Content uploaded by Phillip Brooker
Author content
All content in this area was uploaded by Phillip Brooker on Apr 09, 2018
Content may be subject to copyright.
1
PRE-PRINT VERSION
Published: Phillip Brooker , Julie Barnett , Timothy Cribbin & Sanjay Sharma (2015) ‘Have we
even solved the first 'big data challenge?' Practical issues concerning data collection & visual
representation for social media analytics’, in H. Snee et al. (Eds) Digital Methods for Social
Science: An Interdisciplinary Guide to Research Innovation. London: Palgrave MacMillan.
ISBN: 9781137453655.
Have We Even Solved the First 'Big Data Challenge?': Practical Issues Concerning Data
Collection and Visual Representation for Social Media Analytics
Abstract
The present chapter explores the technical and computational processes
through which social media data is shaped into research findings. The authors
make this argument by depicting the effects of two practical issues - API rate
limiting in Twitter data collection and the use of spatial mapping algorithms in
visualising those data - on resulting analyses. Such issues are not problematic
to social media analytic research; rather, they can be used as resources for
helping to characterise and understand the data at hand. Hence, the authors
work to demonstrate the value in incorporating these reflexive analyses of
technical and computational processes into our accounts; to advocate
thinking in assemblages as a requirement for making analytic claims with 'big'
social media data.
Introduction
Thanks to an influx of data collection and analytic software, harvesting and visualising 'big'
social media data
1
is becoming increasingly feasible as a method for social science
researchers. Yet whilst there is an emerging body of work utilising social media as a data
resource, there are a number of computational issues affecting data collection. These issues
may problematise any conclusions we draw from our research work, yet for the large part,
they remain hidden from the researcher's view. We contribute towards the burgeoning
literature which critically addresses various fundamental concerns with Big Data (see boyd
and Crawford, 2012; Murthy, 2013; and Rogers, 2013). However, rather than focus on
epistemological, political or theoretical issues – these areas are very ably accounted for by
the authors listed above, and others – we engage with a different concern: how technical
aspects of computational tools for capturing and handling social media data may impact our
readings of it. This chapter outlines and explores two such technical issues as they occur for
data taken from Twitter.
2
Throughout the chapter, we demonstrate a perspective consistent with Procter et al.'s
(2013) view that social researchers wishing to make sense of 'big' social media data should
have sufficient knowledge of the underlying concepts of the computational methods and
tools they are using, so as to be able to decide when and where to appropriately apply
them. Furthermore, we take heed of boyd and Crawford's suggestion that 'When
researchers approach a data set, they need to understand – and publicly account for – not
only the limits of the data set, but also the limits of which questions they can ask of a data
set and what interpretations are appropriate' (2012: 669-70). To this end, we highlight how
certain technical characteristics and constraints pertaining to the collection and processing
of Twitter data can impact on research and how an understanding of these factors might
lead to more robust accounts of such data.
Our aim is to demonstrate the mainstream relevance of a commonplace methodological
procedure in the social sciences; namely the self-critical reflexive analyses of our methods in
terms of their impact on our accounts of the subjects we study. Our goal here is to show the
importance of understanding the effects that technical processes may have on our readings
of data for all social scientists, not just for those with a background in computer science.
Without this understanding it is impossible to make sense of the data at hand. Hence, we
promote the idea of thinking of the investigative process as an 'assemblage' (Langlois, 2011;
Sharma, 2013) that draws together various social and technical (and other) factors into a
unified research process. Here, we refer to the ways in which the research process comes to
feature not only conceptual theoretical knowledge and inductive empirical findings, but also
how technical issues (such as API rate limiting and spatial mapping algorithms) contribute
towards the production of knowledge in multifarious complex ways. How such an
assemblage might operate will become clearer as we present our selected two technical
issues, and in the discussion that follows.
