Algorithmic Personalization as a Mode of Individuation

Abstract

Recognizing that many of the modern categories with which we think about people and their activities were put in place through the use of numbers, we ask how numbering practices compose contemporary sociality. Focusing on particular forms of algorithmic personalization, we describe a pathway of a-typical individuation in which repeated and recursive tracking is used to create partial orders in which individuals are always more and less than one. Algorithmic personalization describes a mode of numbering that involves forms of de- and re- aggregating, in which a variety of contexts are continually included and excluded. This pathway of a-typical individuation is important, we suggest, to a variety of domains and, more broadly, to an understanding of contemporary economies of sharing where the politics of collectivities, ownership and use are being reconfigured as a default social.

Keywords

algorithm individuation numbering optimization pathway personalization

The less the determinism, the more the possibilities for constraint. (Hacking, 1991: 194)

This is the age of personalization. Personalizing practices permeate everyday life in the UK – we are invited to participate in personalized medical, health and care services, to benefit from personalized customer experiences, to find our way with personalized maps, acquire a personalized education, keep up-to-date with personalized news, get a bargain with personalized prices and so on.¹ To give some more concrete examples: in 2007 the UK government published Putting People First: A Shared Vision and Commitment to Finding New Ways to Improve Social Care in England.² The pamphlet outlined the government’s intention to transform adult social care so ‘that every person who receives support, whether provided by statutory services or funded by themselves, will have choice and control over the shape of that support in all care settings’. This ‘vision’ describes itself as a totally different approach to an historic ‘one size fits all’ system. With an initial focus on transforming social care and support services, the pamphlet proposes that principles of personalization be embedded in a range of other service areas such as health and education. An example from the field of health and well-being is PatientsLikeMe,³ a website that combines features of traditional qualitative online patient communities with quantitative data-collection; the (trade-marked) strap-line is ‘Live better, together!’ This website has 300,000 members, who ‘share’ over 23,000 diseases, and have contributed over 25 million data points about their diseases, resulting in over 50 published research studies. The website says:

By sharing health data on PatientsLikeMe, people not only help themselves, but help others who can learn from their experiences, and advance research.… Learn from others, connect with people like you, track your health.

The platform is also described as a tool that helps patients find a ‘just-in-time’, ‘someone-like-me’ peer who can be relied upon to compare options and aid decision-making. A final example of personalization is the recommendation service Stitch Fix,⁴ a website that describes itself as ‘Your partner in style’ and which seeks to recommend clothing for women on a personal basis. The business proposition is that the recommendation service – a composite of algorithmic and human reasoning – knows better than the customer herself what clothes she will like: selected items of clothing are sent directly to her, without a preview, as otherwise her prejudices might prevent her from choosing items that she, unknowingly, will really like.

The question this article seeks to address is: what kind of individuation (Foucault, 2001; Simondon, 1992) is personalization? We ask this question in order to explore the implications of personalization for how we live together, that is, for forms of sociality. We start from the assumption that personalization is not only personal: it is never about only one person, just me or just you, but always involves generalization. Indeed, our argument will be that it is a mode of individuation in which entities are precisely specified by way of recursive inclusion in types or classes as part of the making of what we describe as an a-typical pathway. To make this argument, we explore the use of recommendation algorithms to sort or classify people, analysing the way in which individuals are addressed as ‘a you’, while their membership of types or classes of person is perpetually revised. Our conclusion is that the familiar recognition that personalization seems to provide – knowing you better than you yourself do – should not be considered as merely a more precise form of individuation. To the contrary, personalization also constrains who and how we can be.

Recognizing that many of the modern categories with which we think about people and their activities were put in place through the use of numbers (Hacking, 1991), we develop our analysis of algorithmic personalization by drawing on an understanding of number as composition (Day et al., 2014). This approach starts from the assumption that numbering is everywhere (Hayles, 2014), even though numbers may not always be visible. As such, it seeks to situate contemporary analyses of algorithms within the wider context of cultures of numbering. As Badiou (2008) remarks, ‘A “cultural fact” is a numerical fact. And, conversely, whatever produces number can be culturally located; that which has no number shall have no name either’ (Badiou, 2008: 2–3). In a similar vein, Totaro and Ninno (2014) also comment on the pervasiveness of numbering, but focus specifically on the performativity of the recursive function, which, they argue, provides ‘an interpretive key to modern rationality’:

The notion that the ‘logic of numbers’ operates exclusively on numbers is misleading. In the second half of the last century, the theory of recursive functions has made it clear that the concept of calculation is very general and does not necessarily imply the manipulation of numerical symbols. (2014: 2; see also Neyland, 2014; Totaro and Ninno, 2016)

More circumspectly, Heintz notes that ‘the observation and regulation of performances today has become mutual and reflexive, generalised and anonymous, and it is now increasingly based on observations and comparisons in terms of numbers’ (Heintz and Vollmer, 2011: 22).

