Finding (Nearly) Duplicate Items in a Data Column

[WARNING – THIS IS A *BAD ADVICE* POST – it describes a trick that sort of works, but the example is contrived and has a better solution – text facet and then cluster on facet (h/t to @mhawksey’s Mining and OpenRefine(ing) JISCMail: A look at OER-DISCUSS [Listserv] for making me squirm so much through that oversight I felt the need to post this warning…]

Suppose you have a dataset containing a list of Twitter updates, and you are looking for tweets that are retweets or modified retweets of the same original tweet. The OpenRefine Duplicate custom facet will identify different row items in that column that are exact duplicates of each other, but what about when they just don’t quite match: for example, an original tweet and it’s appearance in an RT (where the retweet string contains RT and the name of the original sender), or an MT, where the tweet may have been shortened, or an RT of an RT, or an RT with an additional hashtag. Here’s one strategy for finding similar-ish tweets, such as popular retweets, in the data set using custom text facets in OpenRefine.

The Ngram GREL function generates a list of word ngrams of a specified length from a string. If you imagine a sliding window N words long, the first ngram will be the first N words in the string, the second ngram the second word to the second+N’th word, and so on:

If we can reasonably expect word sequences of length N to appear in out “duplicate-ish” strings, we can generate a facet on ngrams of that length.

It may also be worth experimenting with combining the ngram function with the fingerprint GREL function. The fingerprint function identifies unique words in a string, reduces them to lower case, and then sorts them in alphabetical order:

If we generate the fingerprint of a string, and then run the ngram function, we generate ngrams around the alphabetically ordered fingerprint terms:

For a sizeable dataset, and/or long strings, it’s likely that we’ll get a lot of ngram facet terms:

We could list them all by setting an appropriate choice count limit, or we can limit the facet items to be displayed by displaying the choice counts, setting the range slider to show those facet values that appear in a large number of columns for example, and then ordering the facet items by count:

Even if tweets aren’t identical, if they contain common ngrams we can pull them out.

Note that we might then order the tweets as displayed in the table using time/date order (a note on string to time format conversions can be found in this postFacets in OpenRefine):

Or alternatively, we might choose to view them using a time facet:

Note that when you set a range in the time facet, you can then click on it and slide it as a range, essentially providing a sliding time window control for viewing records that appear over a given time range/duration.

Grabbing Twitter Search Results into Google Refine And Exporting Conversations into Gephi

How can we get a quick snapshot of who’s talking to whom on Twitter in the context of a particular hashtag?

Here’s a quick recipe that shows how…

First we need to grab some search data. The Twitter API documentation provides us with some clues about how to construct a web address/URL that will grab results back from a particular search on Twitter in a machine readable way (that is, as data):

  • is the base URL, and the format we require is json, which gives us
  • the query we want is presented using the q= parameter:
  • if we want multiple search terms (for example, library skills), they need encoding in a particular way. The easiest was is just to construct your URL, enter it into the location/URL bar of your browser and hit enter, or use a service such as this string encoder. The browser should encode the URL for you. (If the only punctuation in your search phrase are spaces, you can encode them yourself: just change each space to %20, to give something like library%20skills. If you want to encode the # in a hashtag, use %23
  • We want to get back as many results as are allowed at any one time (which happens to be 100), so set rpp=100, that is:
  • results are paged (in the sense of different pages of Google search results, for example), which means we can ask for the first 100 results, the second 100 results and so on as far back as the most recent 1500 tweets (page 15 for rpp=100, or page 30 if we were using rpp=50 (since 15*100 = 30*50 = 1500):

Clicking on Next provides us with a dialogue that will allow us to load the data from the URLs into Google Refine:

Clicking “Configure Parsing Options” loads the data and provides us with a preview of it:

If you inspect the data that is returned, you should see it has a repeating pattern. Hovering over the various elements allows you to identify what repeating part of the result we want to import. For example, we could just import each tweet:

Or we could import all the data fields – let’s grab them all:

If you click the highlighted text, or click “Update Preview View”, you can get a preview of how the data will appear. To return to the selection view, click “Pick Record Nodes”:

“Create Project” actually generates the project and pulls all the data in… The column names are a little messy, but we can tidy those:

Look for the from_user and to_user columns and rename them source and target respectively… (hovering over a column name pops up tooltip that shows the full column name):

For the example I’m going to describe, we don’t actually need to rename the columns, but it’s handy to know how to do it;-)

We can now filter out all the rows with a “null” value in the target column. It seems a bit fiddly at first, but you soon get used to the procedure… Select the text facet to pop up a window that show the unique elements in the target column and how often they occur. Sort the list by count, and click on the “null” element – it should be highlighted and its setting should appear as “exclude”. The column will now be showing elements in the column that have the null value:

Click on the “Invert” option and the column will now filter out all the “null” elements and only show the elements that have a non-null value – that is, tweets that have a “to_user” value (which is to say, those tweets were sent to a particular user). Here’s what we get:

Let’s now export the source and target data so we can get it into Gephi:

Deselect all the columns, and then select source and target columns; also deselect the ‘output column headers’ – we don’t need headers where this file is going…

Export the custom layout as CSV data:

We can now import this data into another application – Gephi. Gephi is a cross platform package for visualising networks. In the simplest case, it can import two column data files where each row represents two things that are connected to each other. In our case, we have connections between “source” and “target” Twitter names – that is, connections that show when one Twitter user in our search sample has sent a message to another.

Launch Gephi and from the file menu, open the file you exported from Google Refine:

We’ve now got our data into Gephi, where we can start to visualise it…

…but that is a post for another day… (or if you’re impatient, you can find some examples of how to drive Gephi here).

Merging Data Sets Based on Partially Matched Data Elements

A tweet from @coneee yesterday about merging two datasets using columns of data that don’t quite match got me wondering about a possible R recipe for handling partial matching. The data in question related to country names in a datafile that needed fusing with country names in a listing of ISO country codes,although the recipe I’m describing here is intended to be a general purpose partial matcher. (See the comments for alternative methods for the special case of reconciling country names with country codes.)

The original data set had the form:

9,New Zealand,86
12,S. Korea,83.8
15,United Kingdom,82
16,Antigua & Barbuda,82

I’m not sure what country code listing was being used, but it probably looked something like this list of ISO Country Codes by Country:


It’s easy enough to reduce all the country names to lowercase characters so that we can try to match them exactly without worrying about any differences in capitalisation, but how do we match country names that don’t match exactly – ANTIGUA AND BARBUDA and Antigua & Barbuda, for example, or Central African Rep. and Central African Republic?

One trick is to use one of the well known partial string matching algorithms, such as the Levenshtein Distance. Here’s a recipe I hacked together that first tries to find an exact match on country names by attempting to merge the two country lists directly, and then tries to partially match any remaining unmatched names in the original list. A signature function is constructed to help out the partial matching attempt that reduces all words in the country name to lower case, sorts them alphabetically, and then concatenates them with no adjoining spaces.

(The signature idea was cribbed from the fingerprint that is available in Google Refine and that I employed in Merging Datasets with Common Columns in Google Refine.)

[@downes commented that the code wasn’t really written with clarity in mind – so I’ve added some comments…]

#Load in the data from the URLs:
PercentageUsingTheNet=read.csv(url(', encoding='MACROMAN'))

##Here's where the algorithm starts...
##I'm going to generate a signature from country names to reduce some of the minor differences between strings
##In this case:
### convert all characters to lower case (tolower())
### split the string into a vector (unlist()) of separate words (strsplit())
### sort the words alphabetically (sort())
### and then concatenate them with no spaces (paste(y,collapse='')).
##So for example, United Kingdom would become kingdomunited
##To extend this function, we might also remove stopwords such as 'the' and 'of', for example (not shown).
  sig=paste(sort(unlist(strsplit(tolower(x)," "))),collapse='')

#The partialMatch function takes two wordlists as vectors (x,y) and an optional distance threshold (levDist)
#The aim is to find words in the second list (y) that match or partially match words in the first (x)
  #Create a data framecontainind the signature for each word
  xx=data.frame(sig=sapply(x, signature),row.names=NULL)
  yy=data.frame(sig=sapply(y, signature),row.names=NULL)
  #Add the original words to the data frame too...
  #We only want words that have a signature...