Reviewing the Field
The State of Social Media Analytics
With the field of social media analytics still in relative infancy there are few methodological
practices taken as standard. The tendency thus far has been to fit digital data to existing
'offline' ways of working. As O'Connor et al. note:
Often the online researcher has little in the way of research precedent to
use as a guide to practice online research and, as a result, online researchers
frequently turn to established offline practices (O'Connor et al., 2008: 276).
Working from this position, several authors characterise social media data as a special kind
of 'offline' social science data. For example, Hine argues that a key concern of social media
analytics is to avoid a loss of quality in data; 'Face-to-face interaction here becomes the gold
standard against which the performance of computer-mediated interaction is judged' (Hine,
3
2006: 4). This quality problem of social media data is a concern shared by many, with
comments being levelled at 'the lack of uniformity in how users fill in forms, fields, boxes
and other text entry spaces' (Rogers, 2013: 205); the representativeness and validity of the
data more generally (Tufekci, 2014); and the fact that the production of data is not
controlled and led by researchers but appears untamed 'in the wild' (Kitchin, 2013). Kitchin
notes:
The challenge of analysing Big Data is coping with abundance, exhaustivity
and variety, timeliness and dynamism, messiness and uncertainty, high
relationality, and the fact that much of what is generated has no specific
question in mind or is a by-product of another activity (Kitchin, 2014: 2).
Given the uncertain relationship between digital and 'offline' methods, it becomes
important to explore possible ways of rendering visible the characteristics of digital methods
to see how and where they fit into existing methodological practices. Our proposed
treatment of such data embraces the 'digitality' of researching in this area by advocating a
greater working familiarity with computational tools and methods. Emphatically, this is not
to say that the work of social media analytics can be reduced to the rote operating of
software (see Keegan (2013)). Kitchin summarises this tension:
For many...the digital humanities is fostering weak, surface analysis, rather
than deep, penetrating insight. It is overly reductionist and crude in its
techniques, sacrificing complexity, specificity, context, depth and critique for
scale, breadth, automation, descriptive patterns and the impression that
interpretation does not require deep contextual knowledge. (Kitchin, 2014:
8)
Yet we do not believe this is necessarily how social media analytics has to operate. On the
contrary, we advocate a mode of reading data that allows computational methods to pick
out areas of potential interest which then might be explored more intimately through closer
readings. To do this requires an understanding of how social media data can be affected by
technical processes as part of a wider assemblage. As Lewis et al. note, 'As scholars
increasingly take up such datasets, they may be better served by interweaving
computational and manual approaches throughout the overall process in order to maximise
the precision of algorithms and the context-sensitive evaluations of human coders' (2013:
49). By outlining how this kind of research process works 'on the shop floor', we hope to
foster a way of thinking about such technical issues which might facilitate the mainstream
usage of digital research methods generally.
We take steps towards respecifying 'big' social media work in this way by concentrating on
two issues. Firstly we demonstrate the effects of a data collection issue – the rate limiting of
Twitter's Application Programming Interfaces (APIs), which is a built-in restriction on the
4
flow of Twitter data that may interfere with analyses in significant ways. Secondly, we
remark upon the ways in which computational models (and the visualisations that represent
them) might shape our analytic readings of data. Those already working in the field are well
aware of these concerns, yet they do not routinely feature in published accounts of relevant
work. Consequently, such issues may stand as a barrier to entry by steepening an already
steep learning curve. Hence, we openly discuss two such issues, not necessarily as
problematic to social media analytics but as presenting an opportunity to make better use
of new and powerful data resources.
Addressing the First 'Big Data Challenge'
Before doing so it is useful to describe what we are referring to in the title as the First 'Big
Data Challenge'. One much vaunted promise of 'Big Data' is that we all now have the means
to access data from sources like Twitter, and to engage in analytic work on large data
corpora through processing that data into easily-digestible visualisations. Moreover, this
work does not necessarily require any special skill with computer science or programming –
there is a wealth of freely-available third-party software tools to do the 'behind the scenes'
technical work for us
2
. In this sense, the First 'Big Data Challenge' refers to a) having easy
access to big data, and b) the availability of tools that facilitate its analysis. Our question in
the title – whether or not we are yet in a position to close the book on this first challenge –
demonstrates our intention to probe such matters further: can we tap vast data resources
unfiltered? Is it really as simple as employing visual models to show us the results? Such
concerns are worked out through the process of doing social media analytics and acquiring
necessary relevant skills along the way. Our approach here is to more accountably explore
this process of doing social media analytics to help figure out what might count as
appropriate methods and methodologies.
