Amoore - Life Beyond Big Data Governing With Little Analytics

Economy and Society
ISSN: 0308-5147 (Print) 1469-5766 (Online) Journal homepage: http://www.tandfonline.com/loi/reso20
Life beyond big data: governing with little analytics

Louise Amoore & Volha Piotukh
To cite this article: Louise Amoore & Volha Piotukh (2015) Life beyond big data: governing with
little analytics, Economy and Society, 44:3, 341-366, DOI: 10.1080/03085147.2015.1043793
To link to this article: http://dx.doi.org/10.1080/03085147.2015.1043793
Published online: 04 Sep 2015.
Submit your article to this journal
Article views: 1922
View related articles
View Crossmark data
Citing articles: 5 View citing articles
Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=reso20
Download by: [92.176.7.158]
Date: 11 January 2017, At: 09:22
Economy and Society Volume 44 Number 3 August 2015: 341366

http://dx.doi.org/10.1080/03085147.2015.1043793
Life beyond big data:

governing with little
analytics
Louise Amoore and Volha Piotukh
Abstract
The twenty-first-century rise of big data marks a significant break with statistical
notions of what is of interest or concern. The vast expansion of digital data has
been closely intertwined with the development of advanced analytical algorithms
with which to make sense of the data. The advent of techniques of knowledge
discovery affords some capacity for the analytics to derive the object or subject
of interest from clusters and patterns in large volumes of data, otherwise
imperceptible to human reading. Thus, the scale of the big in big data is of less
significance to contemporary forms of knowing and governing than what we will
call the little analytics. Following Henri Bergsons analysis of forms of perception
which cut out a series of figures detached from the whole, we propose that
analytical algorithms are instruments of perception without which the extensity
of big data would not be comprehensible. The technologies of analytics focus
human attention and decision on particular persons and things of interest, whilst
annulling or discarding much of the material context from which they are
extracted. Following the algorithmic processes of ingestion, partitioning and
memory, we illuminate how the use of analytics engines has transformed the
nature of analysis and knowledge and, thus, the nature of the governing of
economic, social and political life.
Keywords: analytics; algorithm; big data; knowledge discovery; Bergson;
technology.
Louise Amoore, Department of Geography, Durham University, South Road, Durham DH1
3LE, United Kingdom. E-mail: [email protected]
Volha Piotukh, Department of Geography, Durham University, South Road, Durham DH1
3LE, United Kingdom. E-mail: [email protected]
Copyright 2015 Taylor & Francis
342 Economy and Society

To sew the pieces together
In an amorphous space he carves out moving figures, or else, he imagines
relations of magnitude which adjust themselves one to another [] But it is not
enough to cut out, it is necessary to sew the pieces together. You must now
explain how those qualities which you have detached from their material
support can be joined to it again (Bergson, 1912, p. 32).
The advantage is we can throw all the data at algorithms and the algorithm sorts
it, picks out the strongest relationships (SAS Analytics, 2013).
In a crowded analytics workshop in London in 2013, a data analyst

demonstrates the techniques that may allow a user, without having to code,
to segment data, to rejoin data together and to get insight into that data. In
a world of big data, he tells the assembled crowd, what matters is the capacity
for businesses and governments to make sense of the data, to throw it at
algorithms such that the strongest relationships between elements can be
identified. Understood in these terms, the work of the analytics is concerned
with cutting out pieces from across a vast array of data sources and types,
before stitching them together in a composite with other data elements.
There can be little doubt that the very idea of big data is having significant
consequences for economy and society, and for human knowledge whether
in the petabytes of scientific data generated by the Large Hadron Collider at
CERN, or in the ongoing debates on the social sciences use of transactional
data for the understanding of human behaviour and social transformation.1
The widely held view that we are living in a world dominated by the 4 Vs
of big data increased volume, variety, velocity, veracity has led to a focus
on the significance of the scale and scope of the digital traces left in the wake
of people, things, money and ideas on the move (Boyd & Crawford, 2012).
As data-generating devices proliferate, and as data storage and processing
power have become more scalable, the rise of twenty-first-century big data
has been described as a goldmine of magical material, a new oil fuelling
innovative forms of economic transaction and circulation whose core assets
are data (Kroes, 2013; OECD, 2013).
And yet, what precisely is meant by the concept big data? What distinguishes,
for example, our contemporary period of vast quantities of digitized data from
what Ian Hacking (1982) more specifically observes as the avalanche of
statistical numbers of the nineteenth century?2 Hackings detailed analyses of
the emergence of statistical knowledge of the trends, rates and patterns
of populations emphasize the importance not only of the novel availability
of data on nineteenth-century population, but also the scientific and calcula
tive techniques that rendered the data available for the making up of people
by the state (1986). Hacking is attentive to the intimate relationship between
rationalities and technologies of governing, and mechanisms for calculat
ing, intervening and acting upon the world (see also Miller & Rose, 1990).
Louise Amoore and Volha Piotukh: Life beyond big data
343
It is precisely such reflection on the situated calculative techniques and processes

used in the gathering, analysis and deployment of big data that we find to be
absent from the contemporary ubiquitous use of the term.
As Bruno Latour notes, the giant in the story is not necessarily a larger
character, or a figure with greater agency, than the dwarf (1988, p. 30). The
spatial scale of the object, we might say with Latour, tells us little about the
capacities or agency that it assembles around it. Notwithstanding such
problems of ontological scale in the designation of big/small, for our purposes
there are two significant aspects to the big in big data. First, big data pushes at
the limits of traditional relational databases as tables of rows and columns, and
requires new ways of querying and leveraging data for analysis, in addition
to the structured query language (SQL) built for relational databases in the
1970s. As larger volumes of data of more diverse types become available for
analysis, algorithmic tools have developed in parallel for the bulk processing
and analysis of stored data at rest, and for so-called real-time analysis of
streaming data in motion (Gupta et al., 2012, p. 43). Second, big data is big to
the extent that it exceeds and changes human capacities to read and make sense
of it. Specifically, the contemporary pluralization of data forms exceeds the
linear techniques of punch card indexes, population sampling and probabilistic
calculation characteristic of Ian Hackings variant of large-scale printed
number. The forms of sense-making that grew up around the collection of
demographic data, such as census, and epidemiological data, such as statistical
rates of mortality, and that persisted in early forms of computing, where the
computer was a human who calculates, are transformed with new forms of
human and machine reading of data (Hayles, 2005, 2012). The study of the
form of sense-making that dominates twenty-first-century big data analysis
must also address, then, the dwarf in the story the little analytical devices
without which the giant of big data would not be perceptible at all.
Perception and the little analytics

Writing on perception, Henri Bergson describes a transformative scene from
fairyland, in which as by a magicians wand perception is conjured so that
it may have nothing in common with the matter from which it started (1912,
p. 32). For Bergson, the inescapable problem of the insufficiency of our faculties
of perception to capture infinite variation is one shared, albeit differently,
by physics and metaphysics, by science and philosophy (1965, p. 132).3 The
difficulty of the problem, as described by Bergson, is that we imagine
perception to be a kind of photographic view of things, taken from a fixed
point by an apparatus, a device or an organ of perception (1912, p. 31). The
knowledge of a material world given to us by perception works a dividing up of
matter that is always too sharply defined, always subordinated to practical
needs, whilst our science, aspiring to the mathematical form, over-accentuates
the spatiality of matter (Bergson, 1965, p. 211). In short, for Bergson perception