  #The first matching pass - are there any rows in the two lists that have exactly the same signature?
  matched=subset(xy,subset=(!( & !(
  #?I think matched=xy[ complete.cases(raw.x,raw.y) ] might also work here?
  #Label the items with identical signatures as being 'Duplicate' matches

  #Grab the rows from the first list that were unmatched - that is, no matching item from the second list appears
  #We've grabbed the signature and original raw text from the first list that haven't been matched up yet
  #Name the columns so we know what's what

  #This is the partial matching magic - agrep finds items in the second list that are within a 
  ## certain Levenshtein distance of items in the first list.
  ##Note that I'm finding the distance between signatures.
  todo$partials= as.character(sapply(todo$sig, agrep, yy$sig,max.distance = levDist,value=T))

  #Bring the original text into the partial match list based on the sig key.

  #Find the items that were actually partially matched, and pull out the columns relating to signatures and raw text
  partial.matched=subset(todo,subset=(!( & !(,select=c("sig","raw.x","raw.y"))
  #Label these rows as partial match items
  #Add the set of partially matched items to the set of duplicate matched items
  #Find the rows that still haven't been matched

  #If there are any unmatched rows, add them to the list of matched rows, but labelled as such
  if (nrow(un.matched)>0){

  #Grab the columns of raw text from x and y from the matched list, along with how they were matched/are unmatched
  #Ideally, the length of this should be the same as the length of valid rows in the original first list (x)


#A rogue character in @coneee's data file borked things for me, so I originally needed to do a character code conversion first
#Loading the CSV with the encoding attribute set (h/t Kent Johnson) seemed to work properly though...

#Call the partial match function using two vectors
#The aim is to find items in the second vector that partially match items in the first
#The function returns the first list annotated with partial match items from the second and a match type

As ever, this code was arrived at by tinkering and searching on Stack Overflow (using search terms along the lines of “partial match R” and “R levenshtein”). If you can improve on it, please do so and paste a link to the improved code, or a code fragment itself, in the comments below:-)

UPDATE: via the comments, the following suggestion that I don’t have time to check right now…
#Bring the original text into the partial match list based on the sig key.

#Label these rows as partial match items
+if (nrow(partial.matched) > 0) partial.matched$pass=”Partial”


When we run the script and look at the contents of the matches dataframe, this is an example of what we get:

This data frame can then be merged with the originally loaded data to give us the required country code annotations:

#Merge the original data set with the ISO country code country name keys
#Merge in the ISO country codes

Something like this, for example:

Unfortunately, not every country in the original data set is matched:

In particular, here are the unmatched items and what they presumably should have been matched with:

Guinea Bissau - GUINEA-BISSAU
Congo (Kinshasa) - CONGO
Slovak Republic - SLOVAKIA

We could try to increase the Levenshtein Distance within which a partial match is suggested, but then we run the risk of introducing false positives.

Removing stopwords as part of the signature function may help, for example in the case of St. Vincent & Grenadines. (Also, I’m not sure what went wrong with Guinea Bissau?) We could presumably also make something of the abbreviations (S., Dem., Rep., Eq., P. D. R.), using these as NGrams or stubs in a regular expression? So for example, convert Lao P. D. R. to a regular expression invoking searches for a three word phrase in which consecutive words start with P, D, R. (Of course, this would require matching across the full country name rather than the signature.)

The Google Refine documentation page Clustering In Depth (Methods and theory behind the clustering functionality in Google Refine) also identifies several other fuzzy/partial matching algorithms that could be tried here. In which case, I guess I need to rewrite the above function so that it can accommodate different matching algorithms? IF you manage to implement any of these other fuzzy matching approaches, please post a link in the comments.

PS If you use any other approaches for partial matching of strings, please feel free to share how you do it in the comments below…:-) For example:
theBioBucket: Merging Dataframes by Partly Matching String. This post also introduces adist(), an R function that returns the Levenshtein distance between two strings.