But why is it important to render transparent what might be argued to be mundane
computational issues? Doing social media analytics with Twitter data necessitates an
interfacing with the mechanisms governing how users access Twitter data: the Twitter
Application Programming Interfaces (APIs). These APIs allow users to request to access
certain slices of Twitter data, according to various search strategies (i.e. by keyword, by
bibliographic/demographic information, by random selection, and so on). Moreover, once
investigators have personal copies of these data, they may then subject them to further
algorithmic processes to make their 'Big Data' analysable, e.g. in the rendering of statistical
information or in the production of visualisations and so on. In this way, computational
processes come to feature as essential elements in the production and construction of our
data and analyses. As Marres notes, this bringing together of different disciplinary ideas can
be equally productive and constricting:
[Digital] social research is noticeably marked by informational practices and
devices not of its own making, from the analytic measures built into online
5
platforms (e.g. the number of links, number of mentionings, follower counts),
to the visual forms embedded in visualization modules (the tag cloud). Online
social research is visibly a distributed accomplishment (Marres, 2012: 160).
This intertwining of technical issues and research methods is foregrounded for the social
sciences by Fielding and Lee, who argue that 'Social science has demonstrated how
technology both shapes and is shaped by social action. Research methods are no exception'
(2008: 505). As such, since there are computational processes governing data collection and
analysis, we may find ourselves better-armed to undertake research in the area if we
understand some of the finer points about how these tools and processes work. How
exactly do they restrict the data we can harvest? And how exactly do they shape the
statistics and visualisations and other analytic outputs which we use to understand that
data?
To this end, we now turn to a more pointed examination of two such issues – the possible
effects of API rate limiting on data collected through Twitter, and the possible effects of
spatial mapping algorithms on data visualisations – as they occur through the usage of a
bespoke social media analytics software suite, Chorus
3
. This serves to demonstrate the kinds
of issues investigators may find themselves contending with, as well as helping figure out
ways of handling them methodologically. Our reflexive focus on the research process itself is
very much a mainstream methodological practice of social scientists
4
– we seek to take a
self-critical view on the (opaque) process of undertaking research involving data collection
through the Twitter APIs and data visualisation using spatial mapping algorithms. However,
our approach sees such limitations not as obstacles to research to be overcome. Rather, we
discuss these issues as an exercise in learning the tools of the trade of social media analytics
and understanding how they construct the data we analyse, so as to be better able to deal
with them as part of our work.
Two Practical Issues
API Rate Limiting in Twitter Data Collection
Twitter data collection is a process mediated through Twitter's various APIs. For the
purposes of social media analytics, the APIs are the tools by which users can make requests
for specific types of data. This process of using Twitter's APIs to access data necessitates
that users write requests as a RESTful statement to return responses in a data interchange
format called JSON
5
, or that users take advantage of a third-party client which facilitates the
task for non-coders. However, though comprehensive collections of data are available for
purchase through data archiving services such as Gnip or DataSift, the Twitter APIs are not
completely openly available to users and developers. In fact, several restrictions on their
usage are in place; ostensibly this is to prevent misuse of the service by individual users.
