Roger Clarke's Web-Site
© Xamax Consultancy Pty Ltd, 1995-2017
|Identity Matters||Other Topics||Waltzing Matilda||What's New|
Revised Version of 27 April 2015
Proc. 28th Bled eConference, Slovenia, 7-10 June 2015
Roger Clarke **
© Xamax Consultancy Pty Ltd, 2014-15
Available under an AEShareNet licence or a Creative Commons licence.
This document is at http://www.rogerclarke.com/EC/BDSA.html
Big data and big data analytics have been the subject of a great deal of positive discussion, not only in traditionally upbeat popular management magazines but also in nominally scientific and therefore professionally sceptical academic journals. Research was undertaken to assess the impact on the quality of inferences drawn from big data of the quality of the underlying data and the quality of the processes applied to it.
Empirical study is difficult, however, because big data is emergent, and hence the phenomena are unstable, and likely to vary considerably across different settings. A research technique was accordingly sought that enabled theoretical treatment to be complemented by consideration of real-world data. The paper introduces Quasi-Empirical Scenario Analysis, which involves plausible story-lines, each of which commences with a real-world situation and postulates lines of plot-development. This enables interactions among factors to be analysed, potential outcomes to be identified, and hypotheses to be generated.
The set of seven scenarios that was developed investigated the nature and impacts of shortfalls in data and decision quality in a range of settings. In all cases, doubts arise about the reliability of inferences that arise from big data analytics. This in turn causes concern about the impacts of big data analysis on return on investment and on public policy outcomes. The research method was found to offer promise in the challenging contexts of technologies in the process of rapid change.
As sensor technologies have matured, and as individuals have been encouraged to contribute data into organisations' databases, more transactions than ever before have been captured. Meanwhile, improvements in data-storage technologies have resulted in the cost of evaluating, selecting and destroying old data being now considerably higher than that of simply letting it accumulate. The glut of stored data has greatly increased the opportunities for data to be inter-related, and analysed. The moderate enthusiasm engendered by 'data warehousing' and 'data mining' in the 1990s has been replaced by unbridled euphoria about 'big data' and 'data analytics'. What could possibly go wrong?
Professional and management periodicals on big data topics have a very strong focus on opportunities, and so do most academic papers in the area to date. Far too little attention has been paid to the threats that arise from re-purposing data, consolidating data from multiple sources, applying analytical tools to the resulting collections, drawing inferences, and acting on them. The research reported on in this paper was conceived as a way to test the key claims made in the 'big data' literature, together with some of its implicit assumptions.
The paper commences by briefly reviewing the emergent theory and current practice of big data. Working definitions are provided that distinguish the processes whereby the data becomes available from the processes whereby inferences are drawn from it. A short summary is then provided of theory relating to the central concepts of data quality and decision quality. Formal empirical research techniques are difficult to apply to emergent phenomena. A research technique is described which is intended to provide a quasi-empirical base to complement the theoretical analysis. A set of scenarios is outlined, and the theoretical material is then used as a lens to gain insights into the nature of big data and big data analytics, and the risks that they entail.
During the 1980s, organisations had multiple data collections that were largely independent of one another. In order to enable the manipulation of data structures and content without disrupting underlying operational systems, copies of data were extracted from two or more collections and stored separately, in what was referred to as a 'data warehouse' - "a copy of transaction data specifically structured for query and analysis" (Jacobs 2010. See also Inmon 1992 and Kimball 1996). The processing of the contents of 'data warehouses' was dubbed 'data mining' (Fayyad et al. 1996, Ratner 2003, Ngai et al. 2009, Hall et al. 2009).
During the same era, government agencies that could gain access to data-sets from multiple sources practised an additional technique called data matching. This involves comparing machine-readable records from different data-sets that contain data that appears to relate to the same real-world entities, in order to detect cases of interest. In most data matching programs, the category of entities that is targeted is human beings (Clarke 1994b, 1995a).
A further technique that has long been applied in both the public and private sectors is profiling. This is "a technique whereby a set of characteristics of a particular class of person is inferred from past experience, and data-holdings are then searched for individuals with a close fit to that set of characteristics" (Clarke 1993). Although profiles can be ad hoc, they have been increasingly "supported by analyses of existing data-holdings within and beyond the organisation, whereby individuals who are known to belong to that class are identified, their recorded characteristics examined, and common features isolated".
During the last few years, there has been a resurgence in enthusiasm for such techniques. The range and intensity of data capture has greatly increased in the intervening years. In addition, the economics of storage and destruction have shifted, such that retaining data is now cheaper than deleting it. This has led a wide variety of media commentators, consultants and excitable academics making enthusiastic pronouncements about revolutions, break-throughs and sparkling new opportunities.
One of the most-quoted authorities on big data is Mayer-Schonberger & Cukier (2013), which had accumulated over 700 citations in its first 21 months after publication. According to those authors, the cornerstone of big data thinking is 'datafication' - the expression of phenomena "in a quantified format so it can be tabulated and analyzed" (p.78). This they say undermines the kinds of analyses conducted in the past: "With enough data, the numbers speak for themselves. Petabytes allow us to say: Correlation is enough ... [W]hen you are stuffed silly with data, you can tap that instead [of experience, expertise and knowledge], and to greater effect. Thus those who can analyze big data may see past the superstitions and conventional thinking not because they're smarter, but because they have the data ... [S]ociety will need to shed some of its obsession for causality in exchange for simple correlations: not knowing why but only what ... Knowing why might be pleasant, but it's unimportant ... [L]et the data speak" (pp. 71, 143, 7, 52, 141).
At its most extreme, this argument postulates that understanding is no longer necessary, and that the ready availability of vast quantities of data justifies the abandonment of reason: "If the statistics ... say it is, that's good enough. No semantic or causal analysis is required. ... [M]assive amounts of data and applied mathematics replace every other tool that might be brought to bear. Out with every theory of human behavior, from linguistics to sociology. Forget taxonomy, ontology, and psychology. ... [F]aced with massive data, [the old] approach to science -- hypothesize, model, test -- is becoming obsolete. ... Petabytes allow us to say: 'Correlation is enough'" (Anderson 2008).
