general. We’ve used similar approaches to understand government and newspaper websites.

Oct 20 2014

Book review: Graph Theory and Complex Networks by Maarten van Steen

graph_theory

This review was first published at ScraperWiki.

My last read, on the Gephi graph visualisation package, was a little disappointing but gave me an enthusiasm for Graph Theory. So I picked up one of the books that it recommended: Graph Theory and Complex Networks: An Introduction by Maarten van Steen to learn more. In this context a graph is a collection of vertices connected by edges, the edges may be directed or undirected. The road network is an example of a graph; the junctions between roads are vertices, the edges are roads and a one way street is a directed edge – two-way streets are undirected.

Why study graph theory?

Graph theory underpins a bunch of things like route finding, timetabling, map colouring, communications routing, sol-gel transitions, ecologies, parsing mathematical expressions and so forth. It’s been a staple of Computer Science undergraduate courses for a while, and more recently there’s been something of a resurgence in the field with systems on the web provided huge quantities of graph-shaped data both in terms of the underlying hardware networks and the activities of people – the social networks.

Sometimes the links between graph theory and an application are not so obvious. For example, project planning can be understood in terms of graph theory. A task can depend on another task – the tasks being two vertices in a graph. The edge between such vertices is directed, from one to the other, indicating dependency. To give a trivial example: you need a chicken to lay an egg. As a whole a graph of tasks cannot contain loops (or cycles) since this would imply that a task depended on a task that could only be completed after it, itself had been completed. To return to my example: if you need an egg in order to get a chicken to lay an egg then you’re in trouble! Generally, networks of tasks should be directed acyclic graphs (or DAG) i.e. they should not contain cycles.

The book’s target audience is 1st or 2nd year undergraduates with moderate background in mathematics, it was developed for Computer Science undergraduates. The style is quite mathematical but fairly friendly. The author’s intention is to introduce the undergraduate to mathematical formalism. I found this useful, since mathematical symbols are difficult to search for and shorthands such  as operator overloading even more so. This said, it is still an undergraduate text rather than a popular accounts don’t expect an easy read or pretty pictures, or even pretty illustrations.

The book divides into three chunks. The first provides the basic language for describing graphs, both words and equations. The second part covers theorems arising from some of the basic definitions, including the ideas of “walks” – traversals of a graph which take in all vertices and “tours” which take in all edges. This includes long standing problems such as the Dijkstra’s algorithm for route finding, and the travelling salesman problem. Also included in this section are “trees” – networks with no cycles – where is a cycle is a closed walk which visits vertices just once.

The third section covers the analysis of graphs. This starts with metrics for measuring graphs such as vertex degree distributions, distance statistics and clustering measures. I found this section rather brief, and poorly illustrated. However, it is followed by an introduction to various classes of complex networks including the original random graphs(connect), small-world and scale-free networks. What is stuck me about complex graphs is that they are each complex in their own way. Random, small-world and scale-free networks are all methods for constructing a network in order to try to represent a known real world situation. Small-world networks arise from one of Stanley Milgram’s experiments: sending post across the US via social networks. The key feature is that there are clusters of people who know each other but these clusters are linked by the odd “longer range” contact.

The book finishes with some real world examples relating to the world wide web, peer-to-peer sharing algorithms and social networks. What struck me in social networks is that the vertices (people!) you identify as important can depend quite sensitively on the metric you use to measure importance.

I picked up Graph Theory after I’d been working with Gephi, wanting to learn more about the things that Gephi will measure for me. It serves that purpose pretty well. In addition I have a better feel for situations where the answer is “graph theory”. Furthermore, Gephi has a bunch of network generators to create random, small-world and scale-free networks so that you can try out what you’ve learned.

Sep 22 2014

Book review: Network Graph Analysis and visualization with Gephi by Ken Cherven

network_gephi

This review was first published at ScraperWiki.

I generally follow the rule that if I haven’t got anything nice to say about something then I shouldn’t say anything at all. Network Graph Analysis and visualization with Gephi by Ken Cherven challenges this principle.

Gephi is a system for producing network visualisations, as such it doesn’t have a great many competitors. Fans of Unix will have used Graphviz for this purpose in the past but Gephi offers greater flexibility in a more user-friendly package. Graph theory and network analysis have been growing in importance over the past few years in part because of developments in the analysis of various complex systems using network science. As a physical scientist I’ve been aware of this trend, and it clearly also holds in the social sciences. Furthermore there is much increased availability of network information from social media such as Twitter and Facebook.

I’ve used Gephi a few times in the past, and to be honest there has been an air of desperate button clicking to my activities. That’s to say I felt Gephi could provide the desired output but I could only achieve it by accident. I have an old-fashioned enthusiasm for books even for learning about modern technology. Hence Network Graph Analysis and visualization with Gephi – the only book I could find with Gephi in the title. There is substantial online material to support Gephi but I hoped that this book would give me a better insight into how Gephi worked and some wider understand of graph theory and network analysis.