is attuned to action, to the carving out of a series of still images, a bringing
to attention that allows action to take place. To be clear, it is not the case that
data analytics can be considered analogous to Bergsons many and varying
accounts of the faculties of perception, but that all forms of perception, human
and non-human, physical and metaphysical, are intractably bound up with
attention and action, with how an image of something of interest is brought to
attention for action (Crary, 1999). The practices of data mining and analysis
institute curious forms of perception, drawing as they do so extensively on
the spatial methods of the mathematical and physical sciences (Mackenzie, 2011;
Parisi, 2013), and yet claiming to produce an enhanced version of perception,
to capture a world of flux, a totality of images, data streams and becoming.4
The question for Bergsons philosophical method, then, is not how
perception arises, but how is it limited, to know how and why this image
is chosen to form part of my perception, while an infinite number of other images
remain excluded from it (1912, p. 34). The task at hand, as Bergson understands
it, is to give up your magicians wand and follow the process to the end,
to understand how a perception that should be the image of the whole becomes
limited and reduced to the image of that which interests you (1912, pp. 3536).
The work of data analytics processes appears to institute a specific form
of carving out and extraction from a broader extensity of big data audio files,
social networking site text, binary image files, GPS data, RFID tag reader data,
digital public records and so on and the stitching together with other data
elements. Like Bergsons fairyland scene, where an image of interest is extracted
from a whole, data analytics are instruments of perception: they carve out images;
reduce heterogeneous objects to a homogeneous space; and stitch together
qualitatively different things such that attributes can be rendered quantifiable.
If the metaphor of big data is to continue to dominate the governing of digital
life, then it cannot be understood without the little analytics that make
data perceptible. As Katherine Hayles suggests, data that are unimaginable in
their totality and too vast to comprehend are rendered more or less tractable
by the algorithms that make them searchable (2012, p. 230). In effect, one of
the many problems with a pervasive focus on big and data is that the finite and
granular minutiae of the analytics are overlooked. From the financial subjects
online access to mobile and adaptive credit scoring (Marron, 2007), to the
business and marketing analysis of personal data generated via mobile devices
about their users (OECD, 2013), to the use of algorithmic and automated border
controls (Amoore, 2011), it is the work of the little analytics that makes ever more
finite interventions in the governing of life itself (Rose, 2006). In the discussion
that follows we focus on how the work of analytics and algorithms not only
transforms the meaning and value of data, but also inscribes the very perception
of the world in which we live, govern and are governed. Animated by Bergsons
injunction to give up the magicians wand and follow the process, we discuss
three processes through which little analytics reshape the landscape of what
can be perceived, known and acted upon: ingestion, partitioning and memory.
345
Ingestion: n=all
Reporting on the rise of analytics-based decision-making, the consultants
Accenture urge their business clients to move beyond traditional sources
of data and seize the opportunities for new insights created by new sources,
such as text analytics from social media and digital interactions (Accenture,
2013, p. 5). What is captured here is a double transformation in the landscape
of big data: a radical expansion in the forms of social interaction and transaction
that can be rendered as data, or what Victor Mayer-Schnberger and Kenneth
Cukier (2013) call datafication, coupled with a novel capacity to analyse across
a variety of types of data. In short, the rise of big data witnesses a transformation
in what can be collected or sampled as data, and how it can be rendered
analysable. In the vocabulary of the computer scientists and data analysts, data
are no longer strictly collected, but rather are ingested, such that everything
becomes available to analysis, the sample being represented as infinite, or n=all.5
In the past, conventional forms of structured data, characterized by numbers,
tables, rows, and columns (Inmon & Nasaevich, 2007, p. 1), were the only
forms of data to inhabit the world of databases, spreadsheets and statistical
tables, and thus were the only data that could be leveraged for analysis. In many
ways, the distinction between structured and unstructured data that dominates
data science discourse and social science accounts is profoundly misleading.
Of course, we might say that all data declared to be unstructured is always
already structured, and certainly remains structured in important ways within
data architectures and digital devices (Berry, 2014; Kitchin, 2014). Yet, while
structured data is territorially indexable, in the sense that it can be queried
on the horizontal and vertical axes of spreadsheets within databases, so-called
unstructured data demands new forms of indexing that allow for analysis to
be deterritorialized (conducted across jurisdictions, or via distributed or cloud
computing, for example) and to be conducted across diverse data forms
images, video, text in chat rooms, audio files and so on.6 In the main this has
implied making unstructured data analysable by the establishment of links
with already indexed structured data and the creation of new indexes.
So, for example, IBMs predictive policing software uses content analytics
that promise to: search and analyse across multiple information sources,
extracting key pieces of information like new addresses, credit cards or passports
that can help resolve identities, build relationship networks and trace patterns
of behaviour (IBM, 2012, p. 2). The linking of the data elements is performed
through joins across data from different data sets, either on the basis of
direct intersections with already indexed data (e.g. via a phone, credit card or
social security number ingested from a database), or probabilistically, through
correlations among data-points from different sources (e.g. text scraped from
a Twitter account correlated with facial biometrically tagged images drawn
from Facebook). Though in many ways the use of the join is not novel and is
commonly used for querying relational databases, today the analysis operates
with a much more diverse pool of data. The allure of unstructured data is that it

is thought to contain patterns heretofore unseen and, therefore, a wealth of
previously hidden insight. The growing use of analytics capable of reading and
making sense of data, of unlocking its potential, is tightly interwoven with
a world of promise and opportunity thought to be buried in a text in need of
an index (Inmon & Nasaevich, 2007). We can sense here a desire for wholeness,
an embrace of the total and comprehensive that ceaselessly generates a politics
of mash-ups, compilation and assemblage (Ruppert et al., 2013, p. 38).
Following the process of ingestion: text and sentiment analysis

Consider a global pharmaceutical company conducting web listening for
sentiment analysis of open-source Twitter and online epilepsy support social
network data.7 Of the 60,000 epilepsy sufferers in the United Kingdom, the
majority take one of two major commercial drugs, one of which is produced
by a global pharmaceutical company, we will call them Alpharm, who wish to
understand the churn in patients moving to the alternative drug. As a problem
addressed by algorithm, this kind of churn is not different ontologically from
what early adopters of sentiment analysis, such as credit card companies or
mobile phone companies, were analysing with respect to customers switching
to new providers, for in essence they all wish to understand what kinds of
sentiments or tendencies signal specific human behaviours. In the case of
Alpharm, text analytics based on R programming language were run in opensource data to build what was described as a whole picture of the relationships
between people, their medication, their family life, their moods and feelings,
their perceptions of disease and the things they do to cope with side-effects
of the epilepsy medication. The text analytics used parsing algorithms to split
the sentences into tokens, which could be words, numbers or symbols, and
stemming algorithms to reduce the words to their base or root (so, for example,
the words thinks, thinkers and thinking were stemmed to a common root think).
A network analysis was then conducted to reveal significant nodes and links
such as, for example, the association between a drug brand-name and a
particular side-effect, or between an affective quality such as fear or panic and
a significant life event. Questions that were asked of the data via the analytics
included: how do people live with their disease?; what are the notable
patterns in the therapeutic use of music, sport or alcohol?; who are the key
opinion leaders in this social network? Significantly, these were not thought
to be queries that Alpharm could make of the more conventional, linear and
structured data otherwise available via patient or medical practitioner surveys.
The subtleties of affects such as anger, rage, depression, melancholy or anxiety
were thought to be retrievable via the social media data in a way that accesses
and reads life in its very emergence, in the unfolding of life.
What one can see in the use of text analytics and sentiment analysis is not
merely a world of more freely available big data, or n=all, but more specifically
a distinct mode of gathering and reading that data. It is the gathering and
347
reading that form part of the work of the little analytics. At first glance,
text analytics do not appear dissimilar from reading as such, and, indeed, the
genesis of text mining has its roots in natural language processing and semantic
structure. However, as Katherine Hayles has argued persuasively, machine
reading is a specific kind of reading that not only allows algorithms to read
text, but also alters irrevocably the way humans read and, consequently,
the way humans think and perceive (2012, pp. 2829). What matters is thus
not strictly whether machines may somehow read like humans, but rather how
the possibilities of digital forms, such as text analytics, change the practice
of reading for humans and machines alike.8 The hyper reading that Hayles
identifies among multiple forms of human and machinic reading consists of
skimming, scanning, fragmenting, and juxtaposing texts, being a mode
of reading attuned to an information intensive environment (2012, p. 12).
The reading involved in text analytics, engaged on the part of the algorithms
and the humans who action a query, is just such hyper reading of multiple
forms and sources of data as though they were a single text.
As in our Alpharm example, in order for the particular object of interest to
be perceptible, a certain damage is done to words and syntax, and to context.
Consider the processes necessary for text analytics to read: the removal of stop
words, including and, prepositions, gender suffixes in some languages, and
the definite and indefinite articles the and a; stemming, whereby words are
reduced to their stems; and the removal of punctuation marks and case
sensitivity. In effect, as one sees in the pharmaceutical companys scraping
of the web for a complete life story of a person, in order for a life to be read
with data analytics, any trace of a context, movement or a story that has a
recognizable narrative must first be pruned out. As Hayles points out, there
remain important differences between narrative-based stories of literature and
data-based story-telling:
The indeterminacy that databases find difficult to tolerate marks another way in
which narrative differs from database. Narratives gesture toward the inexplicable,
the unspeakable, the ineffable, whereas databases rely on enumeration, requiring
explicit articulation of attributes and data values. (Hayles, 2012, p. 179)
The parsing and stemming of text, then, is intrinsic and necessary to the
capacity of analytics to read at all. The stories about the lives of epilepsy
sufferers, or the purchases of retail loyalty-card holders, or the transactions
of online banking customers that can be read by algorithms are not the
indeterminate narratives of life stories. They are lives that are flattened and
reduced to their common stems, connected with others only through correlations, links and associations. On the basis of these analytics-derived life stories,
decisions are made about people, policies are implemented, resources are
allocated and interventions are targeted.
Because text analytics and sentiment analysis conduct their reading by a
process of reduction to bases and stems, their work exposes something of the