Data Shaping in Google Refine – Generating New Rows from Multiple Values in a Single Column

One of the things I’ve kept stumbling over in Google Refine is how to use it to reshape a data set, so I had a little play last week and worked out a couple of new (to me) recipes.

The first relates to reshaping data by creating new rows based on columns. For example, suppose we have a data set that has rows relating to Olympics events, and columns relating to Medals, with cell entries detailing the country that won each medal type:

However, suppose that you need to get the data into a different shape – maybe one line per country with an additional column specifying the medal type. Something like this, for example:

How can we generate that sort of view from the original data set? Here’s one way, that works when the columns you want to split into row values are contiguous (that is, next to each other). From the first column in the set of columns you want to be transformed, select Transpose > Transpose cells across columns into rows:

We now set the original selected column headers to be the cell value within a new column – MedalType – and the original cell values the value within a Country column:

(Note that we could also just transform the data into a single column. For example, suppose we had columns relating to courses currently taken by a particular student (Course 1, Course 2, Course 3), with a course code as cell value and one, two or three columns populated per student. If we wanted one row per student per course, we could just map the three columns onto a single column – CourseCode – and assign multiple rows to each student, then filtering out rows with a blank value in the CourseCOde column as required.)

Ticking the Fill down in other columns checkbox ensures that the appropriate Sport and Event values are copied in to the newly created rows:

Having worked out how to do that oft-required bit of data reshaping, I thought I could probably have another go at something that has been troubling me for ages – how to generate multiple rows from a single row where one of the columns contains JSON data (maybe pulled from a web service/API) that contains multiple items. This is a “mate in three” sort of problem, so here’s how I started to try to work it back. Given that I now know how to map columns onto rows, can I work out how to map different results in a JSON response onto different columns?

For example, here’s a result from the Facebook API for a search on a particular OU course code and the word open in a Facebook group name:

{“data”:[{“version”:1,”name”:”U101 (Open University) start date February 2012″,”id”:”325165900838311″},{“version”:1,”name”:”Open university, u101- design thinking, October 2011″,”id”:”250227311674865″},{“version”:1,”name”:”Feb 2011 Starters U101 Design Thinking – Open University”,”id”:”121552081246861″},{“version”:1,”name”:”Open University – U101 Design Thinking, Feburary 2011″,”id”:”167769429928476″}],”paging”:{“next”:…etc…}}

It returns a couple of results in the data element, in particular group name and group ID. Here’s one way I found of creating one row per group… Start off by creating a new column based on the JSON data column that parses the results in the data element into a list:

We can then iterate over the list items in this new column using the forEach grel command. The join command then joins the elements within each list item, specifically the group ID and name values in each result:


You’ll notice that for multiple results, this produces a list of joined items, which we can also join together by extending the GREL expression:


We now have a column that contains ‘||’ and ‘::’ separated items – :: separates individual group results from each other, || separates the id and name for each particular group.

Given we know how to create rows from multiple columns, we could try to split this column into separate columns using Edit column > Split into separate columns. This would create one column per result, which we could then transform into rows, as we did above. Whilst I don’t recommend this route in this particular case, here’s how we could go about doing it…

A far better approach is to use the Edit cells > split multi-valued cells option to automatically create new rows based on splitting the elements in a single column:

Note, however that this creates blanks in the other columns, so we need to Edit cells > Fill down to fill in the blanks in any other columns we want to refer to. After doing that, we end up with something like this:

We could now split the groupPairs column using the || separator to create two columns – Group ID and group name – giving us one row per group, and separate columns identifying the course, group name and group ID.

If the above route seems a little complicated, fear not…Once you apply it, it starts to make sense!