Our compositional approach acknowledges the pervasiveness of numbering in contemporary society by looking at what numbering does, rather than what numbering is. Adopting a felicitous analogy from Verran, we think of numbers in the same way as anthropologists do kin: numbers both are and have relations, just as people are and have relations (Verran, 2010: 171; see also Mackenzie, 2014; Urton, 1997). In other words, we propose that it is as working relations that numbers are able to perform: to travel, to make possible comparison, conversion, and exchange, to be stored, to inform, and to make the same or different. By looking at how numbers are composed or formed in relations, and how social and cultural practices are formed (in part) by number, we aim to show how numbering is a re-presentation – in this case, of persons – that always holds more than one presentation.

To understand how it is that we have become habituated to declaring, measuring and sharing our personal characteristics, behaviours and opinions in the UK in order to carry out mundane activities, we begin by situating our analysis in the context of what has been called a ‘like economy’ (Gerlitz and Helmond, 2013). This context is important, we suggest, insofar as it makes relational value available for computational calculation. Drawing on our compositional approach to numbering, we then develop a set of terms – tracking, bordering, folding and pausing – that lead us to describe forms of personalization that are performed by recommendation algorithms as the making of a pathway of a-typical individuation. Critically, this pathway creates ‘a’ person or individual that is always provisional and corresponds only partially with the type or category in which it is included, whether this concerns what a person might buy, like, share or possess. The term ‘pathway’ is intended to capture this category of person, a category that is never static but always changing and always in motion. While our analytic focus is on the example of algorithmic personalization, and in particular algorithmic personalization that involves the use of collaborative filters, we also make references to other examples which share the same logic.

Liking and Likeness

The rise of a ‘like economy’ begins, so Gerlitz and Helmond (2013) argue, with the arrival of Google in the late 1990s. It is widely known that Google’s early success stemmed from its use of a search engine that shifted the value determination of websites from hits alone to hits and links. The hyperlink analysis algorithm, PageRank, enabled calculation of the relative importance and ranking of a page within a larger set of pages, based on the number of in-links to the page and, recursively, the value of the pages linking to that page and so on. All links do not have equal value in this type of search engine, as links from authoritative sources or links from sources receiving many in-links are weighted in the algorithm.

The use of weighted measures of linking was a first step towards inscribing the capacity to identify and intensify ‘relational value’ in search engine algorithms (Gerlitz and Helmond, 2013; see also Feuz et al., 2011). And it is this relational value, we propose, that is central to personalization insofar as it makes relations between people available for computational calculation. Since this first step, the capacity to make relations of linking – or sharing – has been significantly extended as the determination of ‘authority’ has changed in line with the participatory features of Web 2.0. More web users now participate in making connections between websites through the creation and exchange of user-generated content (as well as gaming and the purchase of position). In particular, social buttons allow users to share, recommend, like or bookmark content, posts and pages across various social media platforms. In 2006, Facebook launched a share icon as a way for someone to share web content and invite re-sharing and then, in 2009, it introduced the Like button. In 2010, the company extended the capacities of the Like button to link by introducing an external Like button, a plugin that could be implemented by any webmaster, potentially rendering all web content like-able. Significantly, the external ‘Like’ button not only captures actual likes, but also aggregates all activities performed on a web object: the number of likes and shares, further likes and comments on stories, and the number of inbox messages containing the object as an attachment. In another important development, Facebook’s Open Graph Protocol opened up their social graph – a representation of people and their connections – for external content, allowing for a controlled way of exchanging preformatted data between Facebook and the external web.

It is through the use of these and other techniques, so Gerlitz and Helmond argue, that Facebook has been able to build a ‘like economy’, that is, an economy that builds on and exploits relational value, mediated by participation. They further suggest that this economy produces what, using Mark Zuckerberg’s own phrase, they call ‘the default social’. To this analysis we want to add the observation that the relations between the individual and the population that characterize this new social are both participatory and participative in that users may participate knowingly (participatory) or unknowingly (participative). Moreover, participation in the default social is mediated by techniques of exclusive inclusion and inclusive exclusion (Agamben, 1998).⁵ On the one hand, the Open Graph is able to include non-users of Facebook as the external Like button cookie can trace non-users and add any information gained as anonymous data to the Facebook database and, on the other hand, a user’s explicitly invited activities may be excluded or rendered invisible to other users if they are not sufficiently highly ranked in the dimensions the graph provides. These oscillating dynamics – of being excluded in ways that inform the ordering of those included, and being included but not in ways that allow you to understand the terms of your membership – were intensified further in 2011 when Facebook expanded the possibilities of ‘invisible’ participation by proliferating custom actions:

When creating an app, developers are prompted to define verbs that are shown as user actions and to specify the object on which these actions can be performed. Instead of being confined to ‘like’ external web content, users can now ‘read’, ‘watch’, ‘discuss’ or perform other actions. (Gerlitz and Helmond, 2013: 1353)

Gerlitz and Helmond conclude their analysis of the emergence of the like economy by suggesting that Facebook is being developed as an ‘infrastructure of decentralised data production and recentralised data processing’ (2013: 1357). While they do not discuss the role of recommendation systems within this infrastructure, these have become increasingly important operators of the de- and re-centralizing practices of the economy Gerlitz and Helmond describe. Structured by the ‘participatory’ practices of inclusive exclusion and exclusive inclusion, recommendation algorithms penetrate all corners of the internet, making personalized recommendations – directly and indirectly – to individuals with interests in a variety of fields, including movies, music, news, books, research publications, restaurants, jokes, financial services, products of all sorts and persons (for example, in online dating). It is to these algorithms that we now turn.