fiction of a clear distinction between structured and unstructured data. Through
processes of parsing and stemming, everything can be recognized and read as
though it were structured text. The significance, then, lies not merely with some
newly abundant unstructured data stream, but with the process of ingestion
itself. From the Latin in-gerere, to carry into, to ingest suggests the drawing in
of quantities of matter into an engine or body, such that the contents can be
filtered, some of them absorbed and others expelled or discarded. For instance,
in his first interview with the film-maker Laura Poitras, Edward Snowden refers
to the ingestion by default of bulk communications data by the National
Security Agency (NSA), wherein the software absorbs that which has value
or interest.9 When one hears government statements that the haystack is
required in order to target the needle, or that analytics are akin to using a
magnet to draw the needle out of the haystack instead of combing through the
straw yourself, it is the process of ingestion that is key (Intelligence and Security
Committee, 2014, 2015).
The analytics solutions that inhabit this world of ingestion such as TIBCO
Spotfire 5.5, capable of analysing 30 different types of data simultaneously,
or IBM Content Analytics, working with 30 sources and 150 formats can
read the data only because they are indifferent to the qualitative differences that
dwell within heterogeneous data. What is most significant about the process
of ingestion for us is not merely the volume of the data that can be drawn into an
analytics engine, but how an object or person of interest emerges via ingestion,
how the target is identified from the mass. Returning to Bergson, who writes
about a form of ingestion, albeit one where plants and animals absorb nutrients in
ways that care little for individual differences (1912, p. 206):
Hydrochloric acid always acts in the same way upon carbonate of lime [] Now
there is no essential difference between the process by which this acid picks out
from the salt its base, and the act of the plant which invariably extracts from the
most diverse soils those elements that serve to nourish it. In short, we can follow
from the mineral to the plant, from the plant to the simplest conscious beings,
from the animal to the man, the progress of the operation by which things
and beings seize from out of their surroundings that which attracts them, that
which interests them practically [] simply because the rest of their
surroundings takes no hold upon them. (Bergson, 1912, pp. 207208)
The analytics promise to leverage all types of data stored across multiple
architectures in order to unveil things that could not otherwise be seen, the
previously unseen, hidden patterns that dwell in the folds and joins between
data forms. Yet, if we understand the work of the analytics in seizing from
the surroundings that which interests or sustains, then we begin to see how
qualitative differences between data forms become obscured by the pursuit of
the object of interest.10 The analytics extract from diverse elements that which
is of interest, indifferent to the heterogeneity that surges beneath that data.
Viewed in this way, the contemporary big data question of how to approach
349
n=all is posed rather differently. In contrast with a word of big data that seeks
out complete data sets never available before (interview 1 October 2013) and
where big data wants n, nothing else (Hildebrandt, 2013, p. 6), n=all appears
instead as an impossible claim. The process of ingestion draws in the data
rather as Bergsons hydrochloric acid acts upon chalk, or a plant acts on
diverse nutrients in the soil, that is to say indifferent to the all with which
it communes. In this specific sense n will never be equal to all. In the so-called
flat files of analytics algorithms, which quite literally flatten the multiple
distinctions among data forms in order to make the data readable and
analysable, the complex temporalities of the life that generated the data are
entirely lost.
Partitioning: transform, select and filter the variables
As IBM describe their Intelligent Miner software, the task of analytics
algorithms is to extract facts, entities, concepts and objects from vast
repositories (2012, p. 2). Understood thus, the work of the analytics can
be conceived as one specific from of sense-making one means by which
subjects and objects of interest are partitioned from a remainder and singled
out for attention. How are qualitatively different entities in a heterogeneous
body of data transformed into something quantitative, something that can be
enumerated? In his early work Henri Bergson differentiates between two ideas
of time, the time of lived experience, or dure relle, and the mechanistic time
of science in which time is a succession of images or spatial frames, as in film
(Ansell Pearson & Mullarkey, 2002, p. 17). In this spatial representation of time
as a series of halts, we begin from a fixed point in the immobile to watch for the
moving reality as it passes instead of putting ourselves back into the moving
reality to traverse with it (Bergson, 1965; see also Connolly, 2011). Understood
thus, the fixed instrument of perception partitions, according to what is of
interest to it, a series of immobile stills from which to derive some picture of
a changing world.
As Gilles Deleuze notes, Bergson calls into question the order of needs,
of action, and of society that predisposes us to retain only what interests us
in things and that tends to obscure differences in kind (Deleuze, 1991, p. 33).
Following Bergson, Deleuze understands the qualitative multiplicity of
duration to bear all of the differences in kind, while space is unable to
present anything but differences of degree (since it is quantitative homogeneity) (1991, p. 31). The patterns of life, so readily claimed as the world captured
by analytics, might be properly thought of as durational, multiple, continuous
and qualitative. Like the modern physics Bergson and Deleuze describe, the
analytics extract and detach data from the whole, drawing a series of dis
continuous spatial images as vantage points on a mobile world. While analytics
claim to afford a vantage point on emergent life patterns and tendencies,

in practice they spatialize time and substitute differences in kind for differences
in degree, collapsing qualitative difference into enumeration and action.
Following the process of partitioning: MapReduce

Let us consider more closely the form of partitioning at work in one widely
deployed programming model that uses distributed or parallel computing,
MapReduce. Originally designed by Google in order to transform the indexing
of Google webpages, MapReduce is a framework for parallel processing across
vast data sets (Dean & Ghemawat, 2004). In the context of data proliferating
in different forms and across different databases and servers, the challenge,
as understood by data scientists, is said to be the feasibility of reasoning over
such large volumes of data (Tachmazidis et al., 2012). Commonly used in opensource software such as Apache Hadoop, the MapReduce architecture makes
it feasible to analyse data precisely via its distributed form, by dividing
computation into two distinct phases.
The first map step breaks down the data into manageable pieces, subdividing
the aggregate problem into multiple discrete elements and automatically spreads
them to different servers (Ohlhorst, 2013, p. 8). The input files are sharded,
that is to say they are not divided according to the existing structure of the
files, but arbitrarily so, for example, text files are split according to byte
boundaries (Tachmazidis et al., 2012). The second reduce step draws on input
from the scattered map nodes, and joins the map results back into a final master
calculation (Ohlhorst, 2013, p. 8). Thus, for example, in the application of
MapReduce to human genome analysis, the first map step would conduct
analysis at the level of the genome, such as genotyping, with the output fed to the
reduce step, where calculations are made across the aggregate data on multiple
points of the genome (McKenna, 2010). In domains such as human genetics,
meteorological data, security intelligence and business intelligence, MapReduce
is thought to supply processing of a vast amount of data in parallel on large
clusters of machines in a fault-tolerant manner (Gupta et al., 2012, p. 49).
In simple terms, MapReduce is significant because it changes the nature
of what can be analysed across multiple data formats and databases, across a
distributed data landscape:
Every call, tweet, e-mail, download, or purchase generates valuable data.
Companies and governments are increasingly relying on Hadoop to unlock the
hidden value of this rapidly expanding data [] Sensor output, videos, log files,
location data, genomics, behavioural data are just a few of the data sources
driving Hadoop use. (MapR for Apache Hadoop, 2011)
MapReduce is thought to unlock hidden value because it makes it possible

to analyse exponentially increasing volumes of diverse data for patterns and
clusters that are not necessarily determined in advance. The partitioning and
351
analysis of data using Hadoop software deploys algorithms in a process