Such bombast is common in papers on big data. More sceptical views do exist, however. For example, big data is "a cultural, technological, and scholarly phenomenon that rests on the interplay of [three elements]" boyd & Crawford (2012, p.663). Their first two elements are technical, and are examined below. Their third element, on the other hand, emphasises the importance of "mythology: the widespread belief that large data sets offer a higher form of intelligence and knowledge that can generate insights that were previously impossible, with the aura of truth, objectivity, and accuracy".
Where definitions of big data are offered, they tend to use imprecise and upbeat explanations, and to use a single term to encompass both the data and the techniques applied to it. A common expression is 'data that's too big, too fast or too hard for existing tools to process', and many writers refer to the distinguishing features being 'volume, velocity and variety'. Even the definition offered by an international agency is vague: "the capacity to analyse a variety of mostly unstructured data sets from sources as diverse as web logs, social media, mobile communications, sensors and financial transactions" (OECD 2013, p.12). An aspect that appears in some discussions is the accumulation of measurements of the same phenomenon over time, resulting in a longitudinal dimension within the data collection.
This paper distinguishes 'big data' from 'big data analytics', and proposes working definitions, listed in Table 1.
'Big Data' refers to a relatively large data collection that is associated with, or purports to be associated with, or can be interpreted as being associated with, particular entities. The following sub-categories can be usefully distinguished:
'Big Data Analytics' refers to the techniques whereby a big data collection is used to draw inferences about an entity, or a category of entities, or relationships among entities
Earlier research performed by the author distinguished a number of categories of purpose to which big data analytics may be applied. These are extracted in Table 2.
After Clarke (2014b)
Big data analysis may test hypotheses about populations, which may be predictions from theory, existing heuristics, or hunches. Inferences may be drawn about the digital personae, which may be further applied to the population of entities that the data purports to represent, or to segments of that population. Profiles may be constructed for categories of entities that are of specific interest, such as heavy weather incidents, risk-prone card-holders, wellness recipes, or welfare cheats.
Further functions that can be performed are concerned not with populations but with individual instances. Outliers of many different kinds can be identified. Inferences can be drawn about individual entities, directly from a particular digital persona, or from apparent inconsistencies within a consolidated digital persona, or through comparison of a digital persona against the population of similar entities, or against a previously-defined profile, or against a profile created through analysis of that particular big data collection.
The performance of these functions places a great deal of reliance on the analytical tools and the big data collections to which they are applied. The purpose of the research reported in this paper is to assess the impact of the quality of the big data, and of the big data analytics, on the quality of inferences drawn. The following section juxtaposes against big data notions the understanding accumulated within the information systems literature of the characteristics of data quality and decision quality.
During its first 50 years, the information systems (IS) discipline has adjusted its focus over time. Underlying many threads of IS theory, however, are characteristics of data that influence the effectiveness and efficiency of decision-making based on it. This section briefly summarises key aspects of data and decision quality.
The computing and information systems literatures contain a body of material relating to data quality factors. See, for example, OECD (1980), Huh et al. (1990), Clarke (1995b, pp. 601-605), Wang & Strong (1996), Rusbridge et al. (2005), English (2006), Piprani & Ernst (2008), and the still-emergent ISO 8000 series of Standards. The primary factors are listed in Table 3.
Each item of data is a measurement against some kind of scale. Some data arises from measurement against a ratio scale, and is capable of being subjected to analysis by powerful statistical tools. Frequently, however, data arises from measurements against cardinal, and against merely ordinal scales, such as Likert-scale data. Such data is capable of analysis using a smaller collection of statistical tools. However, researchers often assume for convenience that the data was gather against a ratio scale, in order to justify the application of more powerful inferencing techniques. Meanwhile, a great deal of data is collected against nominal scales, including text, sound, images and video. This supports only weak analytical tools, fuzzy matching and fuzzy logic. A further challenge arises where data that has been measured against different kinds of scales is consolidated and analysed. The applicability of analytical tools to mixed-scale data is more of a murky art than a precise science.
A further consideration when assessing the quality of data is that what is collected, against what scales, and with what trade-offs between data quality and collection costs, all reflect the purpose of collection and the value-judgements of the sponsor. Particularly where data is collected frequently over time, data collection may also involve compression, through sampling, filtering and averaging. Further, where interesting outliers are being sought, compression is likely to ensure that the potentially most relevant data is absent from the collection.
To be useful, data needs to be associated with real-world entities, with the data currently held about any particular entity being that entity's 'digital persona' (Clarke 1994, 2014a). The reliability of the association depends on the attributes that are selected as identifiers, and the process of association has error-factors. In some circumstances the link between the digital persona and the underlying entity is unclear (the condition referred to as pseudonymity), and in some cases no link can be achieved (anonymity). In order to protect important interests and comply with relevant laws, the link may be broken (the process of de-identification or anonymisation - UKICO 2012, DHHS 2012). On the other hand, rich data-sets are vulnerable to re-identification procedures. These problems afflict all big data collections, particularly those that are intended to support longitudinal studies. Where the data is sensitive, significant public policy issues arise.
Over time, many threats arise to data integrity, including the loss of metadata such as the scale against which data was originally collected, the data's definition at the time of collection, the data's provenance, any supporting evidence for the data's quality, undocumented changes in meaning over time, and loss of contextual information that would have enabled its appropriate interpretation.
The big data movement commonly features the use of data for a purpose extraneous to its original purpose - and indeed many big data proponents champion this proposition. The many data quality issues identified above are exacerbated by the loss of context, the lack of clarity about the trade-offs applied at the time of collection, and the greatly increased likelihood of misinterpretation.
The big data movement commonly involves the further step of physically or virtually consolidating data from multiple sources. This depends on linkages among data-items whose semantics and syntactics are different, perhaps substantially, perhaps subtly. The scope for misunderstandings and misinterpretation multiply.