Recommendation Algorithms

While very many different kinds of algorithms are used in recommendation systems, two main kinds are distinguished: collaborative filtering algorithms and content sharing algorithms. Sometimes, as in Netflix, they are combined. The former group of algorithms is based on large amounts of digital data on users’ behaviour, activities or preferences and leads to predictions of what users will like based on their similarity to others (see further below). An example is Last.fm,⁶ a music ‘station’ or streaming service that personalizes the music it transmits by observing the music an individual has listened to on a regular basis and comparing those tracks with the listening behaviour of other individuals. The calculative process involved in this group of algorithms is sometimes described as ‘leveraging’ the behaviour of users⁷ since it requires the participation of many users to produce personalized recommendations for one person.⁸ Content-based filtering methods, in contrast, are based on a description of an item in terms of discrete characteristics; the algorithm is then designed to produce recommendations for individual users of items that have similar properties to those that the individual liked in the past (or is examining in the present). Pandora Internet Radio (currently restricted to listeners in the USA, Australia and New Zealand because of licensing restrictions) is an example: it makes use of an algorithm that uses properties of a song or artists (a subset of the 450 attributes provided by the Music Genome Project) in order to seed a station to transmit personalized music (Prey, 2017).

We focus on collaborative filtering algorithms, partly because their ability to make successful predictions across fields is held to be stronger than that of content-based algorithms, but also because they require and exploit ‘participatory’ methods to develop novel classificatory techniques. As such, they allow us to identify distinctive aspects of personalization as a mode of individuation. Crucially for the use of such algorithms, information relating to ‘pre-existing’ or demographic qualities of the person or entity concerned is not required to produce personalized recommendations. Instead the information required is produced through the aggregation of the ongoing participation of both the individual to whom the recommendations are made and other users of social media. Rather than allocating users to a pre-existing class, group or type (typically a socio-demographic stratum), the properties of which are presumed to be known in advance, the operations of collaborative filtering start from the premise that (individual) customers who share (that is, have in common) some preferences will also share others. This single but powerful assumption is of value for those developing algorithms in that ‘the only information [they need] to work is a set of numerical ratings – specific information about users or items to be recommended is not necessary’ (Seaver, 2012).

It is worth exploring how these numerical ratings are turned into personalized recommendations in some detail. A helpful analysis is provided by Seaver (2012), who describes the ‘signature action’ of collaborative filtering algorithms in terms of the operation of a grid, with items along one side, users along the other, and numerical ratings at their intersections (see also Bowker, 2014; http://personalisedcommunication.net/the-project/). Significantly, this grid is a matrix, that is, a grid with the formatted capacity to map (or perform) network transformations:

This matrix is mostly empty (or ‘sparse’), since most users will have not rated most items. The work of the collaborative filtering algorithm … is to predict what values will show up in the empty spaces of the matrix. These predictions are then provided in some form to the user as recommendations. (Seaver, 2012)

Seaver continues:

At any given time, the matrix is in an anticipatory flux: new ratings from users arrive constantly, displacing their predicted values and shifting the others. The calculative operations involved in this in-filling process is the signature action within the matrix – blank values are replaced by predictions, which are then replaced by actual ratings.⁹

In the next stage described by Seaver, the numbers from the matrix are statistically analysed and their variance is mapped to a number of dimensions or axes. Users who are ‘near’ each other on this multi-dimensional coordinate system are held to be similar (like each other), and a user will be recommended items from the neighbourhood around them. It is this calculative activity that leads to the paradigmatic claim of such algorithms to specify the individual in the complex conjugated personalized address: ‘People like you like things like this.’

But how, if at all, does this description of the calculative logic in recommendation algorithms help us understand what is distinctive about personalization as a mode of individuation? Seaver concludes that preference and similarity are collapsed in this calculative system since ‘liking’ and ‘being like’ are equated. We consider that an understanding of the composition of these relations will help us see how numbers are rendered consequential for the making up of persons (Hacking, 1991). How is this equation accomplished or, in our terms, composed? If we are to address the specificity of personalization as a mode of individuation we need to see the particular ways in which numbers both are and have relations.

Pathways of A-typical Individuation

Seaver’s claim about the equivalence between liking and likeness in recommendation algorithms is less critical to our argument than his observation that this calculative matrix is in a constant state of anticipatory flux. Indeed, we propose that the emphasis on perpetual renewal means that the equation of ‘liking’ and ‘being like’ that is accomplished by these recommendation algorithms is not about establishing relations of absolute equivalence. Instead, we suggest, the calculative activity that produces the anticipatory flux of the matrix involves an ongoing series of approximations in which ‘being like’ and ‘liking’ are continually made more and less like each other in a variety of ways. Such approximations, we would emphasize, are designed to be subject to constant testing.