described as knowledge discovery. In contrast to the deductive production of
knowledge from a priori queries or hypotheses, in this case the data analytics
use inductive and abductive steps to identify previously unknown patterns in a
large volume of data (Dunham, 2002). The significance here is that knowledge
discovery increasingly does not begin with a set of search queries against which
the data will be run. Instead, the process inductively generates queries, such
that the analytics are said to let the data speak (Ohlhorst, 2013).
How do analytics techniques like Hadoop MapReduce transform the nature
of what can be rendered perceptible and analysed? In contrast with the more
strictly statistical structured query language (SQL), designed in the 1970s
for use with relational databases, advanced analytics work with the uncertainties of possible links and connections. As the author of the worlds most highly
cited computer science paper on data mining software, Rakesh Agrawal,
explains, past forms of data analytics used a statistical notion of what was
interesting, such that the prevailing mode of decision making was that
somebody would make a hypothesis, test if it was correct, and repeat the
process (Agrawal & Winslett, 2005, p. 5). With the advent of large databases,
distributed computing and extensive unstructured data sources, however,
the decision making process changed, and a series of algorithms would
generate all rules, and then debate which of them was valuable (2005, p. 8).
While in statistical forms of large-volume data the object of interest emerges
from the testing of probabilistic assumptions or queries, with the output being
a subset of a database, in knowledge discovery the matter of interest is iterative
and emergent, with the output consisting of previously unknown patterns and
relationships.
In this way, the work of the analytics is to discover, aggregate and interpret
rules for items within a subset of data. It is the rules generated by the rules within
the analytics that will determine what is of possible value. Let us imagine, for
the purposes of illustration, a rudimentary knowledge discovery process in
which it is inferred that transactions in a database which contain item x also
contain item y. So, it may be that 24 per cent of customers who purchase a
novel x from an online retailer will also purchase music download y, and that
5 per cent of the total transactions in a given period contain both purchases. In
this example, 24 per cent expresses the confidence score of the rule xy, and 5 per
cent expresses the support for the rule. The problem of knowledge discovery
across a large volume of transactions, then, is to run the analytics in order to
identify all rules that satisfy some predetermined level of support and
confidence (Zhang & Wu, 2001). What matters is thus not the intrinsic value
or content of x or y, but how these data items can be associated together and
what can be known about their relations with a wider set of bulk data (Agrawal
et al., 1993). Importantly, the use of association rules for large transaction data
sets is not unproblematic, as their analysis can reveal hundreds of thousands of
rules at reasonable levels of support or confidence, with many of them being
redundant or obvious and therefore not interesting (Klemettinen et al., 1994).
352
Economy and Society
A process of pruning is therefore required in order to reduce the things

of interest to the strongest relationships in the data (Raeder & Chawla, 2011,
p. 100).
If the advent of what has come to be known as big data analytics is changing
the nature of what questions or queries can be asked, or of what can be
calculated, or the nature of analysis itself, then what is the significance of this?
Does it matter to political and social life, to how we govern and are governed?
The interception and analysis of terabytes of unstructured data for security
purposes has attracted a great deal of political attention in the wake of Edward
Snowdens revelations about the NSAs and GCHQs PRISM and TEMPORA
programmes (Greenwald, 2014; Harding, 2014).11 The locus of the debate,
though, has resided primarily with the question of mass surveillance and the
collection and storage of personal communications and transactions data in
bulk (LIBE, 2013). There has been scant attention paid to the processes of data
partitioning and reassembly, and what these might mean for the relationship
between a mass or bulk volume of data and an object or person of interest.
Somewhat hidden and unremarked upon in one of the leaked PRISM slides
on collection and dataflow is the scissors process, which sorts data types
(http://www.washingtonpost.com/wp-srv/special/politics/prism-collectiondocuments/m/). Though we can observe very little of the classified processes
at work in the partitioning and sorting of the data in PRISM, there are
materials in the public domain that make it possible to understand the
processes of cutting and stitching involved in analytics processes such as
scissors.
In June 2013, the US Government Accountability Office (GAO) published
a decision on the technology corporation IBMs challenge to the CIAs
award of a contract for data analytics services to Amazon Web Services. IBM
had challenged the award of the contract to Amazon on the grounds that the
CIAs evaluation team did not fairly evaluate the technical and financial aspects
of the bids. At issue was the two companies materially different interpretations of the scenario requirements (GAO, 2013, p. 4). While Amazon priced
their bid on the basis of continual 24/7 analysis of a volume of 100 terabytes
of data, IBM had envisaged analysis of a series of batches of 100 terabytes
of data. In preparing their bids, the software design teams were required to
address, and to price, the CIAs scenario for data analysis:
This scenario centers around providing and hosting an environment for
applications which process vast amounts of information in parallel on large
clusters (1000s of nodes) in a fault tolerant manner using MapReduce. The solution
to this scenario should automatically compute for the segmentation and parallel
processing of datasets via the MapReduce framework [] Assume a cluster
large enough to process 100 TB of raw input data. Assume 6 reads/second and
2 writes/second. Assume 100% duty cycle on all machines. (GAO, 2013, p. 5,
emphasis added)
353
When IBM queried the temporality of the 100 terabytes and 100 per cent duty
cycle, they specifically asked how many data analysts would make simultaneous
queries of the data in the scenario. The CIA responded to the query by
appealing to the existing practices of data analytics in the commercial sphere,
inviting the bidders to bring to the state the techniques already thought to
be best practice in economy and commerce: The contractor should propose
commercial best practices derived from their commercially available solutions
to provide data analytics via the MapReduce software framework to concurrent
users from multiple organizations (GAO, 2013, p. 7). Here the divergent
responses of IBM and Amazon to the scenario reveal rather more than two
competing interpretations of the requirements. They afford a glimpse of how
the data analytics in processes such as scissors sort large volumes of data, and
the proximity of security applications such as PRISM and TEMPORA to the
commercial data analytics used every day to tell us which book we might like
to buy next. Amazons established commercial practice of analysing clickstream
data on its customers in close to real time and on a continuous cycle, it seems,
better met the CIAs requirement for analytics to deal with large volumes
of unstructured Internet data to be queried by multiple concurrent users, from
border and immigration control to counter-terrorism officers.
The capacity to integrate data analytics across multiple analysts, and to map
and reduce across multiple nodes, exhibited here by Amazon, contrasts with
IBMs extraction of batches of data for analysis. One response to the CIAs
scenario appears to sustain a somewhat conventional social science approach
to sampling, and a particular relation between the subset and the whole of big
data. In the other response one can see the ceaseless stream of ingestion,
partitioning and reassembly that affords novel iterative approaches to sample
and whole, where, in effect, people and objects continually cross back and forth
across the sample and the whole. The distributed analysis of data streams,
as David Berry writes, sustains some form of relationship with the flow of data
that doesnt halt the flow, but rather allows the user to step into and out of a
number of different streams in an intuitive way (2011, p. 143). In Amazons
MapReduce framework for the CIA, it is the identification of patterns of note
across different data streams that gives rise to a threshold at which a target or
person of interest is identified.
The mobile thresholds of support and confidence for an association rule the
very key to setting the gauge for the analytics have become highly significant
political boundaries for our times. The threshold is the moment when the
strongest relationships are identified, the moment when someone or something
of interest becomes perceptible. The moving of the threshold changes who or
what is surfaced from the data and brought to attention. In the historical origins
of data analytics this threshold was commonly defined in terms of a frequent set,
where the co-occurrence of retail consumer items in patterns of purchases
met a predetermined level of support and confidence. Co-occurrence in itself
is not always a matter of interest, for example, milk co-occurring with bread
in basket data would have high levels of support and confidence, but would
354
Economy and Society
not constitute an object of interest. Where similar MapReduce processes are

used to set the threshold of risk for border controls, or the threshold for a nexus
to terrorism (de Goede, 2012), the threshold of support and confidence
becomes a border in itself, where the co-occurrence of particular data elements
will give rise to a person or object of interest. Understood thus, the little analytics
are instrumental in what is called target discovery, the defining of a political
threshold of perceptibility where a person of interest comes into view.
The partitioning and assembly processes not only structure something of
the threshold of political visibility, but also redefine the lines of sovereign
authority. In the design of software for sovereign information sharing (SIS)
for the US Transportation Security Administration (TSA), for example, the
analytics are said to enable computation across autonomous data sources in
such a way that no information other than the intersection results is revealed
(Agrawal et al., 2005, para 2.1). Such insights into the work of the analytics
are critical to the contemporary form of governing life. Though the classified
nature of programmes such as PRISM and TEMPORA makes it impossible
definitively to identify whether sovereign information sharing is among the
techniques used, the computer scientists reveal clearly how their analytics
make it possible to share the subset data on persons who cross a threshold
of interest, while annexing the big data sample from which it was drawn. In the
TSA example available in the public domain, the airlines encrypt their PNR
data, and the security authorities encrypt their watch-lists, with the analytics
running the intersection results for patterns, associations and matches. The
TSA agrees that the use of the intersection results will be limited to the
purpose of identifying suspects, write the computer scientists, but it will
store all the metadata (Agrawal, et al., 2005, para 6.5). In the TSAs sovereign
information sharing system, the analytics become part of the condition of
possibility for sovereign power. Not only does SIS appear to make possible
sovereign decisions on who or what poses a risk to US transportation security,
but it also establishes the threshold at which data on persons of interest are
pulled to the perceptible surface in the map process and extradited to the
reduce step, where a calculation is made about them.
The partitioning work of contemporary analytics such as MapReduce,
then, is political in the sense that it defines the threshold of perceptibility.
This is a threshold where, as Deleuze reminds us with Bergson, differences in
kind can no longer appear, and where science no longer presents anything
but differences of degree, of position, of dimension, of proportion (1991, p. 34).
Analytics are technologies of degree par excellence. They subdivide a heterogeneous data set such that no original whole could ever meaningfully be
reassembled. The partitioning has significant implications for any critical
response to the pervasive use of programmes, such as Hadoop at our
borders, in our consumer databases, inside the risk calculations of banks and so
on. For the work of the analytics is not at all challenged by the demand that the
context of data be respected in its analysis (Nissenbaum, 2010), nor that one has
the right to be forgotten or deleted in a digital age (Mayer-Schnberger, 2009).
355
The processes of partitioning and analysis precisely do not require a context, nor
do they need individuals who can be remembered. The critical demand for
a contextual limit to the analysis of life data, or a deletion of the digital subject,
gains little purchase in a world where attributes are extracted from their qualities
and afforded numeric values. The analytics that partition big data make it
possible to forget the person and the context, but to remember the position,
the distance or proximity of association.
Memory: near real-time analytics