Theorists and practitioners perceive deficiencies in the data, such as missing elements, and inconsistencies among seemingly similar data-items gathered from two or more sources. To address these concerns about the analysts' particular perceptions of data quality, they devise data 'scrubbing', 'cleansing' or 'cleaning' processes. Some of these processes use an external, authoritative reference-point, such as a database of recognised location-names and street-addresses. Most, however, lack any external referent, and are merely based on 'logical data quality', i.e. internal consistency within the consolidated data-sets: "rules are learned from data, validated and updated incrementally as more data is gathered and based on the most recent data" (Saha & Srivastava 2014. See also Jagadish et al. 2014). Such process guidance as exists overlooks intrinsic and contextual data quality, and omits controls and audit (e.g. Guo 2013). As a result, the notion of 'cleanliness' primarily relates to the ease with which analytical tools can be applied to the data, rather than to the quality of the data.
It is feasible for big data analytics to be applied as a decision system. This may be done formally, by, for example, automatically sending infringement or 'show cause' notices. However, it is also possible for big data analytics to become a decision system not through conscious intent by an organisation, but by default. This can arise where a decision-maker becomes lazy, or is replaced by a less experienced person who is not in as good a position to apply human intelligence as a means of checking the reasonableness of the inferences drawn by software.
Where decisions are made by analytics, a number of concerns arise about decision-quality. In Clarke (2014b), three were highlighted:
Those theoretical aspects lead to practical questions. On what scale, with what accuracy and what precision, was the data collected that was instrumental in leading to the decision, and was the inferencing mechanism that was used really applicable to those categories of data? Did the data mean what the inferencing mechanism implicitly treated it as meaning? To the extent that data was consolidated from multiple sources, were those sources compatible with one another in respect of the data's scale, accuracy, precision, meaning and integrity?
In many circumstances to which big data is claimed to be applicable, real-world decisions depend on complex models that feature confounding, intervening and missing variables. Correlations are commonly of a low grade, yet may nonetheless be treated, perhaps implicitly, as though the relationships were causal, and causal in one direction rather than the other. Big data proponents blithely dismiss causality and champion correlation instead. Yet mens rea (i.e. intention to cause the outcome) is a fundamental element of most criminal prosecutions, and many decisions in civil jurisdictions also focus on the proximate cause of an event. Big data proponents are swimming against a strong tide of cultural and institutional history. If decisions are being made that have real impacts, is the decision-process, and are the decision-criteria, transparent? And are they auditable? Are they subject to appeal processes and review? And can society have confidence that risks and liabilities are appropriately allocated?
Big data analytics may be more commonly used as a form of decision support system, whereby a human decision-maker evaluates the inferences before applying them to any real-world purpose. However, the person may have great difficulty grasping the details of data provenance, data quality, data meaning, data relevance, and the rationale that have given rise to the recommendation. An even more problematic situation arises where the nominal decision-maker is not in a position to appreciate the rationale underlying the 'recommendation' made by the analytical procedure, and hence feels themselves to be incapable of second-guessing the system.
With what were once called 'third-generation' development tools, the rationale was evident in the form of an algorithm or procedure, which may have been explicitly documented externally to the software, but was at least extractable by any person who had access to the source-code and who had the capacity to read it. The fourth generation of development tools merely expressed the decision-model in a more generally-applicable manner. The advent of the fifth-generation adopted a different approach, however. Rather than a model of the decision, this involved a model of the problem-domain. It became much more difficult to understand how the model (commonly expressed in logic, rules or frames) applied to particular circumstances. Some tools can provide a form of explanation of the rationale (or at least a list of the rules that were fired), but many cannot.
Subsequently, with the sixth generation, an even greater barrier to understanding arose. With a neural network, there is no formal model of a decision or even of a problem-domain. There is just an empirical pile, and a series of inscrutable weightings that have been derived through mathematical processes, and that are then applied to each new instance (Clarke 1991). There have been expressions of concern from many quarters about the delegation of decision-making to software whose behaviour is fundamentally unauditable (e.g. Roszak 1986, Dreyfus 1992, boyd & Crawford 2012).
An examination of the big data literature, reported in Clarke (2014b), concluded that, to date, there is limited evidence of the body of knowledge about data quality and decision quality being applied by the 'big data' movement, either by practitioners or researchers. The research question addressed by this research was accordingly:
What is the impact of the quality of big data, and of big data analytics, on the quality of inferences drawn?
In order to address that question, it is strongly desirable to conduct studies of real-world phenomena. However, conventional empirical research techniques are founded on some key assumptions, most relevantly that:
The big data research domain has a number of characteristics that present serious challenges to the conduct of empirical research. These are summarised in Table 4.
These characteristics present serious theoretical challenges. The selection and application of a body of theory is made difficult because each theory is relevant only to particular entities, relationships and/or environmental circumstances, and the underlying model may map to the reality at one time, but not at another. In addition, inferences drawn from the new context may or may not be relevant to the body of theory that was used to guide the research design. The characteristics also present practical challenges, because there are difficulties in reliably defining populations, population segments, and sampling frames, and in defining what is to be observed and what the terms mean that are used in interviews, questionnaires and case study reports.
Another category of research can be envisaged, which is referred to here as 'quasi-empirical'. This draws on the technique of scenario analysis that has been used in futurology and long-term strategic planning since the 1960s, and is well-described in Wack (1985), Leemhuis (1985), Mobasheri et al. (1989), Schwarz (1991) and van der Heijden (1996).
A scenario is a story-line that represents a composite or 'imagined but realistic world'. The prefix 'quasi' is an appropriate qualifier, because it is from the Latin, in use in English since the 15th century, meaning 'resembling', or 'seemingly but not actually'. Scenario analysis involves the preparation of a small set of scenarios that have the potential to provide insights into an emergent new context.