As Seaver points out, such approximations vary hugely depending on the calculative space in which they are produced (3 or 9 axes or dimensions, for example). Their value – that is, their ability to produce personalized recommendations in terms of criteria of accuracy, diversity (of recommendations), privacy protection and trust – is realized as they are tested repeatedly in relation to data collected via a whole variety of participatory methods and metrics as part of what have been called experiments in participation (Lezaun et al., 2016). In other words, the personalized address to ‘a you’ is not achieved through the collapse of liking and likeness, preference and similarity, but through a carefully calibrated sequencing of their possible inter-relationship. Crucially, this process not only involves the statistical making of proximity or nearness but also the turning of near-ness into next-ness, a process of bordering or adjoining. We conclude therefore that personalization is not the collapse of liking and likeness but the making of a pathway, a dynamic series of approximations of similarity and preference that makes persons.

Indeed, this pathway can be described as a mode of a-typical individuation. What do we mean, though, in our use of a-typical? Certainly it might seem counter-intuitive to use the term at all if it is understood to mean ‘not typical’ or ‘not of a type’, since we have argued throughout that personalization is a mode of individuation that involves generalization through the (repeated or recursive) inclusion of an entity in a type or class. The term ‘a-typical’ is thus used to describe a mode of recursive inclusion, in which both the individual and the type are repeatedly specified anew. In doing so, it draws on multiple – etymologically unrelated – meanings of ‘a’.

The first set of meanings is associated with the use of ‘a’ as the indefinite article, since this use directly indicates membership in a type or class of people, things or events (‘this is a cat’, ‘this is a friend of mine’). The indefiniteness of this inclusion, while appearing to indicate a lack of determination, has its own logic: for example, as well as meaning ‘one single’ or ‘any’, ‘a’ is also commonly used to introduce someone or something for the first time. It allows for a mode of inclusion in a type or category on the basis of criteria that are not pre-given but rather open to further (indefinite) specification (‘If that is what you think, then you are not a friend of mine because a friend of mine would not think that’). As the indefinite article, ‘a’ is also used to specify both someone or something as being like someone or something else (‘you are a star’) and to express rates or ratios, as in ‘for each’ or ‘per’. These meanings of the ‘a’ in the term ‘a-typical’ thus call up the operation of the two analytically distinct, but historically intertwined, understandings of analogy identified by Stafford (2001): participation (similitude, mimesis, likeness) and proportion (ratio).¹⁰ In our use, they are combined to produce principles of inclusion that are subject – recursively – to further revision: their combination is the means by which the you that is ‘a you’ becomes a recursive shifter (Chun, 2011, 2016).

To these meanings of ‘a’ as the indefinite article, however, we add a further meaning, that is, ‘a’ as a variant spelling of ‘ad-’, denoting motion or direction, a reduction or change into, an addition, increase or intensification, as in ‘adjoin’ or ‘adjacent’. The etymology of these terms relates to the Latin adjacentem, adjacens; from adjacere, ‘to lie at, to border upon, to lie near’; from ad-, ‘to’ + jacere, ‘to lie, to rest’; literally, ‘to throw’. Our use of the term a-typical to describe pathways of individuation is thus intended to describe the ways in which collaborative filtering algorithms are designed to allow for the ongoing redefinition of principles of inclusion and exclusion via the recursive activity of adjoining or the work of adjacency: what we describe as the compositional practice of bordering or framing.¹¹ In this practice, the aim is to create, not equivalence, but a topological invariance: that is, the aim is to achieve a continuity of a recursive function¹² such that likeness (‘People like you’) is iteratively produced as a pathway through a massively aggregated de- and re-contexting of liking.

How this is accomplished in the multi-dimensional calculative space of recommendation algorithms can be illustrated by way of a consideration of ‘the next adjacent possible’, a term developed by the theoretical biologist Stuart Kauffman (2000).¹³ Put briefly, Kauffman understands life in terms of autonomous agents,¹⁴ by which he means ‘something that can act on its own behalf in an environment’. This living entity is ‘something that can both reproduce itself and do at least one thermodynamic work cycle’ (Kauffman, 2000: 64). He says:

That bacterium, sculling up the glucose gradient, flagellum flailing in work cycles, is busy as hell doing ‘it’, reproducing and carrying out one or more work cycles. So too are all free-living cells and organisms. We do, in blunt fact, link spontaneous and nonspontaneous processes in richly webbed pathways of interaction that achieve reproduction and the persistent work cycles by which we act on the world. Beavers do build dams; yet beavers are ‘just’ physical systems. (Kauffman, 2000: 64)

In making this argument, Kauffman emphasizes the importance of the role of adjoining or bordering and points to the constitutive role of the particular material constraints (or context) in which any entity individuates. He also identifies the work of adjacency as the activity of ‘constructing constraints that can manipulate constraints’, thus drawing attention to the role of the border as the operator of the relation of the inside of an entity to its outside.