One such restriction is Twitter's 'rate limiting' of an individual account's API usage, or the
6
rate at which a user can poll the API with requests for data matching a search query. Each
API has different rate limits and different software tools will handle those limits in different
ways
6
. Providing concrete and quantifiable definitions of these limits is, however, a hopeless
task. The ‘Search API’ (that our example below draws upon) is fairly well-documented in
terms of its limits, as we go on to discuss. However, the usage restrictions of all Twitter APIs
are variably dependent on contextual information; chiefly the volume of data any queries
will yield. Hence, for APIs other than Search, Twitter does not provide information as to the
exact limits they will impose. All of this makes understanding and navigating through the
Twitter APIs a labyrinthine process. We use data collected via the Search API – which
handles data based on keywords appearing in tweets and is the most commonly-used
Twitter API – to demonstrate how the API itself inevitably comes to feature in a burgeoning
assemblage built up by the research as it is undertaken.
At the time of writing, Twitter's Search API allows Chorus' data collection programme,
TweetCatcher Desktop (TCD), to make 450 requests every fifteen minutes. Each such
request has the potential to capture a maximum of 100 tweets. The Search API allows users
to retrieve historical data up to seven days prior to the initial request. On 23 July 2014, using
TCD we ran a very general search query for all usages of the term 'black', as a way of
exploring the topics and sub-topics of racialised tweet content. To capture the data, we
refreshed the query at various points over an approximately four-hour period so as to
ensure as comprehensive a dataset as the API would allow. This resulted in a dataset of
28,317 tweets featuring the term 'black'. Plotting the data over time in half-hour intervals
within Chorus' visual analytic programme TweetVis, it was clear that there were gaps in the
chronology of the conversation:
Figure 1: Chart to show volume of tweets mentioning “black” across time (half-hour
intervals)
What we see in Figure 1 is a striking reminder that Twitter's APIs are restricted, and that this
may have significant effects on the data we wish to capture through them. For high-volume
queries it is easy to come up against Twitter's API rate limits, such as during searches for
7
general terms like 'black', as well as for trending terms (e.g. 'Obama' in the run-up to the
2012 US presidential election). In this example, we were simply unable to keep up with the
pace of peoples' tweets; we were able to capture an average 118 tweets-per-minute over
the four hour period, whereas the actual conversation skipped along between 450-550
tweets per minute. Naturally, this left a sizeable chunk of data missing from our dataset (see
Figure 1). However, what is less immediately obvious in this rendering of the data are the
presence of other breaks in the flow of the conversation we captured, which become more
apparent when viewing at a finer level of granularity:
Figure 2: Chart to show volume of tweets mentioning “black” across time (two minute
intervals)
In Figure 2, we see the same data grouped into intervals of two minutes. Here we can
identify the same gap in the data as in Figure 1, but also an earlier gap which was previously
obscured when viewed with our earlier half-hour intervals. Hence, we now can detect a
probable disruption to the flow of data between 14:44:59 and 14:58:59 where only a
consistent chronology was visible (or at least presumed) before. It may simply be that
people tweeted fewer times during these minutes, though it is equally possible that it is at
this point we were reaching the limits of what the API would allow us to see. It is in fact
impossible to figure out what has happened from the data or visualisations themselves.
What does this ambiguity mean for social media analytics and social research involving
Twitter data? A key insight to draw from this demonstration is that comprehensive
collections of Twitter data are not freely available to researchers. Even where we may
assume we are capturing the entirety of a conversation, drilling down into finer levels of
granularity may show us otherwise. Furthermore, failing to recognise when these rate
limiting issues have occurred may be detrimental to the analyses we draw from our data.
Without due care and attention, we may find ourselves using falsely-derived
conceptualisations of data as chronologically consistent
7
.