The Quasi-Empirical Scenario Analysis (QuESA) technique proposed here involves a set of scenarios within which analytics are applied to big data collections. Each scenario is empirical to the extent that it commences with a setting-description that is drawn from contemporary phenomena. A story is then developed from the elements within the setting, by overlaying further data that is surmised, inferred or postulated, but is plausible or at least tenable, and by applying social and economic process sequences that are commonly observed in comparable settings.
The purpose of the QuESA technique is emphatically not to generate assertions of predictive, explanatory or even descriptive standing. The intention is to assist in the emergence of insights, and in the formulation of hypotheses for testing by means of conventional empirical techniques as phenomena emerge, stabilise and become observable. In some respects, the QuESA technique might be compared with focus group research, in that the purpose of both is not to formally test hypotheses, but to gain insights and to bring to the surface potential hypotheses.
In the QuESA technique, each scenario is built from a factual base, utilises real or at least realistic story elements, and is presented using narrative logic, i.e. plausible interactions and sequences. A single scenario is too limiting, firstly because it cannot capture sufficient of the richness of the research domain, and secondly because of the risk that the researcher or readers of reports on the research will lapse into predictive thinking rather than recognising the technique's purposes and limitations. Hence multiple scenarios are developed, addressing a requisite diversity of contexts, postulating events, interactions and sequences that are plausible within the particular context associated with the factual base.
This section describes the manner in which the QuESA technique was applied to big data and big data analytics, and presents the findings that resulted from the work.
A range of contexts was identified from the big data literature. In most cases, reports exist of applications of the specific kind referred to in the scenario, although in a few cases the application may be at this stage only aspirational. An endeavour was made to achieve diversity in the settings, in the nature of the data, in the nature of the analytical tools applied to it, and in the type of function being performed. It is infeasible to attempt representativeness in the sample, because the population is as-yet ill-defined, and indeed the dimensions across which the population varies are still under investigation.
The text for the seven Scenarios is in Appendix 1. Some Scenarios involve data that identifies or may identify specific individuals. This is the case with (2) Creditworthiness, (4) Foster Parenting, (6) Fraud Detection and (7) Insider Detection. In (3) Ad Targeting, the data relates to online identities that may but may not relate to a particular human being. One instance depends on de-identified patient data - (5) Cancer Treatment; whereas in Scenario (1) Precipitation Events, the data relates to environmental phenomena.
In each case, starting with the factual base, various events or trajectories were postulated, which appeared to be tenable for that particular context. A narrative was then developed, and a story-line written. For example, Scenario (5) Cancer Treatment is based on a widely-quoted case in Mayer-Schonberger & Cukier (2013). An impact is postulated on research funding policies, which is a natural corollary of such a discovery. Because of the tight linkage between medical science research and industry, an associated development in the pharmaceutical industry is postulated. The impact on the population of young researchers is a natural demographic consequence. It is commonly the case in empirical research that subsequent work establishes that the correlations that were initially discovered were not as they seemed, and that the inferences drawn from the correlations need substantial qualification; and it is surmised that just such an eventuality will occur in this context as well.
The inferences that are drawn from Quasi-Empirical Scenario Analysis are by definition not empirically based, but are 'insights' that arise from a plausible story-line, not from observation of the real world or of some more or less carefully-controlled proxy for the real world. The 'findings' reported in this section are accordingly hypotheses. The credibility of the hypotheses is weaker than that which arises from the application of a demonstrably relevant theory to a research domain - which are of course to be preferred if such a theories are available. On the other hand, the credibility of the hypotheses is considerably greater than that of ad hoc propositions and hunches, and greater than ideas generated on the basis of anecdotes alone.
The first data quality issue discussed above related to the analysis of big data that was gathered against varying data scales. Challenges of these kinds are evident in Scenarios (3) Ad Targeting, (6) Fraud Detection and (7) Insider Detection. For example, data about online identities includes nominal data (interests), binary data although possibly with some data missing (gender), ordinal data (age-group), ratio-scale data (transaction counts and proportions), and longitudinal data. Few analytical techniques that are supported by mathematical statistics theory are available to cope with such mixed-mode data.
The relationship between purpose of collection and the investment in original data quality was identified as a factor that may undermine the quality of inferences drawn. Administrative actions in relation to fraud, and particularly prosecutions, are undermined by poor-quality evidence. In some contexts, commercial liability might arise, e.g. Scenario (2) Creditworthiness, while in others a duty of care may be breached, e.g. Scenario (4) Foster Parenting.
Loss of quality through inappropriate or inconsistent data compression, and through inappropriate or inconsistent handling of missing data over a timeline, arise in Scenario (1) Precipitation Events.
Various uncertainties arise about identity, and hence about whether the data from different sources, and data collected over time, actually relate to the same real-world entity. This is significant in big health data contexts generally, e.g. Scenario (5) Cancer Treatment, and big social data applications, e.g. Scenario (4) Foster Parenting. The issues also loom large in Scenario (3) Ad Targeting, and - given the prevalence of identity fraud - in Scenario (2) Creditworthiness.
Public demands for anonymisation/deidentification, and active endeavours to obfuscate and falsify identity, can be reasonably expected to exacerbate these challenges. Retrospective studies across long periods, as in Scenario (1) Precipitation Events, (4) Foster Parenting and (6) Fraud Detection, are confronted by problems arising from multiple sources with inconsistent or unclear syntax and semantics.
The vagaries of 'data cleansing' techniques ensure that there will be instances of worsened data quality, resulting in both wrong inferences and spurious matches. This is a problem shared by all big data projects that depend on data consolidated from multiple sources. It afflicts Scenarios as diverse as (1) Precipitation, (5) Cancer Treatment and (7) Insider Detection.
Of the decision quality issues, dubious relevance appears as an issue in (2) Creditworthiness, (5) Cancer Treatment, (6) Fraud Detection and (7) Insider Detection. In Scenario (6) Fraud Detection, data meaning is an issue, in that suspects may be unjustifiably identified as a result of inconsistencies arising from attempts to deceive, and from semantic issues even within a single database, let alone within a consolidation of multiple, inherently incompatible databases. Meanwhile, transparency of the decision mechanism results in the misallocation of resources in Scenario (5) Cancer Treatment, and in unfair discrimination in (2) Creditworthiness and (7) Insider Detection.