Drawing on this analysis, we return to our observation on testing. Personalized recommendations are based on the making of nearness or adjacency in a multi-dimensional space, but the implementation of collaborative filtering algorithms requires that they be subject to repeated testing in the specific kinds of relations to context that are commonly called participation. The aim of this testing is to identify constraints that can manipulate constraints; for example, in A/B tests or tests of ‘liking’, to identify a changed ordering of likes so that some future preference becomes more likely – or predictable. It is only insofar as a population’s relations to multiple contexts (including data relating to liking, sensing and sharing as well as to time and space) are registered by the algorithm that the mode of individuation we are describing can happen at all. In other words, the (numerical-cultural) process of folding a whole into, across or within itself to make parts, of de- and re-contexting what Zuckerburg describes as the default social, is fundamental to the making of pathways of a-typical individuation. As Seaver (2015) observes, while it is sometimes claimed that big data has no context, ‘context is everything’ for recommendation algorithms.¹⁵

It has been widely observed that algorithms do not operate in isolation from context-aware techniques of data capture and collection as they are organized in particular calculative infrastructures (Hayles, 2002; Verran, 2011). Dourish, for example, notes:

If the database is malleable, extensible, or revisable, it is so not simply because it is represented as electrical signals in a computer or magnetic traces on a disk; malleability, extensibility, and revisability depend too on the maintenance of constraints that make this specific collection of electrical signals or magnetic traces work as a database; and within these constraints, new materialities need to be acknowledged. (Dourish, 2014)

Similarly, Amoore and Piotukh highlight the changing role of indexing practices in data-collection activities:

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. (Amoore and Piotukh, 2015: 345)

As they also observe, the activity of ‘(ad-)joining’ is of particular importance in the deployment of these new forms of indexing. They give the example of IBM’s predictive policing:

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). (Amoore and Piotukh, 2015: 345)

It is not just that there is more than one relevant context for recommendation algorithms. Different contexts are deliberately made to appear or disappear in different practices of context-ing. Indeed, this emphasis on context – what is sometimes called context-awareness – provides another compelling reason to describe personalization as a pathway of a-typical individuation. A trajectory is not established in advance – as when we travel with the aim of moving from A to B, already knowing where B is – but in response to contexts that emerge in the making of a path.

Becoming Normal by Being Better than You

We turn now to a discussion of the consequences of personalization for the making of the default social, by considering the practice of normalization (Agamben, 1998; Canguilhem, 1991; Foucault, 1991; Hacking, 1991).¹⁶ In his discussion of modes of governance linked to earlier forms of statistical normalization, Hacking (1991) argues that debates concerning the setting of boundary conditions were fundamental to the way in which a population was governed by statistical laws. Updating this argument, we suggest that the work of adjoining in the personalization practices described above involves an ongoing reorganization of boundary conditions (operating the relation between inside and outside, inclusion and exclusion through techniques of contexting) that transforms conditions of governmentality. This is especially clear in relation to the way in which practices of normalization now require the achievement of transitivity.¹⁷ On the one hand, the verbs of the vocabulary of participation – liking, sharing, linking – describe activities in which objects are repeatedly attached to persons; that is, they promote an algorithmic kind of linguistic transitivity (as in ‘things like this like people like you’). On the other hand, the data collected through the tracking of participation are then ordered transitively – in a mathematical sense – in an n-dimensional space of likeness or similitude. In these practices, the ‘new normal’ of individuation appears as a function of the ideal of transitive closure, an internal limit, in relation to which every possible relation (between verb and object) is partially ordered in such a way that the you that is ‘a you’ is similar to other ‘yous’, that is, nearly but not quite the same as other ‘yous’, and never quite able to be consolidated as an ‘us’.

While this limit can never be reached since it involves a never-ending in-filling in relation to a constantly changing population,¹⁸ we are nonetheless witnessing a proliferation of models of optimization across the fields of medicine, marketing, project management and operational research (the last of which is sometimes described as ‘the science of better’, the significance of which will be made apparent below). In such models, optimal pathways of a-typical individuation are commonly identified in ‘experiments in participation’ in relation to specific objectives, often through software that merges data with parameters (as in the case of the parametric algorithms discussed by Parisi, 2013) or employs evolutionary modelling. As described above, one of the novel aspects of such techniques is the calculative deployment of recursion such that the aim of the action of ad-joining is not set in relation to a predefined target; rather pathway and target emerge together.

Indeed, the term ‘precision medicine’ is sometimes preferred to the synonyms personalized or stratified medicine because it acknowledges the significance of the necessarily dynamic fit between, for example, a cancer, drug target, resistance and side effects through repeated monitoring and the operationalization of the feedback loop between evaluation and intervention.¹⁹ In some cases, the methods of operational research are applied in conjunction with computational biology with the aim of identifying a pathway that has a ‘biologically meaningful objective’: a network is ‘designed (or revised) optimally’ to find ‘the natural circumstances that trigger one particular pathway but not others’.²⁰ An example of findings based upon pathways defined in molecular terms rather than by anatomy or traditional disease classification is the recently reported study (Mateo et al., 2015) of the efficacy of the drug Olaparib, approved for treating ovarian cancers with BRCA1/2 mutations. This study built upon the finding that cancers are significantly heterogeneous at the molecular level and discovered that the variation within one, such as ovarian cancer, can be more marked than between cancers, such as ovarian and prostate, when tracked in terms of their differential sensitivity to particular treatments.