On the positive side I now have a good understanding of the superficial side of the interface, a feel for how a more expert user thinks about Gephi, and some tricks to try.

I discovered from Network Graph Analysis that the “Overview” view in Gephi is what you might call “Draft”, a place to prepare visualisations which allows detailed interaction. And the “Preview” view is what you might call “Production”, a place where you make a final, beautiful version of your visualisations.

The workflow for Gephi is to import data and then build a visualisation using one of a wide range of layout algorithms. For example, force-based layouts assume varying forces between nodes for which an arrangement of nodes can be calculated by carrying out a pseudo-physical simulations. These algorithms can take a while to converge, and may get trapped in local minima. The effect of these layout algorithms is to reveal some features of the network. For example, the force layouts can reveal clusters of nodes which might also be discovered by a more conventional statistical clustering algorithm. The concentric layout allows a clearer visualisation of hierarchy in a network.

It’s clear that the plugin ecosystem is important to the more experienced user of Gephi. Plugins provide layout algorithms, data helpers, new import and export functionality, analysis and so forth. You can explore them in the Gephi marketplace.

Cherven recommends a fairly small, apparently well-chosen set of references to online resources and books. The Visual Complexity website looks fabulous. You can read the author’s complete, pre-publication draft of Networks, Crowds and Markets: Reasoning about a highly connected world by David Easley and Jon Kleinberg here. It looks good but it’s nearly 800 pages! I’ve opted for the rather shorter Graph Theory and Complex Networks: An Introduction by Maarten van Steen.

On the less positive side, this is an exceedingly short book. I read it in a couple of 40 minute train journeys. It’s padded with detailed descriptions of how to install Gephi and plugins, including lots of screenshots. The coverage is superficial, so whilst features may be introduced the explanation often tails off into “…and you can explore this feature on your own”.

Network Graph Analysis is disappointing, it does bring a little enlightenment to a new user of Gephi but not very much. A better book would have provided an introduction to network and graph analysis with Gephi the tool to provide practical experience and examples, in the manner that Data Mining does for weka and Natural Language Processing with Python does for the nltk library.

This book may be suitable for someone who is thinking about using Gephi and isn’t very confident about getting started. The best alternative that I’ve found is the online material on GitHub (here).

Sep 18 2014

Of Matlab and Python

I’ve been a scientist and data analyst for nearly 25 years. Originally as an academic physicist, then as a research scientist in a large fast moving consumer goods company and now at a small technology company in Liverpool. In common to many scientists of my age I came to programming in the early eighties when a whole variety of home computers briefly flourished. My first formal training in programming was FORTRAN after which I have made my own way.

I came to Matlab in the late nineties, frustrated by the complexities of producing a smooth workflow with FORTRAN involving interaction, analysis and graphical output.

Matlab is widely used in academic circles and a number of industries because it provides a great deal of analytical power in a user-friendly environment. Its notation for handling matrix (array) calculations is slick. Its functionality is extended by a range of toolboxes, and there is a community of scientists sharing new functionality. It shares this feature set with systems such as IDL and PV-WAVE.

However, there are a number of issues with Matlab:

  • as a programming language it has the air of new things being botched onto a creaking frame. Support for unit testing is an afterthought, there is some integration of source control into the Matlab environment but it is with Source Safe. It doesn’t support namespaces. It doesn’t support common data structures such as dictionaries, lists and sets.
  • The toolbox ecosystem is heavily focused on scientific applications, generally in the physical sciences. So there is no support for natural language processing, for example, or building a web application based on the powerful analysis you can do elsewhere in the ecosystem;
  • the licensing is a nightmare. Once you’ve got core Matlab additional toolboxes containing really useful functionality (statistics, database connections, a “compiler”) are all at an additional cost. You can investigate pricing here. In my experience you often find yourself needing a toolbox for just a couple of functions. For an academic things are a bit rosier, universities get lower price licenses and the process by which this is achieved is opaque to end-users. As an industrial user, involved in the licensing process, it is as bad as line management and sticking needles in your eyes in the “not much fun thing to do” stakes;
  • running Matlab with network licenses means that your code may stop running part way through because you’ve made a call to a function to which you can’t currently get the license. It is difficult to describe the level of frustration and rage this brings. Now of course one answer is to buy individual licenses for all, or at least a significant surplus of network licenses. But tell that to the budget holder particularly when you wanted to run the analysis today. The alternative is to find one of the license holders of the required toolbox and discover if they are actually using it or whether they’ve gone off for a three hour meeting leaving Matlab open;
  • deployment to users who do not have Matlab is painful. They need to download a more than 500MB runtime, of exactly the right version and the likelihood is they will be installing it just for your code;

I started programming in Python at much the same time as I started on Matlab. At the time I scarcely used it for analysis but even then when I wanted to parse the HTML table of contents for Physical Review E, Python was the obvious choice. I have written scrapers in Matlab but it involved interfering with the Java underpinnings of the language.