Social Interest Positioning – Visualising Facebook Friends’ Likes With Data Grabbed Using Google Refine

What do my Facebook friends have in common in terms of the things they have Liked, or in terms of their music or movie preferences? (And does this say anything about me?!) Here’s a recipe for visualising that data…

After discovering via Martin Hawksey that the recent (December, 2011) 2.5 release of Google Refine allows you to import JSON and XML feeds to bootstrap a new project, I wondered whether it would be able to pull in data from the Facebook API if I was logged in to Facebook (Google Refine does run in the browser after all…)

Looking through the Facebook API documentation whilst logged in to Facebook, it’s easy enough to find exemplar links to things like your friends list ( or the list of likes someone has made (; replacing me with the Facebook ID of one of your friends should pull down a list of their friends, or likes, etc.

(Note that validity of the access token is time limited, so you can’t grab a copy of the access token and hope to use the same one day after day.)

Grabbing the link to your friends on Facebook is simply a case of opening a new project, choosing to get the data from a Web Address, and then pasting in the friends list URL:

Google Refine - import Facebook friends list

Click on next, and Google Refine will download the data, which you can then parse as a JSON file, and from which you can identify individual record types:

Google Refine - import Facebook friends

If you click the highlighted selection, you should see the data that will be used to create your project:

Google Refine - click to view the data

You can now click on Create Project to start working on the data – the first thing I do is tidy up the column names:

Google Refine - rename columns

We can now work some magic – such as pulling in the Likes our friends have made. To do this, we need to create the URL for each friend’s Likes using their Facebook ID, and then pull the data down. We can use Google Refine to harvest this data for us by creating a new column containing the data pulled in from a URL built around the value of each cell in another column:

Google Refine - new column from URL

The Likes URL has the form which we’ll tinker with as follows:

Google Refine - crafting URLs for new column creation

The throttle control tells Refine how often to make each call. I set this to 500ms (that is, half a second), so it takes a few minutes to pull in my couple of hundred or so friends (I don’t use Facebook a lot;-). I’m not sure what limit the Facebook API is happy with (if you hit it too fast (i.e. set the throttle time too low), you may find the Facebook API stops returning data to you for a cooling down period…)?

Having imported the data, you should find a new column:

Google Refine - new data imported

At this point, it is possible to generate a new column from each of the records/Likes in the imported data… in theory (or maybe not..). I found this caused Refine to hang though, so instead I exprted the data using the default Templating… export format, which produces some sort of JSON output…

I then used this Python script to generate a two column data file where each row contained a (new) unique identifier for each friend and the name of one of their likes:

import simplejson,csv



data = simplejson.load(open(fn,'r'))
for d in data['rows']:
	#'interests' is the column name containing the Likes data
	for i in interests['data']:
		print str(id),i['name'],i['category']

[I think this R script, in answer to a related @mhawksey Stack Overflow question, also does the trick: R: Building a list from matching values in a data.frame]

I could then import this data into Gephi and use it to generate a network diagram of what they commonly liked:

Sketching common likes amongst my facebook friends

Rather than returning Likes, I could equally have pulled back lists of the movies, music or books they like, their own friends lists (permissions settings allowing), etc etc, and then generated friends’ interest maps on that basis.

[See also: Getting Started With The Gephi Network Visualisation App – My Facebook Network, Part I and how to visualise Google+ networks]

PS dropping out of Google Refine and into a Python script is a bit clunky, I have to admit. What would be nice would be to be able to do something like a “create new rows with new column from column” pattern that would let you set up an iterator through the contents of each of the cells in the column you want to generate the new column from, and for each pass of the iterator: 1) duplicate the original data row to create a new row; 2) add a new column; 3) populate the cell with the contents of the current iteration state. Or something like that…

PPS Related to the PS request, there is a sort of related feature in the 2.5 release of Google Refine that lets you merge data from across rows with a common key into a newly shaped data set: Key/value Columnize. Seeing this, it got me wondering what a fusion of Google Refine and RStudio might be like (or even just R support within Google Refine?)

PPPS this could be interesting – looks like you can test to see if a friendship exists given two Facebook user IDs.

PPPPS This paper in PNAS – Private traits and attributes are predictable from digital records of human behavior – by Kosinski et. al suggests it’s possible to profile people based on their Likes. It would be interesting to compare how robust that profiling is, compared to profiles based on the common Likes of a person’s followers, or the common likes of folk in the same Facebook groups as an individual?