Presenting their analytics for stream-based event processing, analytics
providers TIBCO introduce their Spotfire solution:
TIBCO helps you understand the past so that you can better anticipate the
future. We call it the two second advantage What good is it to know that
youve lost your customer after the customer has left your premises? What good
is it to know that there is a power outage after a city is in darkness? Or that
fraud has occurred after the money has left the bank? This is the power of the
two second advantage. This is the importance of being able to both understand
the past and anticipate the future.12
TIBCOs development of analytics that promise to identify emergent trends

and to enable anticipatory action to uncover opportunities nobody else can
see signals something of the temporality of contemporary analytics. TIBCO
Spotfire promises to turn data into actionable insights with dashboards,
apps and analytics, so that data on unfolding events can be used to enable fast
and strategic near-real-time decisions. The software for such stream-based
analysis identifies links between events coming from multiple data sources for
example, Twitter trends, smartphone transactional data, Facebook likes
stitching together the data signals to anticipate near futures, what they call
the two second advantage. Such methods are becoming ubiquitous in the
commercial world, for identifying people with a propensity to churn and transfer
their custom to a new company, and in the security domain, where radicalization
or attack planning is thought to be identifiable at the intersection of multiple data
events. Spotfire also signals a broader move to simplify the interface between
the analytics and the user, such that she doesnt need to understand the R-code
running in the background; the software automatically chooses the most
appropriate forecasting algorithm; and the analyst can seek help from what does
it mean pop-ups, without having to understand how the forecast was created.13
The relationship between past data, decisions made in the present and actions
taken on a future that is seconds away, then, is significantly reconfigured by
advanced analytics. What form of machinic memory do the analytics access in
order to anticipate the future? How do simplified drag and drop interfaces, which

hide the complexity from the analyst, change the orientation to decision and
action?
The claim of the analytics capacities to analyse data in near real time, or with
a two second advantage, is based upon data stream processing. In effect,
the stream of unstructured data is conceived as a continuous flow whose
speed exceeds conventional perceptions of an event. For example, when a 5.8
magnitude earthquake struck the US state of Virginia in August 2011, it was
suggested that the first Twitter messages reached New York in advance of the
first measurable shock waves (Emerson et al., 2012). Similarly, on the night of
the raid on Osama Bin Ladens compound in Abbottabad in 2011, IT consultant
Sohabib Athor began a string of Twitter messages with Helicopter hovering
over Abbottabad at 1am, is a rare event; 82.68 million retweets and 21 hours
later, Athor tweeted Uh oh, now Im the guy who live blogged the Osama
raid without knowing it (Emerson et al., 2012). The Topsy Labs analytics
identified the virality of the message exposure within the social web, demonstrating how open-source stream data can be analysed to reveal otherwise
unknown events in their unfolding.
Advanced event stream analytics, such as those deployed in high-frequency
financial trading (MacKenzie, 2011), and in applications like TIBCO Spotfire, imply a significant transformation of the relationship between past,
present and future, a transformation that is not fully captured by the idea of
a real time. In his commentaries on the relation between past and present,
Henri Bergson signals the specific temporalities of memory. You define the
present in an arbitrary manner as that which is, he writes, whereas the present is
simply what is being made, and nothing is less than the present moment, if you
understand by that the indivisible limit which divides the past from the future
(1912, p. 193). If, as Bergson understands it, perception of the present is more
precisely located in the immediate past practically we perceive only the past
then could it be that any claim to a real-time present can only hope to engage
the invisible progress of the past gnawing into the future (1912, p. 194)? Or,
does the machinic memory of advanced analytics find novel ways to contain and
access images of the past, such that action in the present is a possibility? But,
if the brain cannot serve such a purpose, asks Bergson, in what warehouse
shall we store the accumulated images? (1912, p. 192). Can we conceive of
analytics such as TIBCO Spotfire as supplying an infinite data warehouse of
images of the past, which may be retrieved and replayed at any time?
Of course, in many ways the accessing of memory depicted in Bergsons
account of conscious perception is a process that is impossible for algorithm
to replicate. In spite of claims to transcend the limits of human memory and
expand capacity for action, big data analytics have limited capacity to incorporate
the durational time of consciousness and experience. What matters in the
algorithmic recalling of a series of past events as data is not strictly their temporal
character, but their spatial distances data-point to data-point, their links, asso
ciations and correlations, one to another. As Katherine Hayles has expressed
the different temporalities at work, the distinction between measured time and
357
time as temporal process can be envisioned as the difference between exterior

spatialization and interior experience (2012, p. 112). The system of memory
appropriate to describe real-time analytics, then, is one of measured time in
which the temporalities or durations that dwell within the life that yields the
data-points are entirely lost. Gilles Deleuze depicts Bergsons duration as
the time in which the present endures: the present that endures divides
at each instant into two directions, one oriented and dilated toward the past,
the other contracted, contracting toward the future (1991, p. 52). It is only in
duration, one might say with Bergson and Deleuze, that the capacity to perceive
past, present and future as qualitatively different things is possible. By contrast,
the data analytics that promise a two second advantage can conceive only of a
spatial point sliced through time, a point where action can be taken.
Following the process of memory: Featurespace

What kind of process of memory is at work in so-called real-time analytics? The
past of the analytics is a preselected collection of actual pasts and, when it is
used, only discrete elements are recalled. Let us illustrate this process at work
in the Adaptive Real-Time Individual Change Identification (ARIC) engine
produced by Featurespace. Described as the only adaptive behavioural analytics
software in existence, ARIC is used in the detection of casino and credit
card fraud, customer churn, marketing opportunities and security threats in
over 60 countries, and in processing over 20 million transactions per day
(Featurespace, 2013). The ARIC engine analyses thousands of disparate data
sources, including SMS data from mobile phones and e-mail metadata,
identifying signals, or small but interesting changes in the patterns of data.
The signals are described as symptoms of human behaviour that can be
converted, via processes of parsing, stemming and partitioning, into features.
The features are observed in their multiple relations to other people, things
and groups with correlating features, such that predictions for individuals are
made based on a propensity to act: for example, to churn, to commit fraud, or to
purchase a product (Featurespace, 2013). Changes in behaviour that deviate
from individual and context profiles are identified, and the results are fed back
into ARIC in a process of self-learning and continuous updating of profiles
in real time (Figure 1). In effect, the analytics in ARIC combine conventional
Bayesian conditional probabilities with a capacity for machine learning on the
basis of small modifications in observed behaviours.
The form of memory at work in Featurespaces analytics engine is an
iterative movement back and forth across past features, present recursive
recalibrations of the rules of the analytics (the machine learning) and future
projections of human propensities to act. This process observes no qualitative
distinctions, not only in the vast and heterogeneous array of data that is
ingested from different architectures, but also in the transformations that take
place in the present as threshold between past and future. As one analyst
358
Economy and Society
described the process to us in interview, we just keep iterating until the results
are satisfying. The processes of memory and iteration in Featurespace, or
TIBCO Spotfire or indeed in the analytics used by GCHQ and the NSA to
build a pattern of life (The New York Times, 2013) are far removed from what
Bergson termed attention to life (1912, p. 63). Where attention to life bears
witness to the adaptation of the past to the present, the utilization of the past
in terms of the present (Deleuze, 1991, p. 70), the features or patterns
of life sought by major supermarket chains and national security agencies know
no limitations, no indeterminacies, nothing that is not available to action.
Where the durational time of consciousness confronts the indeterminate
future by shedding some light gathered from selected past states, combining
with present states, it does so in the knowledge that the rest remains in the
dark (Bergson, 1912, p. 194). Amid their claims to predict human propensities,
by contrast, analytics engines, such as ARIC and Spotfire, confront an
indeterminate future in order precisely to leave nothing in the dark and nothing
undetermined. Though the analytics share with consciousness the selection
Figure 1 Featurespaces ARIC engine.