More broadly, we can see the operation of principles of optimization modelling in the now ubiquitous ordinal tropes of ranking, which ensure that what counts as best is not given in advance, but rather emerges in a participative fashion with the (continually changing) requirement to do and be better (Esposito, 2013; Gerlitz and Lury, 2014; Guyer, 2010). In these practices the you that is addressed is both specific and a you ‘that is like everyone else’ (Chun, 2011), only more or less so. The exhortation to ‘Believe in Better’²¹ pervades contemporary culture and might be seen as an appropriation of ‘optimism of the will’, recursively calibrating relations between individuals and populations to establish new forms of stratification (Fourcade and Healey, 2013; Ruppert, 2011). In the requirement to be like but better than each other established in relation to such optimizing practices, you and I are not just different to each other but different-er: our differences are such that we are always both more and less different to each other. As the Optimizely commercial platform informs us, ‘Being personal is no longer optional’,²² or, as the name of a British financial services comparison website says, GoCompare.²³ Indeed, it is not just persons that are invited – or obliged – to participate in bettering themselves in the compositional practices of personalization: universities, hospitals, museums, police forces, hotels, holidays, restaurants, brands and schools are also now frequently placed in dynamic relations of competitive comparison with each other by often mandatory or non-voluntary inclusion in the recursive partial orderings of ranking systems. While normalization techniques sometimes provide a statistical snapshot, a one-off cross-section of a population fixed in relation to a single environment (the nation, for example), personalization is noteworthy for the way that it establishes (constantly shifting) grounds for dynamic stratification in relation to multiple norms in multiple environments.

Signature Pathways

We consider one further aspect of the making of a pathway of a-typical individuation by exploring the use of ‘you’ as a shifter. In linguistic terms, shifters such as ‘this’ and ‘that’ as well as ‘I’ and ‘you’ can only be understood by reference to the context in which they are uttered. In other words, a shifter, sometimes also called a place-holder, is an indexical term whose meaning cannot be determined without referring to the message that is being communicated. The ‘you’ in a pathway of personalization designates both the person to whom a message is directed and the ‘you’ that is contained in the message that is sent. In relation to our description of algorithmic personalization, it is the suturing of this doubling in the shifter that makes a personalized address to the individual possible and also organizes the activity of shifting as adjoining, creating constraints that can manipulate constraints in the making of a pathway.

For Jakobson (1971 [1957]), enunciation is encoded in a shifter in the statement itself. While Jakobson defines the shifter as an indexical symbol, Lacan defines it as an indexical signifier in order to problematize the distinction between enunciation and statement. As a signifier, the shifter ‘I’ is normally part of a statement. As an index, it is also normally part of the enunciation. For Lacan (1977), this division or distribution of the ‘I’ or ‘you’ does not merely illustrate the splitting of a subject; it is that split. Drawing on these understandings of shifters, it seems that the indexical signifier is not stopped or ‘arrested’ by (representatives of) the symbolic order (Fenves, 2002) in the anticipatory flux of personalizing practices.²⁴ In the context of (algorithmic) personalization, it seems that the shifter is rather paused. Temporary halting incites participation or the folding of a context into the pathway. Indeed, it is this pausing, the marking of an interval, a stopping and starting that repetitively gathers a collectivity. In assembling observers and observed, pausing allows for both observation and the observation of the observing (Kaldrack and Röhle, 2014).

Given that a pathway is a process of stopping and starting that repetitively gathers a collectivity, it is not surprising that the ability to identify some pathways but not others – the signature action Seaver describes – is currently the source of considerable interest. Frow’s discussion of signature and brand (2002) is illuminating in this respect. He describes the signature as a shifter that sets up ‘a tension between representation and the represented’ and observes that the signature is not only an index of the act of framing (of adjoining or bordering), but also designates a naming right. Specifically, Frow argues that the power of the signature stems from the elision of the difference between the signature as an index and the taxonomic function of the proper name. This elision is effected in a particular way by the brand, he asserts, since ‘the “Name”, when one abstracts it from the signature it indicates, loses its “index” character and becomes a “trademark”. Like the trademark, the name is of a symbolic order’ (Gandelman, quoted in Frow, 2002: 63).

As Frow observes, the brand’s economic significance as a ‘nexus between high-speed, continuous flow manufacturing and the reshaping of people’s habits and lives’ (Ohmann, 1996: 61 in Frow, 2002: 64) is growing. The detachment from indexicality is what provides the basis for using the signature as a claim to ownership. Importantly, however, Frow argues that the brand is in principle reducible to neither a product nor a corporation. As a quasi-signature or signature-effect, a brand name is routinely attached to a product range, and even to generations of product ranges, rather than to singular objects. It is precisely the divisibility of brand from product (in practices of bordering or framing) that makes possible the transfer of brand loyalty from one generation of a product to another. With Frow’s insights, we propose that recommendation algorithms create pathways of a-typical individuation that are always distinct (divisible and detachable) from both object and person. In consequence, the ways in which such pathways acquire autonomy or not, and how that autonomy is recognized,²⁵ constitute the heart of current debates on the sharing economy. It is here that the politics of collectivity, ownership and use are being reconfigured.