Comparing Columns in Google Refine

A reader (Cosmin Cabulea) writes: “I have two columns (A and B) and want to identify identical cells.”

I think I misapprehended the point of the question, but it prompted me to create this simple example.

In something like Google Spreadsheets, we could use an if statement to set the value of cells in a new column based on a comparison of the values of two other columns in the same row. In column C, cell C1, for example, we might use a formula of the form:


In Google Refine, we can use a GREL expression to achieve a similar effect. Create a new column and then use an expression of the form:

if(cells["A"].value == cells["B"].value, "similar", "different")

where A and B are the appropriate column headings.

Google refine - compare two columns

If you’re generating the new comparison column from one of the two columns you’re comparing (column with header B, say), you can reference the values of the original column directly:

if(cells["A"].value == value, "similar", "different")

Google refine compare two columns

It strikes me that the pattern scales to comparisons across multiple columns and of arbitrary complexity. For example, using a nested if control flow statement:

if( value == cells["Host"].value, if( cells["amount"].value > 75, 2, 1 ), 0 )

Or using a Boolean operator:

if( and( value=="May",cells['amount'].value > 0 ), 2, 0 )

Ref: Google Refine Basic blog: Compare values from two columns

Aggregating Values for Recurring Column Values

So this, it turns out (I think?!), was more in line with what Cosmin was after. Given something like:

1 2
1 3
2 1
2 3
2 4


1 2,3
2 2,3,4

Here’s a way of doing that using R (I use the R-Studio environment).

Using some (guess what) F1 data, loaded into the dataframe hun_2011proximity, let’s pull out a sample of laptime data (say the first 10 laps of a race), featuring just the car numbers, and the laptimes (ref: R: subsetting data). First we grab just those rows where the lap column value is less than 11, then we create a frame containing only a couple of the columns (car and laptime) from the dataset (the original hun_2011proximity data frame contained 20 or so columns, including two with headers car and laptime, and 70 laps worth of data):


(Thinks: would it be more efficient to do this the other way round, and reduce the data set to 2 cols first before extracting just the first 10 laps worth of data?)

samp2 now contains 240 rows describing 10 laps of data, each row containing data for one car from one lap; each row contains car and laptime data (2 cols).

Now we can run down one column, looking for recurring elements, and generate a new column that contains the aggregate values from another column for each unique element in the first column:

samp3=aggregate(samp2$laptime, samp2['car'],paste,collapse=',')

Here’s what we get as a result:

Combining common column value data elements

Ref: [R] aggregate text column by a few rows

A Couple of Alternative Approaches
Chatting to Cosmin, it turns out the actual requirement was to identify common followers of a set of Twitter accounts. So for example, with columns TwitterID FollowedBy, extract the unique FollowedBy Twitter IDs and then aggregate the TwitterID values (something like aggdata=aggregate(twData$TwitterID, twData[‘FollowedBy’],paste,collapse=’,’)).

One approach to this would be to look at the data in Gephi, plotting edges as a directed graph from FollowedBy to TwitterID, sizing the nodes according to out degree (so we could see how many of the target accounts each person in the union follower set was following). We could then use filters to reduce the set to just people following lots of the accounts.

Following this line of thought, we could also use a network flavoured representation (e.g. using something like networkx) to construct a graph and run stats on it. (So we could e.g. pull out reports describing the distribution of how many people were following how many of the target accounts, etc.)

Of course, on those occasions where the Google Social API returns Twitter follower names rather than redirect IDs, my Common Friends or Followers on Twitter hack will show common followers of two twitter accounts.

Yet another approach, if we have all the data in a single file, is to do a simple bit of counting using a Unix command line tool. For example, if we have comma separated file containing TwitterID (column 1) and FollowedBy (column 2) columns, we can sort the names in the FollowedBy column and count the number of times they reoccur:

cut -d "," -f 2 twitterdata.csv | sort | uniq -c

Ref: Playing With Large (ish) CSV Files, and Using Them as a Database from the Command Line

Merging Datasets with Common Columns in Google Refine

It’s an often encountered situation, but one that can be a pain to address – merging data from two sources around a common column. Here’s a way of doing it in Google Refine…

Here are a couple of example datasets to import into separate Google Refine projects if you want to play along, both courtesy of the Guardian data blog (pulled through the Google Spreadsheets to Yahoo pipes proxy mentioned here):

University fees data (CSV via pipes proxy)

University HESA stats, 2010 (CSV via pipes proxy)

Load one data file into a Google Refine/OpenRefine project called Merge Test A, the other into a separate project called Merge Test B.