8
It is important here to acknowledge that access to social media data is a highly politicised
issue largely driven by commercial concerns (boyd and Crawford, 2012; Rogers, 2013). In
this sense, it is a fallacy to believe that any data which is collected through Twitter's APIs
(rather than purchased) is complete: incompleteness and unrepresentativeness are
fundamental features purposefully built into the APIs to protect the primacy of Twitter's
approved data providers. Recognising, understanding and accounting for this is a key step in
acknowledging the research process as an unfurling assemblage of interconnected socio-
technical entities (of which the API is one, alongside any software and hardware used in the
undertaking of the work, the social media users whose posts make up the data, any social
theories we use to interpret the resulting findings, and so on). However, the incompleteness
and unrepresentativeness of social media data does not prevent us from accessing
meaningful insights. It is worth questioning our fetishising of data in this respect – what do
we need a chronologically complete dataset for? And what can we do without one? Rather
than bemoan the purposes for what our data cannot be used, it may be more productive to
explore what it can do. Though the methodological and analytic possibilities are impossible
to encapsulate fully in the present chapter (in that they will depend largely on the questions
being addressed), one such approach is advanced in the following section. However, the
salient point remains that perhaps the best way to make sense of data is to attain a deep
understanding of how a dataset has been constructed, and use that understanding as a
resource for designing appropriate analytic approaches with which it may be dealt.
Spatial Mapping Algorithms in Twitter Data Visualisation
Clearly there are issues concerning data collection of which researchers in social media
analytics would do well to be aware. However, our endeavours in the field have also
revealed similar technical issues in data visualization, where collected data is given an
analytic relevance through algorithmic processing. There are already multitudinous tools for
visualising social media data – Gephi, NodeXL and Chorus for instance. Some of these utilise
spatial mapping algorithms – computational processes through which entities such as
individual words or connected arrays of tweeters (or indeed any other kind of 'node') are
located on a 2D visual plane in relation to each other. Though each software package
operates uniquely, a unifying feature of these algorithms is their use of mathematical
reasoning as a way of representing distinctions between nodes. For example, the Chorus
visual analytic suite – TweetVis – features (amongst other visualisations) a topical 'cluster
map' which uses a spatial mapping algorithm to plot individual words contained within
tweets, in regards to the frequency of co-occurrences words have with other words in the
dataset. In this map, words which commonly co-occur in tweets cluster together, thereby
forming distinct topical clusters through which users can navigate and explore. Here, the
algorithm is what constructs and constrains the map, and for users trying to read the
visualisation, the constructions and constraints of the map become an integral part of the
resulting analysis. We demonstrate the possible effects of this algorithmic constructing and
9
constraining on analyses, showing how an understanding of the technical goings-on of a
data visualisation is a necessary requirement for those wishing to view it sensibly through
the lens of an assemblage.
To exemplify what the effects of a spatial mapping algorithm might look like in the
undertaking of a social media analytics project, we draw on previous work
8
on 'racialised
hashtags' (in particular, the hashtag #notracist). With a dataset collecting all usages of the
term #notracist over an eight-month period (resulting in 24,853 tweets), we plotted a
cluster map of hashtags, to see which hashtags featured together more commonly:
Figure 3 - hashtag map of #notracist (labels given to hashtags featuring in =>1% of tweets)
10
What this map showed us is that there are two distinct types of hashtag being used across
the #notracist dataset: firstly, a collection of 'comedy' hashtags (including #funny and #lol
and others) located in a tightly-organised cluster near the centre of the map; and secondly,
an array of 'truth' hashtags (including #justsayin, #justsaying, #iswear, #truth, #fact and
others) appearing on the fringes of topical branches around the outskirts of the
visualisation
9
. Given that these different themes are located in different areas of the map –
'comedy' in a tight central cluster and 'truth' out towards the fringes of branches – we
developed an interest in finding out what exactly this difference might be. To drill into the
data further, we filtered the master #notracist dataset into two sub-sets: one containing
'comedy' hashtags and one containing 'truth' hashtags. We then plotted cluster maps for all
terms contained within either dataset:
Figure 4 - 'Comedy' term map (left) and 'truth' term map (right) (labels given to terms
featuring in =>4% of tweets)
It is clear that these two visualisations are very different from one another – the 'comedy'
map is a messy aggregation of highly interconnected terms, whereas terms in the 'truth'
map are densely populated around the outer fringes of connecting branches. In order to
interpret what the two contrasting maps were telling us, we relied on an understanding of
what the algorithm had done with the data points. For the 'comedy' map, the terms used in
tweets were highly related to each other showing that #notracist comedy was a practice of
tweeting done with lots of different terms and hashtags being used in similar ways (i.e.