Source: http://www.featurespace.co.uk/technology/aric.
359
of some discrete past events, there the commonality ends. For the analytics take
the light of some past states and project it forward as though there could be no
dark corners remaining all propensities will be known, all future acts
anticipated. When analytics like ARIC are being used to monitor social media
in the Arab Spring uprisings, or to monitor Twitter in the predictive policing
of urban protest, it is of great significance that all memory of every past infraction
is thought to be retrievable, all futures foreseeable. Indeed, the event in event
stream analysis is annulled as such, along with the real in real-time analytics.
For nothing new or eventful can emerge, such is the machine time of the iterative
replay of the past state, modified for recent deviations and gnawing into the
uncertain future.
Conclusions: the loss of a fallible world

Asked to explain the distinction between the advanced analytics deployed in
the commercial spheres of retail and banking and those used for state-oriented
security purposes, an analyst reflects on the problem. We license software for
companies to detect fraud and we license it to governments for counterterrorism, he says, but we dont know what they do with it because that
would be classified. I assume, he concludes, they are using the most
advanced analytics (interview, 1 October 2013). What is perhaps most notable
here is that, at least from the standpoint of the software designers and analysts,
the overriding process is one of ingestion, partitioning, retrieving and
analysing large volumes of data in order to identify people, objects or patterns
of interest. In terms of computer science and mathematics, it scarcely matters
what those data are, or from whence they came they could be financial
transactions, supermarket purchases, or national security watch-lists what
matters is the capacity to identify something of interest amid the mass, and the
capacity to action that analysis.
It is precisely because of the growing ubiquity of big data analytics across
diverse spheres of life that social science should pause in the rush to exploit the
value of big data and attend carefully to what form of analysis is engaged by
advanced analytics. The apparent big in big data departs significantly from the
avalanche Ian Hacking (1990) observes in nineteenth-century social statistics.
Hacking describes in detail how people of interest emerge from the volume
of structured data on economy and society in the nineteenth century. From
the average man or lhomme typique of Adolphe Quetelet to the biometric
composites of criminality of Francis Galton, the statisticians identified their
object of interest from probabilistic calculation (see also Daston, 1995).
In short, they began their inquiries with a statistical notion of what was
interesting this crime rate, or these rates of infant mortality, or this
probability of suicide.
The twenty-first-century rise of big data marks a significant break with
conventional statistical notions of what is of interest. The vast expansion of

unstructured digital data, much of it considered to be open source, has been
closely intertwined with the development of advanced analytical algorithms to
make some sense of that data. And so, amid the cacophony of noise around the
big in big data, we urge careful attentiveness to the work of the little analytics.
Rather as the growth of statistical probabilistic methods made data on murder,
health, employment or war perceptible and amenable to analysis, so contemporary analytics are instruments of perception without which the extensity of big
data would not be perceptible at all. As art historian Jonathan Crary explains in
his compelling account of historical changes in the nature of perception,
perception transformed alongside new technological forms of spectacle, display,
projection (1999, p. 2). As instruments of perception, the technologies of
analytics not only focus human attentiveness on particular persons or things
of interest, but they also annul and discard as redundant much of the material
context from which these persons and things emerged. In this sense, the little
analytics are one element of a contemporary experience that requires that
we effectively cancel out or exclude from our consciousness much of our
immediate environment (Crary, 1999, p. 16). And yet, they are not merely
the twenty-first-century manifestation of Benjamins mechanical reproduction
or Lorraine Daston and Peter Galisons (2007) mechanical objectivity. As
we have proposed, the invention of analytics engines has transformed the nature
of analysis and, with it, the nature of what and how life can be rendered
governable.
By way of conclusion, we draw out three implications of a focus on the work
of analytics, as they transform the governing of economic, social and political
life. First, the advent of advanced analytics ushers in a specific and novel
epistemology of population. At first glance, the formulation n=all appears to
render the whole of population as the sample all data on all of lifes
transactions are, at least in theory, available to analysis. But the population as
the sample in contemporary analytics processes does not imagine the
population as a curve of normality or a Gaussian bell curve of plotted
attributes (Foucault, 2007, p. 63). Once we give up on the magicians wand
and follow the process, as advised by Bergson, we can see how the object of
interest becomes detached from the population as such. As the mathematicians
and analysts tell us, advanced analytics work not merely with a statistical
notion of what is interesting, but also via an inductive process of knowledge
discovery, in which the process generates the rules. Thus, in the chain
analysis of air transportation security algorithms, the person of interest
emerges from the links of activities funded; member of; listed; acquainted
with; travelled to; countries visited; and geopolitical events (GAO, 2007).
The population in the n=all formulation is a curious entity a series of
possible chains of association, in which there are no standard bell curves of
normality, and from which anyone or anything could become a matter of
interest or concern.
Second, the advent of data analytics brings significant ontological implications
for thought and practice. The processes of ingestion, partitioning and machinic
361
memory reduce heterogeneous forms of life and data to homogenous spaces

of calculation. Algorithmic technologies, such as those we have described, tend
to reduce differences in kind to differences in degree, as Bergson and Deleuze
might say, or to distances data-point to data-point. From these processes of
reduction and flattening it is thought that different kinds of life stories
patterns of life emerge, and that interventions and decisions can be made on
their basis. The affective world of an epilepsy sufferer, a sub-prime borrower,
a border crosser, a terrorist or a criminal is thought to be excavable from the
seams and joins of multiple data sources. But, what kinds of stories can be
told with analytics? What happens to the things that cannot be spoken, or that
which is not fully accessible to us even of ourselves? In current debates on
protecting people from the worst vicissitudes of data mining and algorithmic
decision, the emphasis is placed overwhelmingly on the restriction of
processes to specified people of interest and specific queries defined in
advance. As we have argued, this critical framing entirely misunderstands the
work of the analytics. The stories of the mass or bulk data are precisely the
means by which the queries are generated the analytics require the mass data
in order to decide on what or who is interesting, and this can only ever be
retroactive.
Finally, the rise of analytics has important consequences for the form of
contemporary politics. In effect, the analytical processes of ingestion, partitioning and reassembly and memory we have described make a particular claim in the
world, they say n=all, this is the world, here it is, all data is rendered tractable.
They carve out and convert radical heterogeneity into flat difference of degree,
such that it appears as though everything is calculable, everything about
the uncertain future is nonetheless decidable. With its applications which
aim only at the convenience of existence, writes Bergson, science gives us
the promise of well-being (1965, p. 129). Todays little analytics promise a
convenience of existence via the detection of all human propensities. What
need, then, for politics? If politics expresses the fallibility of our world, the
impossibility of resolution of all matters economic, social, ethical, then it exists
because not everything is reducible and resolvable. Politics in our times confronts
a ubiquitous analytics that imagines an infallible world where even the most
turbulent of situations can be rendered tractable. Perhaps the most striking and
troubling matter of scale in big data is the recognition of an extensity that always
exceeds the capacity of human knowledge to collect and apprehend it. The
promise and allure of the little analytics is to see those things that would otherwise
be invisible, to perceive the imperceptible and to feed the insights to those who
would action them. Confronted with this claim to reduce the fallibility of
governing and decision, to do so in a manner that is fault tolerant,
a political response must be mindful of the perils of the magicians wand, and
point to the material, contingent and fallible processes that make this claim
possible.