Many applications of big data and big data analytics are not currently able to be studied because the organisations conducting them do not provide ready access. This paper has introduced and applied a new research technique, dubbed QuESA. Its purpose is to enable research to be undertaken in circumstances in which the gathering of the data necessary to support empirical research is not possible, not practicable, or not economic.
The scenarios that have been used in this research are not case studies of specific instances of big data at work. They are story-lines, devised in order to encompass a range of issues not all of which are likely to arise in each particular real-life application. Each plot-line builds on factual foundations, then infiltrates into the stories additional elements that are plausible in the particular context. The intention of the scenarios was to test the assumptions underlying the big data value-proposition, not to pretend to be a substitute for deep case studies of actual experience.
The big data movement involves data-collections that are of uncertain original quality, and lower current quality, and that have uncertain associations with real-world entities. Data collections are combined, by means of unclear validity, and modified by unaudited means, in order to achieve consolidated digital personae that have uncertain validity, and have uncertain relationships with any particular real-world entity. To this melange, powerful analytical tools are then applied. Theoretical analysis, complemented by the findings from the QuESA study, identifies a wide range of circumstances in which these problems can arise.
Big data proponents claim that 'more trumps better'. More specifically: "With big data, the sum is more valuable than its parts, and when we recombine the sums of multiple datasets together, that sum too is worth more ... [W]e no longer need to worry so much about individual data points biasing the overall analysis" (Mayer-Schonberger & Cukier 2013, pp. 108, 40). The over-claiming by those authors extends beyond the general (where, in some circumstances, the much-overstated 'law of large numbers' actually does apply) to the specific: "Big data gives us an especially clear view of the granular" (p.13). As summarised by Leonelli (2014, p.3): "Big Data is viewed, through its mere existence, as countering the risk of bias in data collection and interpretation". These assertions lack adequate support by either theoretical argument or empirical evidence. In the face of the analysis reported on in this paper, they must be regarded as being at best wishful thinking, or self-delusional, or - because they are so often framed as recommendations to executives - reckless or even fraudulent.
Given the uncertain quality of data and of decision processes, it appears that many inferences from big data projects may currently be being accorded greater credibility than they actually warrant. If that is the case, then resources will be misallocated. Within corporations, the impact will ultimately be felt in lower return on investment, whereas in public sector contexts there will be negative impacts on public policy outcomes, such as unjustified discrimination against particular population segments.
When big data analytics are inappropriately applied to population inferencing and profile-construction, the harm that can arise includes resource misallocation and unjustified discrimination. When, on the other hand, inappropriate inferencing is about specific individuals, the costs are inevitably borne not by the organisation(s) involved but by the individuals, sometimes in the form of inconvenience, but sometimes with financial, service-availability, psychological, discrimination, or natural justice dimensions.
When profiles generated by big data analytics are applied, in order to generate suspects, the result is an obscure and perhaps counter-intuitive "predetermined characterisation or model of infraction" (Marx & Reichman 1984, p.429), based not on 'probable cause', but on a merely 'probabilistic cause' (Bollier 2010, pp.33-34). Not only does this result in unjustified impositions on the individuals concerned, but it also denies them natural justice because the lack of transparency relating to data and decision criteria means that the accusations are mysterious and even undefendable and they cannot get a fair hearing.
The computer science and management literatures are remarkably lacking in discussion of the issues and the impacts examined in this paper. As a result, they have yet to mature into a literature on appropriate business processes for acquiring and consolidating big data, and applying big data analytics. The information systems literature could be expected to be more sceptical, and more helpful to decision-makers in business and government. Yet, as at the end of 2014, the entire AIS electronic library contained but one paper whose Abstract included both 'big data' and either 'risk assessment' or 'risk management'. The results of the research reported in this paper, limited though it was to an analytical and quasi-empirical base, make clear that much more care is warranted.
Historical rainfall data has been gathered from many sources, across an extended period, and across a range of geographical locations. The collectors, some of them professional but mostly amateurs, used highly diverse collection methods, with little calibration and few controls. The data is consolidated into a single collection. A considerable amount of data manipulation is necessary, including the interpolation of data for empty cells, the arbitrary disaggregation of long-period data into the desirable shorter periods, and, in some cases, arbitrary disaggregation and reaggregation of data (e.g. to reconcile midnight-to-midnight with dusk-to-dusk recording times).
Attempts are made to conduct quality audits against such sources as contemporaneous newspaper reports. However, these prove to be too slow and expensive, and are curtailed. Analytical techniques are applied to the data. Conclusions are reached about historical fluctations and long-term trends, with appropriate qualifications expressed. Analysts then ignore the qualifications and apply the data as though it were factual rather than a mix of facts and interpolations. Climate-change sceptics point to the serious inadequacies in the database, and argue that climate-change proponents, in conducting their crusade, have played fast and loose with scientific principles.
A financial services provider combines its transactions database, its records of chargebacks arising from fraudulent transactions, and government statistics regarding the geographical distribution of income and wealth. It draws inferences about the risks that its cardholders create for the company. It uses those inferences in its decision-making about individual customers, including credit-limits and the issue of replacement and upgraded cards.
Although not publicised by the company, this gradually becomes widely known, and results in negative media comments and recriminations on social media. Questions are raised about whether it conflicts with 'redlining' provisions in various laws. Discrimination against individuals based on the behaviour of other customers of merchants that they use is argued to be at least immoral, and possibly illegal, but certainly illogical from an individual consumer's perspective. The lender reviews the benefits arising from the technique, the harm done to its reputation, and the trade-off between the two.