Python has matured since my early use. It now has a really great system of libraries which can be installed pretty much trivially, they extend far beyond those offered by Matlab. And in my view they are of very good quality. Innovation like IPython notebooks take the Matlab interactive style of analysis and extend it to be natively web-based. If you want a great example of this, take a look at the examples provided by Matthew Russell for his book, Mining the Social Web.

Python is a modern language undergoing slow, considered improvement. That’s to say it doesn’t carry a legacy stretching back decades and changes are small, and directed towards providing a more consistent language. Its used by many software developers who provide a source of help, support and an impetus for an decent infrastructure.

Ubuntu users will find Python pre-installed. For Windows users, such as myself, there are a number of distributions which bundle up a whole bunch of libraries useful for scientists and sometimes an IDE. I like python(x,y). New libraries can generally be installed almost trivially using the pip package management system. I actually use Python in Ubuntu and Windows almost equally often. There are a small number of libraries which are a bit more tricky to install in Windows – experienced users turn to Christoph Gohlke’s fantastic collection of precompiled binaries.

In summary, Matlab brought much to data analysis for scientists but its time is past. An analysis environment built around Python brings wider functionality, a better coding infrastructure and freedom from licensing hell.

Sep 10 2014

Inordinately fond of beetles… reloaded!

sciencemuseum_logo

This post was first published at ScraperWiki.

Some time ago, in the era before I joined ScraperWiki I had a play with the Science Museums object catalogue. You can see my previous blog post here. It was at a time when I was relatively inexperienced with the Python programming language and had no access to Tableau, the visualisation software. It’s a piece of work I like to talk about when meeting customers since it’s interesting and I don’t need to worry about commercial confidentiality.

The title comes from a quote by J.B.S. Haldane, who was asked what his studies in biology had told him about the Creator. His response was that, if He existed then he was “inordinately fond of beetles”.

The Science Museum catalogue comprises three CSV files containing information on objects, media and events. I’m going to focus on the object catalogue since it’s the biggest one by a large margin – 255,000 objects in a 137MB file. Each object has an ID number which often encodes the year in which the object was added to the collection; a title, some description, it often has an “item name” which is a description of the type of object, there is sometimes information on the date made, the maker, measurements and whether it represents part or all of an object. Finally, the objects are labelled according to which collection they come from and which broad group in that collection, the catalogue contains objects from the Science Museum, Nation Railway Museum and National Media Museum collections.

The problem with most of these fields is that they don’t appear to come from a controlled vocabulary.

Dusting off my 3 year old code I was pleased to discover that the SQL I had written to upload the CSV files into a database worked almost first time, bar a little character encoding. The Python code I’d used to clean the data, do some geocoding, analysis and visualisation was not in such a happy state. Or rather, having looked at it I was not in such a happy state. I appeared to have paid no attention to PEP-8, the Python style guide, no source control, no testing and I was clearly confused as to how to save a dictionary (I pickled it).

In the first iteration I eyeballed the data as a table and identified a whole bunch of stuff I thought I needed to tidy up. This time around I loaded everything into Tableau and visualised everything I could – typically as bar charts. This revealed that my previous clean up efforts were probably not necessary since the things I was tidying impacted a relatively small number of items. I needed to repeat the geocoding I had done. I used geocoding to clean up the place of manufacture field, which was encoded inconsistently. Using the Google API via a Python library I could normalise the place names and get their locations as latitude – longitude pairs to plot on a map. I also made sure I had a link back to the original place name description.

The first time around I was excited to discover the Many Eyes implementation of bubble charts, this time I now realise bubble charts are not so useful. As you can see below in these charts showing the number of items in each subgroup. In a sorted bar chart it is very obvious which subgroup is most common and the relative sizes of the subgroup. I’ve coloured the bars by the major collection to which they belong. Red is the Science Museum, Green is the National Rail Museum and Orange is the National Media Museum.

image

Less discerning members of ScraperWiki still liked the bubble charts.

image

We can see what’s in all these collections from the item name field. This is where we discover that the Science Museum is inordinately fond of bottles. The most common items in the collection are posters, mainly from the National Rail Museum but after that there are bottles, specimen bottles, specimen jars, shops rounds (also bottles), bottle, drug jars, and albarellos (also bottles). This is no doubt because bottles are typically made of durable materials like glass and ceramics, and they have been ubiquitous in many milieu, and they may contain many and various interesting things.

image

Finally I plotted the place made for objects in the collection, this works by grouping objects by location and then finding latitude and longitude for those group location. I then plot a disk sized by the number of items originating at that location. I filtered out items whose place made was simply “England” or “London” since these made enormous blobs that dominated the map.

 

image

 

You can see a live version of these visualisation, and more on Tableau Public.

It’s an interesting pattern that my first action on uploading any data like this to Tableau is to do bar chart frequency plots for each column in the data, this could probably be automated.

In summary, the Science Museum is full of bottles and posters, Tableau wins for initial visualisations of a large and complex dataset.