Acknowledgements
An early version of this paper was presented at the Calculative Devices in the
Digital Age conference, Durham University, November 2013. Our thanks go
to all participants at that event for the stimulating discussion that took place.
The three anonymous reviewers for Economy & Society and Managing Editor
Fran Tonkiss have been extraordinarily helpful and generous with their
comments and insights; thank you. The research field-work from which the
empirical elements of the paper are drawn, conducted during 2013, involved
observations of data analytics industry and governmental events, and interviews with software engineers, analytics consultants and data scientists. We are
grateful to everybody who has so generously given time to the project.
Disclosure statement
No potential conflict of interest was reported by the authors.
Funding
This work was supported by the RCUK Global Uncertainties Fellowship Securing
against Future Events: Pre-emption, Protocols and Publics [Grant number ES/
K000276/1].
Notes
1 In response to the UK governments announcement of the second phase of funding
for Big Data centres, Chief Executive of the ESRC, Professor Paul Boyle, welcomed
the sheer volume of data that is now being created, a significant resource that can
shape our knowledge of society and help us prepare and evaluate better government
policies in the future (ESRC, 2014).
2 It is not our purpose here to map a linear history of practices of data collection and
analysis. Rather, we juxtapose two moments when a specific set of claims are made
regarding the scale and scope of social data and its effects on the governing of societies.
3 Bergsons reflections on perception in science are present throughout his body of
work. Of particular significance here is his insistence on the shared categories of
thought and sensing across science and prosaic perception, so that ordinary knowledge
is forced, like scientific knowledge, to take things in a time broken up into particles,
pulverized so to speak, where an instant which does not endure follows another without
duration (1965, p. 120).
4 Indeed, by 1930, Bergson himself appreciated the growing capacity of modern
mathematics and physics to capture something of perpetual and indivisible change, to
follow the growth of magnitudes and to seize movement from within (1965, p. 211).
5 The earliest use of the concept of ingestion for analysis of data in multiple formats
can be found in papers from IBMs research on smart surveillance and web architecture
(Chiao-Fe, 2005; Gruhl et al., 2004). The use of a vocabulary of ingestion coincides
363
with an expansion of analysable samples of digital data, such that it is said that n=all, or
the sample is equal to everything.
6 The concept of index is used here in the sense proposed by Deleuze and Guattari to
denote the capacity to designate the state of things, territorially locatable in time and
space (1987, p. 124). Understood thus, for example, extraction algorithms are required
in order territorially to index unstructured objects, as in the use of biometric templates
derived from Facebook. It is the extracted template that makes the object searchable in
time and space.
7 The case is derived from field-work conducted in London in 2013. For further
examples and detailed descriptions of text mining and sentiment analysis, see Bello et al.
(2013); Zhao et al. (2013); and Anjaria and Gudetti (2014).
8 Hayles defines the concept of technogenesis as the idea that humans and technics
have coevolved together, such that our very capacity for thought and action is bound
up with epigenetic changes catalysed by exposure to and engagement with digital
media (2012, pp. 1012). The idea is present also in Walter Benjamins famous essay
on art in the age of mechanical reproduction, where he notes that the mode of human
sense perception changes with humanitys entire mode of existence (1999, p. 216).
9 Retrieved from http://www.theguardian.com/profile/laura-poitras. See also Harding (2014, pp. 110, 204).
10 Though the focus of this essay is not on the interface between data architectures
and software, the flattening of differences at this interface is significant. See Galloway
(2012); Berry (2011).
11 Despite substantial interest in the automated analysis of large data sets for security
purposes in the wake of Edward Snowdens disclosures, the use of algorithmic
techniques to analyse Passenger Name Record (PNR) and SWIFT financial data has
been known and documented for some time (Amoore, 2013; de Goede, 2012).
12 Insights drawn from observations at TIBCO Spotfire event, London, 13
June 2013.
13 Insights drawn from observations at TIBCO Spotfire event, London, 13 June
2013, and SAS Analytics How to workshops, 19 June 2013.
References
Accenture. (2013). Accenture analytics in
action. Retrieved from http://www.
accenture.com/sitecollectiondocuments/
pdf/accenture-analytics-in-actionsurvey.pdf
Agrawal, R., Asonov, D., Baliga, P.,
Liang, L., Porst, B. & Srikat, R. (2005).
A reusable platform for building sovereign
information sharing applications. SIGMOD
Proceedings. Retrieved March 25, 2015,
from http://www.rsrikant.com/papers/
divo04.pdf
Agrawal, R., Imielinski, T. & Swami,
A. (1993). Mining association rules between
sets of items in large databases. SIGMOD

Proceedings (pp. 207217).
Agrawal, R. & Winslett, M. (2005). An
interview with Rakesh Agrawal. SIGMOD.
Retrieved March 24, 2015, from http://
www.sigmod.org/publications/interview/
pdf/D15.rakesh-final-final.pdf
Amoore, L. (2011). Data derivatives: On
the emergence of a risk security calculus
for our times. Theory, Culture & Society,
28(6), 2443.
Amoore, L. (2013). The politics of possibility: Risk and security beyond probability.
Durham, NC: Duke University Press.
364
Economy and Society
Anjaria, M. & Gudetti, R. M. R. (2014).

A novel sentiment analysis of social networks using supervised learning. Social
Network Analysis and Mining, 4(3),
181193.
Ansell Pearson, K. & Mullarkey, J.
(Eds.). (2002). Henri Bergson: Key writings. London: Bloomsbury.
Bello, G., et al. (2013). Extracting collective trends from Twitter using socialbased data mining. In C. Badica, N. T.
Nquyen & M. Berezovan (Eds.), ICCCI
2013. LNAI 8083 (pp. 622630).
Benjamin, W. (1999). Illuminations.
London: Random House.
Bergson, H. (1912). Matter and memory.
Mineola, NY: Dover Philosophical
Classics.
Bergson, H. (1965). The creative mind.
(M. Andison, Trans.). Totowa, NJ: Littlefield, Adams & Co.
Berry, D. M. (2011). The philosophy of
software: Code and mediation in the digital
age. Basingstoke: Palgrave Macmillan.
Berry, D. M. (2014). Critical theory and
the digital. New York, NY: Bloomsbury.
Boyd, D. & Crawford, K. (2012).
Critical questions for big data. Information, Communication & Society, 15(5),
662679.
Chiao-Fe, S. (2005). IBM smart surveillance system: an open framework for event
based surveillance. IEEE Proceedings on
advanced video and signal surveillance
(pp. 318373).
Connolly, W. (2011). A world of becoming. Durham, NC: Duke University Press.
Crary, J. (1999). Suspensions of perception:
Attention, spectacle, and modern culture.
Cambridge, MA: MIT Press.
Daston, L. (1995). Classical probability in
the enlightenment. Princeton, NJ: Princeton University Press.
Daston, L. & Galison, P. (2007). Objectivity. New York, NY: Zone Books.
Dean, J. & Ghemawat, S. (2004).
MapReduce: Simplified data processing
on large clusters. Proceedings of the
OSDI. Retrieved March 25, 2015, from
http://static.googleusercontent.com/
media/research.google.com/en//archive/
mapreduce-osdi04.pdf
Deleuze, G. (1991). Bergsonism. New

York, NY: Zone Books.
Deleuze, G. & Guattari, F. (1987).
A thousand plateaus: Capitalism and
schizophrenia. Minneapolis, MN: University of Minnesota Press.
Dunham, M. H. (2002). Data mining:
Introductory and advanced topics. Upper
Saddle River, NJ: Prentice Hall.
Emerson, T., Ghosh, R. & Smith, E.
(2012). Practical Text Mining and Statistical Analysis for Non-Structured Text
Data Applications. London: Elsevier.
ESRC (Economic and Social Research
Council). (2014). ESRC big data network:
Phase 2: Business and local government data
research centres. Retrieved April 27, 2014,
from http://www.esrc.ac.uk/research/
major-investments/Big-Data/BDNPhase2.aspx
Featurespace. (2013). Our technology.
Retrieved November 10, 2013, from
http://www.featurespace.co.uk/technology/our-technology/
Foucault, M. (2007). Security, territory,
population: Lectures at the Collge de
France, 197778. Basingstoke: Palgrave
Macmillan.
Galloway, A. R. (2012). The interface
effect. Cambridge: Polity.
GAO (Government Accountability
Office). (2007). Data mining: early attention
to privacy in developing a key DHS program
could reduce risks, GAO-07-293. Washington, DC: Government Accountability Office.
GAO (Government Accountability
Office). (2013). Decision: IBM v US
Federal, B-407073. Washington, DC:
Government Accountability Office.
de Goede, M. (2012). Speculative security: The politics of pursuing terrorist monies.
Minneapolis, MN: University of Minnesota Press.
Greenwald, G. (2014). No place to hide:
Edward Snowden, the NSA and the surveillance state. London: Penguin.
Gruhl, D., Chavet, D. & Gibson, D.
(2004). How to build a WebFountain.
Utility Computing, 43(1), 6470.
Gupta, R., Gupta, H. & Mohania, M.
(2012). Cloud computing and big data
analytics. Computer Science, 7678, 4261.