A social media services provider accumulates a vast amount of social transaction data, and some economic transaction data, through activity on its own web-sites and those of strategic partners. It applies complex data analytics techniques to this data to infer attributes of individual digital personae. Based on the inferred attributes of online identities and the characteristics of the available materials, the service-provider allocates third-party ads and its own promotional materials to the available space on web-pages.
The 'brute force' nature of the data consolidation and analysis means that no account is taken of the incidence of partial identities, conflated identities, and obfuscated and falsified profiles. This results in mis-placement of a significant proportion of ads, to the detriment mostly of advertisers, but to some extent also of individual consumers. It is challenging to conduct audits of ad-targeting effectiveness, and hence advertisers remain unaware of the low quality of the data and of the inferences. The nature of the data exploitation achieves a considerably higher level of public consciousness as a result of the increasing incidence of inappropriate content appearing on childrens' screens.
A government agency responsible for social welfare programs consolidates data from foster-care and unemployment benefits databases, and discovers a correlation between having multiple foster parents and later being chronically unemployed. On the basis of this correlation, it draws the inference that the longstanding practice of moving children along a chain of foster-parents should be discontinued. It accordingly issues new policy directives to its case managers.
Because such processes lack transparency, and foster-children are young and largely without a voice, the new policy remains 'under the radar' for some time. Massive resistance then builds from social welfare NGOs, as it becomes apparent that children are being forced to stay with foster-parents who they are fundamentally incompatible with, and that accusations of abuse are being downplayed because of the forcefulness of policy directives based on mysterious 'big data analytics'.
Millions of electronic medical records reveal that cancer sufferers who take a certain combination of aspirin and orange juice see their disease go into remission. Research funding agencies are excited by this development, and transfer resources to 'big health data analytics' and away from traditional systemic research into causes, pathways and treatments of disease.
Pharmaceutical companies follow the trend by purchasing homeopathic suppliers and patenting herb genes. The number of doctoral and post-doctoral positions available in medical science drops sharply. After 5 years, enough data has become available for the conclusion to be reached that the health treatments 'recommended' by these methods are ineffectual. A latter-day prophet emerges who decries 'the flight from reason', fashion shifts back to laboratory rather than digital research, and medical researchers slowly regain their previous high standing. The loss of momentum is estimated to have delayed progress by 10-15 years and generated a shortage of trained medical scientists.
A company that has large sums of money flushing through its hands is under pressure from regulators, knows that stock exchanges run real-time fraud detection schemes, and accepts at face value the upbeat claims made by the proponents of big data analytics. It combines fraud-detection heuristics with inferences drawn from its large transaction database, and generates suspects. It assigns its own limited internal investigation resources to these suspects, and refers some of them to law enforcement agencies.
The large majority of the cases investigated internally are found to be spurious. Little is heard back from law enforcement agencies. Some of the suspects discover that they are being investigated, and threaten to take their business elsewhere and to initiate defamation actions. The investigators return to their tried-and-true methods of locating and prioritising suspicious cases.
A government agency receives terse instructions from the government to get out ahead of the whistleblower menace, with Macbeth, Brutus, Iago, Judas Iscariot, Manning and Snowden invoked as examples of trusted insiders who turned. The agency increases the intrusiveness and frequency of employee vetting, and lowers the threshold at which positive vetting is undertaken. It applies big data analytics to a consolidated database comprising all internal communications, and all postings to social media gathered by a specialist external services corporation. To increase the pool of available information, it exercises powers to gain access to border movements, credit history, court records, law enforcement agencies' persons-of-interest lists, and financial tracking alerts.
The primary effect of these measures is to further reduce employee loyalty to the organisation. To the extent that productivity is measurable, it sags. The false positives arising from data analytics explode, because of the leap in negative sentiments expressed on internal networks and in social media, and in the vituperative langauge the postings contain. The false positives greatly increase the size of the haystack, making the presumed needles even harder to find. The poisonous atmosphere increases the opportunities for a vindictive insider to obfuscate their activities and even to find willing collaborators. Eventually cool heads prevail, by pointing out how few individuals ever actually leak information without authority. The wave of over-reaction slowly subsides, leaving a bruised and demotivated workforce with a bad taste in its mouth.
[It is intended that this Annex be deleted from the final version of the paper, but posted on the author's web-site, with a link from the paper to it, to facilitate checking by readers of the research method and findings.]
Elements of the scenarios are readily found throughout the formal literatures and media reports. However, the immediate inspiration for each of the Scenarios is outlined below:
This is based on personal experience sourcing and analysing rainfall data relating to the author's conservation property, combined with the climate science debates of the last decade.
See, for example, the boldness of this paper:
Reek T., Doty S.R. & Owen T.W. (1992) 'A Deterministic Approach to the Validation of Historical Daily Temperature and Precipitation Data from the Cooperative Network' Bull. Amer. Meteor. Soc. 73, 6 (June 1992) 753-762, at http://journals.ametsoc.org/doi/pdf/10.1175/1520-0477%281992%29073%3C0753%3AADATV%3E2.0.CO%3B2
This issue is the subject of occasional media reports, including the one cited below. Because the story-line is mostly already documented, the only aspects that have been extrapolated are the generalisation to jurisdictions that have broad consumer protection and anti-discrimination legislation, and the impact of morally dubious behaviour on corporate reputation and brand value.
Cuomo C. et al. (2009) ''GMA' Gets Answers: Some Credit Card Companies Financially Profiling Customers' ABC News, 28 January 2009, at http://abcnews.go.com/GMA/TheLaw/gma-answers-credit-card-companies-financially-pofiling-customers/story?id=6747461
This is a straightforward extrapolation from the large volume of material published in recent years about 'behavioural targeting', 'web analytics' and 'interest-based targeting'. One aspect of difference is that it reflects both the consumer marketing literature and the grey literature on how people can combat web analytics.