We can now merge data from the two projects by creating a new column from values an existing column within one project that are used to index into a similar column in the other project. Looking at the two datasets, both HESA Code and institution/University look like candidates for merging the data. Which should we go with? I’d go with the unique identifier (i.e. HESA code in the case) every time…

First, create a new column:

Now do the merge, using the cell.cross GREL (Google Refine Expression Language) command. Trivially, and pinching wholesale from the documentation example, we might use the following command to bring in Average Teaching Score data from the second project into the first:

cell.cross("Merge Test B", "HESA code").cells["Average Teaching Score"].value[0]

Note that there is a null entry and an error entry. It’s possible to add a bit of logic to tidy things up a little:

if (value!='null',cell.cross("Merge Test B", "HESA code").cells["Average Teaching Score"].value[0],'')

Here’s the result:

Coping with not quite matching key columns

Another situation that often arises is that you have two columns that almost but don’t quite match. For example, this dataset has a different name representation that the above datasets (Merge Test C):

There are several text processing tools that we can use to try to help us match columns that differ in well-structured ways:

In the above case, where am I creating a new column based on the contents of the Institution column in Merge Test C, I’m using a couple of string processing tricks… The GREL expression may look complicated, but if you build it up in a stepwise fashion it makes more sense.

For example, the command replace(value,"this", "that") will replace occurrences of “this” in the string defined by value with “that”. If we replace “this” with an empty string (” (two single quotes next to each other) or “” (two double quotes next to each other)), we delete it from value: replace(value,"this", "")

The result of this operation can be embedded in another replace statement: replace(replace(value,"this", "that"),"that","the other"). In this case, the first replace will replace occurrences of “this” with “that”; the result of this operation is passed to the second (outer) replace function, which replaces “that” with “the other”). Try building up the expression in realtime, and see what happens. First use:
(what happens?); then:
and then:

The fingerprint() function then separates out the individual words that are left, orders them, and returns the result (more detail). Can you see how this might be used to transform a column that originally contains “The University of Aberdeen” to “aberdeen university”, which might be a key in another project dataset?

When trying to reconcile data across two different datasets, you may find you need to try to minimise the distance between almost common key columns by creating new columns in each dataset using the above sorts of technique.

Be careful not to create false positive matches though; and also be mindful that not everything will necessarily match up (you may get empty cells when using cell.cross; (to mitigate this, filter rows using a crossed column to find ones where there was no match and see if you can correct them by hand). Even if you don’t completely successful cross data from one project to another, you might manage to automate the crossing of most of the rows, minimising the amount of hand crafted copying you might have to do to tidy up the real odds and ends…

So for example, here’s what I ended up using to create a “Pure key” column in Merge Test C:

And in Merge Test A I create a “Complementary Key” column from the University column using fingerprint(value)

From the Complementary Key column in Merge Test A we call out to Merge Test C: cell.cross("Merge Test C", "Pure key").cells["UCAS ID"].value[0]

Obviously, this approach is far from ideal (and there may be more “correct” and/or efficient ways of doing this!) and the process described above is admittedly rather clunky, but it does start to reveal some of what’s involved in trying to bring data across to one Google Refine project from another using columns that don’t quite match in the original dataset, although they do (nominally) refer to the same thing, and does provide a useful introductory exercise to some of the really quite powerful text processing commands in Google Refine …

For other ways of combining data from two different data sets, see:
Merging Two Different Datasets Containing a Common Column With R and R-Studio
A Further Look at the Orange Data Playground – Filters and File Merging
Merging CSV data files with Google Fusion Tables

Fragments: Glueing Different Data Sources Together With Google Refine

I’m working on a new pattern using Google Refine as the hub for a data fusion experiment pulling together data from different sources. I’m not sure how it’ll play out in the end, but here are some fragments….