Conclusion

We have argued that personalization is a mode of a-typical individuation that is produced in techniques of recursive divisibility (the drawing of lines of inclusive exclusion and exclusive inclusion). As such, it provides an entry point into the constitution of what, following Mark Zuckerburg, we have called the ‘default social’. Crucially, as a numbering practice, personalization does not involve zooming (Day et al., 2014), a performative gesture that operates the dynamism of moving from big to small, that is, a slide from one to many and back again, as if the only difference to be registered was that of an increase in a uniform quantity (as in what Badiou calls the count of one). Instead, this is a mode of numbering that constitutes a default social through forms of de- and re-aggregating, in which a variety of contexts are included and excluded, such that one is always more and less than one. In a recursive process that involves tracking bordering, folding and pausing, the individual is precisely and momentarily specified as ‘a you’ (Chun, 2016), that is, as a dividual (Raunig, 2015; Strathern, 1998). At the same time, pausing allows for the composition of heterogeneous (numerical-cultural) quantities, in which qualitative differences of mass are recognized at different levels of observation as matters of dimension and scale. Put somewhat differently, the person who is addressed as a you is refracted in multiple partial orderings that allow for specific forms of comparison and competition (of better-ing) while the folding of contexts into the pathway creates new ways of configuring relations between participation and proportion, sharing, ownership and use in the identification of signature pathways.

Importantly, our argument does not suggest that personalization is replacing other modes of individuation. Rather, it introduces new techniques that combine in a variety of ways to transform and intensify contemporary forms of individualism.²⁶ As such, it merely confirms Hacking’s observation in relation to the history of the making up of people: ‘The less the determinism, the more the possibilities for constraint’ (1991: 194).

Footnotes

Notes

Celia Lury is Professor and Director of the Centre for Interdisciplinary Methodologies, Warwick University. The topic of her ESRC Professorial Fellowship was ‘Order and Continuity: Methods for Change in a Topological Society’, ES/K010689/1.

Sophie Day is Professor of Anthropology at Goldsmiths, and Visiting Professor at the School of Public Health, Imperial College London. She is currently working with Celia Lury and Helen Ward in a collaborative project, ‘People Like You’: Contemporary Figures of Personalisation (Wellcome Trust, 205456/Z/16/Z).

This article is part of the Theory, Culture & Society special issue on ‘Thinking with Algorithms: Cognition and Computation in the Work of N. Katherine Hayles’, edited by Louise Amoore.

References

Agamben

(1998) Homer Sacer: Sovereign Power and Bare Life, Stanford, CA: Stanford University Press.

Agamben

(2011) The Signature of All Things: On Method, Cambridge: Zone Books, MIT Press.

Althusser L (1971) Lenin and Philosophy and other Essays, trans. Brewster B. New York: Monthly Review Press.

Amoore

Piotukh

(2015) Life beyond big data: Governing with little analytics. Economy and Society 44(3): 341–366.

Arvidsson

(2016) Facebook and finance: On the social logic of the derivative. Theory, Culture & Society 33(6): 3–23.

Badiou

(2008) Number and Numbers, Cambridge: Polity Press.

Bateson

(1999) Steps to an Ecology of Mind, Chicago: University of Chicago Press.

Bourdieu P (1987) Distinction: A Social Critique of the Judgement of Taste, trans. Nice R. Cambridge, MA: Harvard University Press.

Bowker GC (2014) Big data, big questions: The theory/data thing. International Journal of Communication 8(2043): 1795–1799. Available at: https://ijoc.org/index.php/ijoc/article/view/2190 (accessed: 12 December 2018).

10.

Canguilhem G (1991) The Normal and the Pathological, trans. Fawcett C. Cambridge: Zone Books, MIT Press.

11.

Chun

WHK

(2011) Programmed Visions, Cambridge, MA: MIT Press.

12.

Chun

WHK

(2016) Updating to Remain the Same, Cambridge, MA: MIT Press.

13.

Coleman

(2017) Developing speculative methods to explore speculative shipping: Mail art, futurity and empiricism. In: Wilkie

Savransky

Rosengarten

(eds) Speculative Research: The Lure of Possible Futures, London: Routledge.

14.

Day

Lury

Wakeford

(2014) Number ecologies: Numbers and numbering practices. Distinktion 15(2): 123–154.

15.

Dourish

(2014) No SQL: The shifting materialities of database technology. Computational Culture 4.

16.

Esposito

(2013) Economic circularities and second-order observation: The reality of ratings. Sociologica 2: 1–20.

17.

Fenves

(2002) Arresting Language: From Leibniz to Benjamin, Stanford, CA: Stanford University Press.

18.

Feuz M, Fuller M and Stalder F (2011) Personal web searching in the age of semantic capitalism: Diagnosing the mechanisms of personalisation. First Monday 16(2, 7 February).

19.

Foucault M (1991) Discipline and Punish: The Birth of the Prison, trans. Sheridan A. Harmondsworth: Penguin.

20.