there are a number of 'stock' formats through which tweeters could accountably claim to be
'doing a joke'). However, for the 'truth' map, our understanding of what we could see in the
visualisation relied upon an understanding of the functions of Chorus' spatial mapping
11
algorithms. In this map, terms are chiefly located on the outer edges, as far apart from each
other as the algorithm will allow. This is demonstrated by the tree-like appearances of
topical branches, with virtually no connecting terms in the centre but a high density of
terms pushed out towards the edges of the 2D plane. This visible pushing of the boundaries
of the algorithm tells us a lot about the data we were working with. The terms used in the
'truth' tweets we identified are typically disconnected from each other and not used
together, and we can begin to characterise #notracist 'truth' tweets as evidence of a topic
that is not implicitly agreed-upon and which reflects a diverse array of strategies for
justifying a #notracist claim as a statement.
Again, we ask: what does this mean for social media analytics? Visualisations such as those
discussed here are designed to fit data into a model which serves to constrain our data and
analytic materials as well as give them visible structure
10
. Hence, our aim in describing the
processes through which such models construct and constrain analyses is to set out a
requirement for social media analytics that it explicitly account for these processes. For
example, in the #notracist data described above, it was fundamental to our understanding
of the data that we could recognise that 'truth' terms were pushing the Chorus cluster map
algorithm to its limits, and consequently use that information as a way of figuring out what
the 'truth' conversation was about. Crucially, the fact of our data being processed through
the algorithm in a certain way is precisely what helped us get to grips with what lay at the
heart of that data. Without this we would have been lost. Consequently we reiterate that
these intrusions of computational, technical and mathematical processes into our analyses
is not something to be resisted. Rather, they are necessary and productive elements of
social media analytics without which we would be unable to characterise the assemblage
through which the analysis had been shaped, and ultimately, unable to make adequate
sense of the data at hand.
Concluding Remarks
In this chapter, we have outlined how social media analytics incorporates data collection
and analytic work in ways which are thoroughly reliant on computational and technical
processes. Despite the provocative nature in which the question in the title of this piece was
posed, we believe investigators in the field are in a position to sensibly account for their
work in terms of these processes. Yet these issues are not something we have seen
discussed explicitly in the accounts produced thus far. Hence, though our points may be
frustratingly mundane to our peers (who may question why researchers would want to
write about the inner workings of APIs and algorithms) we nonetheless think it valuable to
discuss such things transparently as a means of promoting healthy and robust
methodologies for the emerging field.
To that end, rather than depict software tools in general and in abstract, we have
exemplified our ideas with reference to two specific issues arising out of the use of just one
12
software package. This, we hope, gives a flavour of the kinds of issues of which researchers
must be aware when working with digital data and associated software tools. It is our hope
that our accounts of two specific examples can demonstrate just how these kinds of issues
intersect with research work in a very direct way. Working in this way, we have shown how
technical and computational processes become a 'necessary evil' of the work. Only there is
nothing 'evil' about them. Rather, these same processes can be used as resources for
conducting (and figuring out how to conduct) analytic work in appropriate ways. However,
the use of these processes as analytic resources relies on our having a deep understanding
of what they are doing with our data, else we risk wrong-footing our analyses before we
even begin. As Manovich notes:
...you must have the right background to make use of this data. The
[analytic] software itself is free and readily available...but you need the
right training...and prior practical experience to get meaningful results
(Manovich, 2012: 472).
It is clear that it now becomes our job as researchers to equip ourselves with these
understandings of the technical processes on which our work relies, however much this may
take us outside of typical disciplinary boundaries.