Grab Data into Google Refine as CSV from a URL (Proxied Google Spreadsheet Query via Yahoo Pipes)

Firstly, getting data into Google Refine… I had hoped to be able to pull a subset of data from a Google Spreadsheet into Google Refine by importing CSV data obtained from the spreadsheet via a query generated using my Google Spreadsheet/Guardian datastore explorer (see Using Google Spreadsheets as a Database with the Google Visualisation API Query Language for more on this) but it seems that Refine would rather pull the whole of the spreadsheet in (or at least, the whole of the first sheet (I think?!)).

Instead, I had to tweak create a proxy to run the query via a Yahoo Pipe (Google Spreadsheet as a database proxy pipe), which runs the spreadsheet query, gets the data back as CSV, and then relays it forward as JSON:

Here’s the interface to the pipe – it requires the Google spreadsheet public key id, the sheet id, and the query… The data I’m using is a spreadsheet maintained by the Guardian datastore containing UK university fees data (spreadsheet.

You can get the JSON version of the data out directly, or a proxied version of the CSV, as CSV via the More options menu…

Using the Yahoo Pipes CSV output URL, I can now get a subset of data from a Google Spreadsheet into Google Refine…

Here’s the result – a subset of data as defined by the query:

We can now augment this data with data from another source using Google Refine’s ability to import/fetch data from a URL. In particular, I’m going to use the Yahoo Pipe described above to grab data from a different spreadsheet and pass it back to Google Refine as a JSON data feed. (Google spreadsheets will publish data as JSON, but the format is a bit clunky…)

To test out my query, I’m going to create a test query in my datastore explorer using the Guardian datastore HESA returns (2010) spreadsheet URL ( which also has a column containing HESA numbers. (Ultimately, I’m going to generate a URL that treats the Guardian datastore spreadsheet as a database that lets me get data back from the row with a particular HESA code column value. By using the HESA number column in Google Refine to provide the key, I can generate a URL for each institution that grabs its HESA data from the Datastore HESA spreadsheet.)

Hit “Preview Table Headings”, then scroll down to try out a query:

Having tested my query, I can now try the parameters out in the Yahoo pipe. (For example, my query is select D,E,H where D=21 and the key is tpxpwtyiYZwCMowl3gNaIKQ; this grabs data from columns D, E and H where the value of D (HESA Code) is 21). Grab the JSON output URL from the pipe, and use this as a template for the URL template in Google Refine. Here’s the JSON output URL I obtained:

Remember, the HESA code I experiment with was 21, so this is what we want to replace in the URL with the value from the HESA code column in Google Refine…

Here’s how we create the URLs built around/keyed by an appropriate HESA code…

Google Refine does its thing and fetches the data…

Now we process the JSON response to generate some meaningful data columns (for more on how to do this, see Tech Tips: Making Sense of JSON Strings – Follow the Structure).

First say we want to create a new column based on the imported JSON data:

Then parse the JSON to extract the data field required in the new column.

For example, from the HESA data we might extract the Expenditure per student /10:

value.parseJson().value.items[0]["Expenditure per student / 10"]

or the Average Teaching Score (value.parseJson().value.items[0]["Average Teaching Score"]):

And here’s the result:

So to recap:

– we use a Yahoo Pipe to query a Google spreadsheet and get a subset of data from it;
– we take the CSV output from the pipe and use it to create a new Google Refine database;
– we note that the data table in Google Refine has a HESA code column; we also note that the Guardian datastore HESA spreadsheet has a HESA code column;
– we realise we can treat the HESA spreadsheet as a database, and further that we can create a query (prototyped in the datastore explorer) as a URL keyed by HESA code;
– we create a new column based on HESA codes from a generated URL that pulls JSON data from a Yahoo pipe that is querying a Google spreadsheet;
– we parse the JSON to give us a couple of new columns.

And there we have it – a clunky, but workable, route for merging data from two different Google spreadsheets using Google Refine.