Hacking, I. (1982). Biopower and the
avalanche of printed numbers. Humanities
in Society, 5(34), 279295.
Hacking, I. (1986). Making up people. In
T. Heller (Ed.), Reconstructing individualism (pp. 222236). Stanford, CA: Stanford
University Press.
Hacking, I. (1990). The taming of chance.
Cambridge: Cambridge University Press.
Harding, L. (2014). The Snowden files.
London: Faber and Faber.
Hayles, N. K. (2005). My mother was a
computer: Digital subjects and literary
texts. Chicago, IL: Chicago University
Press.
Hayles, N. K. (2012). How we think:
Digital media and contemporary technogenesis. Chicago, IL: Chicago University
Press.
Hildebrandt, M. (2013). Slaves to big
data. Or are we? Keynote at the 9th
Annual Conference on Internet, Law &
Politics (IDP, 2013, Barcelona). Retrieved
January 14, 2014, from http://works.
bepress.com/mireille_hildebrandt/52/
IBM. (2012). IBM content analytics:
Rapid insight for crime investigation.
Retrieved November 19, 2013, from
http://public.dhe.ibm.com/common/ssi/
ecm/en/zzs03073usen/ZZS03073USEN.
PDF
Inmon, W. H. & Nasaevich, A. (2007).
Tapping into unstructured data: Integrating
unstructured data and textual analytics into
business intelligence. Upper Saddle River,
NJ: Prentice Hall.
Intelligence and Security Committee.
(2014). Transcript of evidence given by
Theresa May MP, Home Secretary.
Retrieved October 25, 2014, from http://
isc.independent.gov.uk/public-evidence
Intelligence and Security Committee.
(2015). Privacy and security: A modern and
transparent legal framework. London:
HMSO.
Kitchin, R. (2014). The data revolution:
Big data, open data, data infrastructures and
their consequences. London: Sage.
Klemettinen, M., Mannila, H.,
Ronkainen, P. & Toivonen, H. (1994).
Finding interesting rules from large data
sets of discovered association rules.
365
Proceedings of the 3rd International

Conference on Information and Knowledge Management, 684699.
Kroes, N. (2013). Speech: The big data
revolution (European Commission Speech/
13/261). Retrieved April 27, 2014, from
http://europa.eu/rapid/press-release_
SPEECH-13-261_en.htm
Latour, B. (1988). The pasteurization of
France. Cambridge, MA: Harvard University Press.
LIBE (2013). Draft report on the US
NSA surveillance programme. European
Parliament Committee on Civil Liberties,
Justice and Home Affairs: 2188.
MacKenzie, D. (2011). How to make
money in microseconds. London Review of
Books, 33(10), 1618.
MapR for Apache Hadoop (2011).
White Paper. Retrieved January 26, 2014,
from https://www.mapr.com/sites/
default/files/mapr_dist_white_paper.pdf
Marron, D. (2007). Lending by numbers: Credit scoring and the constitution
of risk within American consumer credit.
Economy and Society, 36(1), 103133.
Mayer-Schnberger, V. (2009). Delete:
The virtue of forgetting in the digital age.
Princeton, NJ: Princeton University Press.
Mayer-Schnberger, V. & Cukier, K.
(2013). Big data: A revolution that will
change how we live, work and think. New
York, NY: Houghton Mifflin Harcourt.
McKenna, A. (2010). The genome analysis toolkit: A MapReduce framework for
analysing DNA sequencing data. Genome
Research, 20(9), 12971303.
Miller, P. & Rose, N. (1990). Governing
economic life. Economy and Society, 19(1),
131.
Nissenbaum, H. (2010). Privacy in context: Technology, policy, and the integrity of
social life. Palo Alto, CA: Stanford University Press.
OECD (Organisation for Economic
Cooperation and Development).
(2013). The OECD privacy framework.
Retrieved April 27, 2014, from http://
www.oecd.org/sti/ieconomy/oecd_
privacy_framework.pdf
Ohlhorst, F. (2013). Big data analytics.
London: Blackwell Wiley.

Parisi, L. (2013). Contagious architecture: Computation, aesthetics and space.
Cambridge, MA: MIT Press.
Raeder, T. & Chawla, N. V. (2011).
Market basket analysis with networks.
Social Networks Analysis and Mining,
1 (2), 97113.
Rose, N. (2006). The politics of life itself:
Biomedicine, power, and subjectivity in the
twenty-first century. Princeton, NJ: Princeton University Press.
Ruppert, E., Law, J. & Savage, M.
(2013). Reassembling social science methods: The challenge of digital devices.
Theory, Culture & Society, 30(4), 2246.
SAS Analytics (2013). Field-work
observations of SAS knowledge
discovery processes, London.
Tachmazidis, I., Antoniou, G. &
Kotoulas, S. (2012). Scalable nonmono-
tonic reasoning over RDF data using

MapReduce. Proceedings of the Joint
Workshop on Scalable Semantic Web
Systems (pp. 7590) 11th November,
Boston, USA.
The New York Times. (2013, September
28). NSA gathers data on social connections
of US citizens.
Zhang, S. & Wu, X. (2001). Large scale
data mining based on data partitioning.
Applied Artificial Intelligence, 15(2),
129139.
Zhao, L., Yingjie, R., Wang, J., Meng,
L. & Zou, C. (2013). Research on the
opinion mining system for massive social
media data. In G. Zhou et al. (Eds.),
NLPCC 2013, CCIS 400 (pp. 424431)
November 15th19th, Chongqing, China.
Louise Amoore is Professor of Political Geography in the Department of

Geography, Durham University. She is RCUK Global Uncertainties Fellow
(20122015) and author of The politics of possibility: Risk and security beyond
probability (Duke University Press, 2013).
Volha Piotukh is currently Postdoctoral Research Associate at Durham
University, working on the Securing Against Future Events project. She is
the author of Biopolitics, governmentality and humanitarianism: Caring for the
population in Afghanistan and Belarus (Routledge, 2015).

Amoore - Life Beyond Big Data Governing With Little Analytics

Uploaded by

Copyright:

Available Formats

Amoore - Life Beyond Big Data Governing With Little Analytics

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Amoore - Life Beyond Big Data Governing With Little Analytics

Uploaded by

Copyright:

Available Formats

Economy and Society

ISSN: 0308-5147 (Print) 1469-5766 (Online) Journal homepage: http://www.tandfonline.com/loi/reso20

Life beyond big data: governing with little analytics

Published online: 04 Sep 2015.

Submit your article to this journal

Article views: 1922

View related articles

View Crossmark data

Citing articles: 5 View citing articles

Full Terms & Conditions of access and use can be found at

Date: 11 January 2017, At: 09:22

Economy and Society Volume 44 Number 3 August 2015: 341366

Life beyond big data:

342 Economy and Society

In a crowded analytics workshop in London in 2013, a data analyst

Louise Amoore and Volha Piotukh: Life beyond big data

It is precisely such reflection on the situated calculative techniques and processes

Perception and the little analytics

344 Economy and Society

Louise Amoore and Volha Piotukh: Life beyond big data

346 Economy and Society

Following the process of ingestion: text and sentiment analysis

Louise Amoore and Volha Piotukh: Life beyond big data

348 Economy and Society

Louise Amoore and Volha Piotukh: Life beyond big data

350 Economy and Society

Following the process of partitioning: MapReduce

MapReduce is thought to unlock hidden value because it makes it possible

Louise Amoore and Volha Piotukh: Life beyond big data

analysis of data using Hadoop software deploys algorithms in a process

Economy and Society

A process of pruning is therefore required in order to reduce the things

Louise Amoore and Volha Piotukh: Life beyond big data

Economy and Society

not constitute an object of interest. Where similar MapReduce processes are

Louise Amoore and Volha Piotukh: Life beyond big data

Memory: near real-time analytics

TIBCOs development of analytics that promise to identify emergent trends

356 Economy and Society

Louise Amoore and Volha Piotukh: Life beyond big data

time as temporal process can be envisioned as the difference between exterior

Following the process of memory: Featurespace

Economy and Society

Figure 1 Featurespaces ARIC engine.

Louise Amoore and Volha Piotukh: Life beyond big data

Conclusions: the loss of a fallible world

360 Economy and Society

Louise Amoore and Volha Piotukh: Life beyond big data

memory reduce heterogeneous forms of life and data to homogenous spaces

362 Economy and Society

Louise Amoore and Volha Piotukh: Life beyond big data

sets of items in large databases. SIGMOD

Economy and Society

Anjaria, M. & Gudetti, R. M. R. (2014).

Deleuze, G. (1991). Bergsonism. New

Louise Amoore and Volha Piotukh: Life beyond big data

Proceedings of the 3rd International

366 Economy and Society

tonic reasoning over RDF data using

Louise Amoore is Professor of Political Geography in the Department of