All of this may make the analysis of social media data an infinitely more complex issue, in
that we are no longer really analysing only the data, but an assemblage (Langlois, 2011;
Sharma, 2013) of technical and social processes which coalesce to form the datasets and
visualisations we find before us. Concerning data collection, we have used the idea of an
assemblage to outline how the technical aspects of API rate limiting become built into social
media analytics research from the very beginning of the research process. Concerning data
visualisation and analysis, our described assemblage relied upon our conceptualisation of
computational processes as having (by necessity) a commitment to numerical
'understandings' of data and how those 'understandings' are translated into images to be
read by human researchers. We have no doubt that Chorus' way of doing things is only one
amongst many, and other such issues will invariably arise in a multitude of different ways
when doing social media analytics with other tools. As with any software tool, Chorus is not
'just a tool' – it engenders a particular way of thinking about social media data which
constructs and constrains analyses in equal amounts. As such, the modest goal of this
chapter has been to encourage readers to consider their research work not only in terms of
the results and findings to be drawn, but in relation to the myriad processes through which
those findings are mediated throughout the endeavour. We advocate thinking in
assemblages as a requirement for social media analytics generally. Furthermore, we have
shown the relevance of assemblages for mainstream purposes, and how the specific
properties of an assemblage might be uncovered through the deployment of a key
methodological principle – reflexivity – which has informed the present chapter from start
13
to finish. In using the idea of assemblages as a frame for undertaking investigative work,
analytic findings would be explicitly situated alongside deconstructions of the processes by
which tools are governed by big data and the processes by which those same tools govern
the generation of empirical findings. In this sense, we may find ourselves in the business of
handling data processes rather than data, and of reading visualisation processes rather than
visualisations. The final result of these processes – the compiled dataset or the visual
representation – are not objects in and of themselves, but are better thought of as a way of
demonstrating how an unfolding combination of human and computational research
processes has resulted in a selection of valid and defensible findings.
14
Bibliography
boyd, d. and K. Crawford (2012) 'Critical Questions for Big Data', Information,
Communication & Society, 15(5), 662-79.
Fielding, N. and R. Lee (2008) 'Qualitative e-Social Science/Cyber-Research' in N. Fielding, R.
Lee and G. Blank (eds.) The Sage Handbook of Online Research Methods. London:
Sage, pp. 491-506.
Hine, C. (2006) 'Virtual Methods and the Sociology of Cyber-Social-Scientific Knowledge' in
C. Hine (ed.) Virtual Methods: Issues in Social Research on the Internet. Oxford: Berg,
1-13.
Keegan, B. (2013) 'C-Level Executives Cry Out for Data Scientists', ComputerWeekly.com,
[ONLINE] Available at: http://www.computerweekly.com/news/2240205984/C-level-
executives-cry-out-for-data-scientists [Accessed 13 November 2014].
Langlois, G. (2011) 'Meaning, Semiotechnologies and Participatory Media', Culture Machine,
12, 1-27.
Lewis, S. C., R. Zamith & A. Hermida (2013) 'Content Analysis in an Era of Big Data: A Hybrid
Approach to Computational and Manual Methods', Journal of Broadcasting &
Electronic Media, 57(1), 34-52.
Lynch, M. (2000) 'Against Reflexivity as an Academic Virtue and Source of Privileged
Knowledge', Theory, Culture & Society, 17(3), 26-54.
Kitchin, R. (2014) 'Big Data, New Epistemologies and Paradigm Shifts', Big Data and Society,
1, 1-12.
Manovich, L. (2012) 'Trending: The Promises and Challenges of Big Social Data', in M. K. Gold
(ed.) Debates in the Digital Humanities. London: University of Minnesota Press, 460-
75.
Marres, N. (2012) 'The Redistribution of Methods: On Intervention in Digital Social Research,
Broadly Conceived', The Sociological Review, 60(S1), 139-65.
Murthy, D. (2013) Twitter. Cambridge: Polity Press.
O'Connor, H., C. Madge, R. Shaw and J. Wellens (2008) 'Internet-Based Interviewing' in N.
Fielding, R. Lee and G. Blank (eds.) The Sage Handbook of Online Research Methods.
London: Sage, 271-89.
15
Procter, R., F. Vis and A. Voss (2013) 'Reading the Riots on Twitter: Methodological
Innovation for the Analysis of Big Data', International Journal of Social Science
Research Methodology, 16(3), 197-214.