See, for example:
O'Reilly (2012) 'Big Data Now' O'Reilly Media, 2012, pp. 56-59, at http://www.oreilly.com/data/free/big-data-now-2012.csp
Park S.-H., Huh S.-Y., Oh W. & Han A.P. (2012) 'A Social Network-Based Inference Model for Validating Customer Profile Data' MIS Quarterly 36, 4 (December 2012) 1217-1237, at http://www.is.cityu.edu.hk/staff/sangphan/mobility/papers/BI_MISQ_%5BHAN,%2520Sag%2520Pil%5D.pdf
Chen J. & Stallaert J. (2014) 'An Economic Analysis of Online Advertising Using Behavioral Targeting' MIS Quarterly 38, 2 (June 2014) 429-449, at http://www.utdallas.edu/~chenjq/Doc/Behavioral%20Targeting.pdf
Kissell J. (2014) 'Take Control of Your Online Privacy' AgileBits, March 2014, at http://email.agilebits.com/t/r-l-ckltdlt-kjiuxtlh-t/
Zhong P. (2014) 'Prism-Break' Prism-Break.org, 2014, at http://prism-break.org/en/http://prism-break.org/en/
This derives from correspondence with a colleague about findings from data analytics reported by a government agency. In mid-2014, the approach may have still been aspirational rather than a delivered project. The attraction is that it is a simple application of the hype being projected by consultants and over-enthusiastic academics, and is in a social rather than a health setting.
The lack of transparency, and the response by NGOs as the method becomes known, are postulated based on the author's experience of the trajectories of a range of similar issues.
The first sentence is taken directly from a passage on p. 14 of a well-known, strongly-promoted and highly-cited book. The notion of research funding agencies shifting significant amounts across to digital research was originally an invention by the author, but was promptly validated by the announcement of Australia's second and third Big Data Research Centres during the same month in which the Scenario was drafted.
The response by pharma companies is pure fiction, but tenable because of the prospect of earning revenue at relatively low cost. The impact on the pool of medical science researchers is an orthodox demographic projection, based on the fact that a reduction in trainee positions is the quickest way in which a research institution can cut its costs when faced with a decline in funds.
Mayer-Schonberger V. & Cukier K. (2013) 'Big Data: A Revolution That Will Transform How We Live, Work and Think' John Murray, 2013
UniSA to lead new $88 million CRC in big data (21 Feb 2014)
UNSW's new Centre for Big Data in Health (12 Sep 2014)
A $20 million Australian Research Council (ARC) Centre of Excellence lfor Mathematical and Statistical Frontiers of Big Data, Big Models, New Insights (19 Sep 2014)
This is an extrapolation from existing approaches in stock exchange contexts, overlaid with 'big data' concepts. The first half is mainstream practice, and it has already been crept across many sectors as a result of anti-money laundering and counter-terrorism laws.
The high rate of spurious hits is speculation, based on the remarkably few success stories that have been reported from decades of fraud detection schemes. The idea of active responses by individuals is an invention. However, these techniques have in the past been focussed primarily on welfare recipients. If the big data movement is successful, it has to be expected that the technique will also be applied to people of influence, affluence and power. Those honest enough, and those brazen enough, can be reasonably expected to fight back.
The initial, factual base of this Scenario is established practice in many jurisdictions. The adaptation draws on an upgrade to staff security assessment processes announced in September 2014 by the Australian Attorney-General.
The remainder is speculative, but reflects the generally low loyalty among government employees, the intemperateness of social media postings, the rarity of whistleblowing, and the rarity with which whistleblowers are actually discovered.
AGD (2014) 'Australian Government Personnel Security Core Policy' Attorney-General's Dept, 2 September 2014, at http://www.protectivesecurity.gov.au/personnelsecurity/Pages/default.aspx
Coyne A. (2014) 'Brandis boosts vetting of APS staff to prevent insider threats' itNews, 2 September 2014, at http://www.itnews.com.au/News/391656,brandis-boosts-vetting-of-aps-staff-to-prevnt-insider-threats.aspx
Anderson C. (2008) 'The End of Theory: The Data Deluge Makes the Scientific Method Obsolete' Wired Magazine 16:07, 23 June 2008, at http://archive.wired.com/science/discoveries/magazine/16-07/pb_theory
Bollier D. (2010) 'The Promise and Peril of Big Data' The Aspen Institute, 2010, at http://www.ilmresource.com/collateral/analyst-reports/10334-ar-promise-peril-of-big-data.pdf
boyd D. & Crawford K. (2012) 'Critical Questions for Big Data' Information, Communication & Society, 15, 5 (June 2012) 662-679, DOI: 10.1080/1369118X.2012.678878, at http://www.tandfonline.com/doi/abs/10.1080/1369118X.2012.678878#.U_0X7kaLA4M
Clarke R. (1991) 'A Contingency Approach to the Software Generations' Database 22, 3 (Summer 1991) 23 - 34, PrePrint at http://www.rogerclarke.com/SOS/SwareGenns.html
Clarke R. (1993) 'Profiling: A Hidden Challenge to the Regulation of Data Surveillance' Journal of Law and Information Science 4,2 (December 1993), PrePrint at http://www.rogerclarke.com/DV/PaperProfiling.html
Clarke R. (1994) 'The Digital Persona and Its Application to Data Surveillance' The Information Society 10,2 (June 1994), at http://www.rogerclarke.com/DV/DigPersona.html
Clarke R. (1995a) 'Computer Matching by Government Agencies: The Failure of Cost/Benefit Analysis as a Control Mechanism' Information Infrastructure & Policy 4,1 (March 1995), PrePrint at http://www.rogerclarke.com/DV/MatchCBA.html
Clarke R. (1995b) 'A Normative Regulatory Framework for Computer Matching' J. of Computer & Info. L. 13,3 (June 1995), PrePrint at http://www.rogerclarke.com/DV/MatchFrame.html
Clarke R. (2014a) 'Promise Unfulfilled: The Digital Persona Concept, Two Decades Later' Information Technology & People 27, 2 (Jun 2014) 182-207, PrePrint at http://www.rogerclarke.com/ID/DP12.html
Clarke R. (2014b) 'Quality Factors in Big Data and Big Data Analytics' Working Paper, Xamax Consultancy Pty Ltd, September 2014, at http://www.rogerclarke.com/EC/BDQF.html
DHHS (2012) 'Guidance Regarding Methods for De-identification of Protected Health Information in Accordance with the Health Insurance Portability and Accountability Act (HIPAA) Privacy Rule' Department of Health & Human Services, November 2012, at http://www.hhs.gov/ocr/privacy/hipaa/understanding/coveredentities/De-identification/guidance.html
Dreyfus H.L. (1992) 'What Computers Still Can't Do: A Critique of Artificial Reason' MIT Press, 1992
English L.P. (2006) 'To a High IQ! Information Content Quality: Assessing the Quality of the Information Product' IDQ Newsletter 2, 3, July 2006, at http://iaidq.org/publications/doc2/english-2006-07.shtml
Fayyad U., Piatetsky-Shapiro G. & Smyth P. (1996) 'From Data Mining to Knowledge Discovery in Databases ' AI Magazine 17, 3 (1996) 37-54, at http://aaaipress.org/ojs/index.php/aimagazine/article/download/1230/1131..