Foucault

(2001) The Order of Things: An Archaeology of the Human Sciences, Abingdon: Routledge.

21.

Fourcade

Healy

(2013) Classification situations: Life-chances in the neoliberal era. Accounting, Organizations and Society 38(8): 559–572.

22.

Frow

(2002) Signature and brand. In: Collins

(ed) High-pop: Making Culture into Popular Entertainment, Malden, MA: Wiley-Blackwell, pp. 56–74.

23.

Gerlitz

Helmond

(2013) The like economy: Social buttons and the data-intensive web. New Media & Society 15(8): 1348–1365.

24.

Gerlitz

Lury

(2014) Social media and self-evaluating assemblages: On numbers, orderings and values. Distinktion 15(2): 174–188.

25.

Guyer

(2010) The eruption of tradition: On ordinality and calculation. Anthropological Theory 10(1–2): 123–131.

26.

Hacking

(1991) How should we do the history of statistics? In: Burchell

Gordon

Miller

(eds) The Foucault Effect, Chicago: University of Chicago Press.

27.

Hayles

(2002) Writing Machines, Cambridge, MA: MIT Press.

28.

Hayles

(2014) Cognition everywhere: The rise of the cognitive nonconscious and the costs of consciousness. New Literary History 45(2): 199–220.

29.

Heintz B and Vollmer H (2011) Globalizing comparisons: Performance measurement and the ‘numerical difference’ in global governance. Available at: http://citation.allacademic.com//meta/p_mla_apa_research_citation/5/0/0/4/8/pages500487/p500487-1.php.

30.

Hood

Friend

(2011) Predictive, personalized, preventive, participatory (P4) cancer medicine. Nature Reviews Clinical Oncology 8: 184–187.

31.

Jakobson R (1971 [1957]) Shifters, verbal categories, and the Russian verb. In: Selected Writings, vol. II: Word and Language. The Hague: Mouton.

32.

Kaldrak I and Röhle T (2014) Divide and share: Taxonomies, orders and masses in Facebook’s Open Graph. Computational Culture 4.

33.

Kauffman

(2000) Investigations, Oxford: Oxford University Press.

34.

Lacan J (1977) Écrits: A Selection, trans. Sheridan A. London: Tavistock.

35.

Lezaun

Marres

Tironi

(2016) Experiments in participation. In: Miller

Smitt-Doer

Felt

Fouche

(eds) Handbook of Science and Technology Studies, Cambridge, MA: MIT Press, pp. 195–222.

36.

Mackenzie

(2014) Multiplying numbers differently: An epidemiology of contagious convolution. Distinktion: Journal of Social Theory 15(2): 189–207.

37.

Mateo J, Carreira S, Sandhu S et al. (2015) DNA-repair defects and Olaparib in metastatic prostate cancer. New England Journal of Medicine online 29 October.

38.

Neyland

(2014) On organizing algorithms. Theory, Culture & Society 32(1): 119–132.

39.

Parisi

(2013) Contagious Architecture: Computation, Aesthetics and Space, Cambridge, MA: MIT Press.

40.

Prey

(2017) Nothing personal: Algorithmic individuation on music streaming platforms. Media, Culture & Society 40(7): 1086–1100.

41.

Rabinow

(2009) Marking Time: On the Anthropology of the Contemporary, Princeton, NJ: Princeton University Press.

42.

Raunig G (2015) Dividuum, Machinic Capitalism and Molecular Revolution, trans. Derieg A. Cambridge, MA: MIT Press.

43.

Ruppert

(2011) Population objects: Interpassive subjects. Sociology 45(2): 218–233.

44.

Seaver N (2012) Algorithmic recommendations and synaptic functions. Limn 2, Crowds and Clouds.

45.

Seaver

(2015) The nice thing about context is that everyone has it. Media, Culture & Society 37(7): 1101–1109.

46.

Simondon

(1992) The genesis of the individual. In: Crary

Kwinter

(ed) Incorporations, Cambridge: Zone Books, MIT Press, pp. 297–319.

47.

Stafford

(2001) Visual Analogy: Consciousness as the Art of Connecting, Cambridge, MA: MIT Press.

48.

Strathern

(1998) The Gender of the Gift, Berkeley: University of California Press.

49.

Tkacz

(2014) Wikipedia and the Politics of Mind, Chicago: University of Chicago Press.

50.

Totaro

Ninno

(2014) The concept of algorithm as an interpretive key of modern rationality. Theory, Culture & Society 31(4): 29–49.

51.

Totaro

Ninno

(2016) Algorithms and the practical world. Theory, Culture & Society 33(1): 139–152.

52.

Urton

(1997) The Social Life of Numbers: A Quechua Ontology of Numbers and Philosophy of Arithmetic, Austin, TX: University of Texas Press.

53.

Verran

(2010) Number as an inventive frontier in knowing and working Australia’s water resources. Anthropological Theory 10(1): 171–178.

54.

Verran

(2011) The changing lives of measures and values: From centre stage in the fading ‘disciplinary’ society to pervasive background instrument in the emergent ‘control’ society. Sociological Review 59(2): 60–72.