Rogers, R. (2013) Digital Methods. London: The MIT Press.
Sharma, S. (2013) 'Black Twitter? Racial Hashtags, Networks and Contagion', New
Formations, 78, 46-64.
Tufekci, Z. (2014) 'Big Questions for Social Media Big Data: Representativeness, Validity and
Other Methodological Pitfalls', Proceedings of the Eighth International AAAI
Conference on Weblogs and Social Media, AAAI Publications, 505-14.
1
We take 'big' social media data to refer to volumes of data too large to handle without computational
processing and which are derived from peoples' everyday usages of social media platforms such as Twitter.
2
The idea that social media analytics requires no specific skill in its practitioners is contestable – for instance,
Keegan (2013) notes that the information technology industry believes itself to be suffering from a lack of
trained data scientists. However, the point remains that there are lots of freely available social media analytics
tools with which investigators from any discipline can explore data, and it is no longer a steadfast requirement
for practitioners to have any significant skills in programming, data mining, data visualisation, and so on.
3
Chorus (see www.chorusanalytics.co.uk for further details) is a free-to-download data collection and visual
analytic software suite dealing with Twitter data for social media analytics. Chorus was developed at Brunel
University by a team including several of the authors of the present chapter (Dr. Tim Cribbin, Dr. Phillip
Brooker and Prof. Julie Barnett). The development of Chorus was supported in part through the MATCH
Programme (UK Engineering and Physical Sciences Research Council grants numbers GR/S29874/01,
EP/F063822/1 and EP/G012393/1).
4
Lynch (2000) however questions the utility of sociology's concern with self-analysis, arguing that it only need
be applied when something particularly interesting will result (as is the case with his own reflexive approach to
reflexivity, and we hope, the present chapter).
5
A RESTful statement is one which is written in adherence with REST (or Representational State Transfer)
principles, REST being the ubiquitous architectural style that standardises and underlies the world wide web. In
regard specifically to handling the Twitter APIs, RESTful statements are the commands by which API users can
speak to Twitter's servers to request specific slices of data, which are returned in JSON format. Readers
wishing to find out more about using the Twitter APIs and writing API requests manually should start with
Twitter’s own developer documentation, available at: https://dev.twitter.com/docs [accessed: 29 July 2014].
6
See https://dev.twitter.com/rest/public/rate-limiting [accessed: 17 November 2014] for a more detailed
account of the rate limiting Twitter applies to its APIs.
7
Other APIs may provide something more like a chronologically complete timeline – for instance, the Twitter
Streaming API pushes ‘real-time’ data matching a query’s criteria. However, the Streaming API only provides a
percentage sample of tweets requested, where the actual percentage is unknowable and dependent on the
volume of tweets requested by the query and concurrent Twitter traffic. Hence, the only way to ensure
comprehensivity of a dataset without running into sampling issues is to purchase Twitter ‘Firehose’ data – this
alone politicises access to data to the extent that few can afford to ever see a comprehensive dataset.
8
See http://www.darkmatter101.org/wiki/notracist_twitter [accessed: 21 January 2015] for an informal
account of this project.
9
Though this chapter is not intended as an empirical study of these tweets, interested readers might wish to
note that the 'comedy' hashtags we identified were tweets designed by tweeters to be humorous, whereas
'truth' hashtags were designed to enforce a point that a tweets was 'just a fact' or 'just an observation' and so
on. Our analytic work explored the different practices through which users attempted to justify their claims
that a tweet was not racist by virtue of it being a joke or a statement of fact.
16
10
This might be likened to filling a glass with water. As with water taking the shape of the glass it is poured
into, the process of collecting and visualising social media data serves to furnish amorphous data with a
structure. However, as much as these technical processes construct data such that they become amenable to
analysis, it can be said that the same processes constrain data into singular readings – a circular glass gives
only a circular shape to the water, but what if other shapes would prove more interesting?