Guo P. (2013) 'Data Science Workflow: Overview and Challenges' ACM Blog, 30 October 2013, at http://cacm.acm.org/blogs/blog-cacm/169199-data-science-workflow-overview-and-challenges/fulltext
Hall M., Frank E., Holmes G., Pfahringer B., Reutemann P. & Witten I.H. (2009) 'The WEKA Data Mining Software: An Update' SIGKDD Explorations 11, 1 (2009), at http://www.sigkdd.org/sites/default/files/issues/11-1-2009-07/p2V11n1.pdf
van der Heijden K. (1996) 'Scenarios: The Art of Stategic Conversation' Wiley, 1996
Huh Y.U., Keller F.R., Redman T.C. & Watkins A.R. (1990) 'Data Quality' Information and Software Technology 32, 8 (1990) 559-565
Inmon B. (1992) 'Building the Data Warehouse' Wiley, 1992
Jacobs A. (2009) 'The Pathologies of Big Data' Communications of the ACM 52, 8 (August 2009) 36-44
Jagadish H.V., Gehrke J., Labrinidis A., Papakonstantinou Y., Patel J.M., Ramakrishnan R. & Shahabi C. (2014) 'Big data and its technical challenges' Communications of the ACM 57, 7 (July 2014) 86-94
Kimball R. (1996) 'The Data Warehouse Toolkit' Wiley, 1996
Leemhuis J.P. (1985) 'Using scenarios to develop strategies' Long Range Planning 18 (1985) 30-37
Leonelli S. (2014) 'What difference does quantity make? On the epistemology of Big Data in biology ' Big Data & Society (April-June 2014) 1-11, at http://bds.sagepub.com/content/1/1/2053951714534395.full.pdf+html
Marx G.T. & Reichman N. (1984) 'Routinising the Discovery of Secrets' Am. Behav. Scientist 27,4 (Mar/Apr 1984) 423-452
Mayer-Schonberger V. & Cukier K. (2013) 'Big Data: A Revolution That Will Transform How We Live, Work and Think' John Murray, 2013
Mobasheri F., Orren L.H. & Sjoshansi F.P. (1989) 'Scenario Planning at Southern California Edison' Interfaces (September/October 1989) 31-44
Ngai E.W.T., Xiu L. & Chau D.C.K. (2009) 'Application of data mining techniques in customer relationship management: A literature review and classification' Expert Systems with Applications, 36, 2 (2009) 2592-2602
OECD (1980) 'Guidelines on the Protection of Privacy and Transborder Flows of Personal Data' OECD, Paris, 1980, mirrored at http://www.rogerclarke.com/DV/OECDPs.html
OECD (2013) 'Exploring Data-Driven Innovation as a New Source of Growth: Mapping the Policy Issues Raised by "Big Data"' OECD Digital Economy Papers, No. 222, OECD Publishing, at http://dx.doi.org/10.1787/5k47zw3fcp43-en
Piprani B. & Ernst D. (2008) 'A Model for Data Quality Assessment' Proc. OTM Workshops (5333) 2008, pp 750-759
Ratner B. (2003) 'Statistical Modeling and Analysis for Database Marketing: Effective Techniques for Mining Big Data' CRC Press, June 2003
Roszak T. (1986) 'The Cult of Information' Pantheon 1986
Rusbridge C., Burnhill P., Seamus R., Buneman P., Giaretta D., Lyon L. & Atkinson M. (2005) 'The Digital Curation Centre: A Vision for Digital Curation' Proc. Conf. From Local to Global: Data Interoperability--Challenges and Technologies, Sardinia, 2005, pp. 1-11, at http://eprints.erpanet.org/archive/00000082/01/DCC_Vision.pdf
Saha B. & Srivastava D. (2014) 'Data Quality: The other Face of Big Data', Proc. ICDE Conf., March-April 2014, pp. 1294 - 1297
Schwartz P. (1991) 'The Art of the Long View: Planning for the Future in an Uncertain World' Doubleday, 1991
UKICO (2012) 'Anonymisation: managing data protection risk: code of practice' Information Commissioners Office, November 2012, at http://ico.org.uk/for_organisations/data_protection/topic_guides/~/media/documents/library/Data_Protection/Practical_application/anonymisation-codev2.pdf
Wack P. (1985) 'Scenarios: Uncharted Waters Ahead' Harv. Bus. Rev. 63, 5 (September-October 1985) 73-89
Wang R.Y. & Strong D.M. (1996) 'Beyond Accuracy: What Data Quality Means to Data Consumers' Journal of Management Information Systems 12, 4 (Spring, 1996) 5-33
This version has benefited from valuable feedback from Kasia Bail, Lyria Bennett Moses, Russell Clarke, David Vaile and Graham Greenleaf. Aspects were presented to a Symposium on Legal Implications of Big Data, run by Prof. Dan Svantesson of Bond University, in Sydney, on 12 December 2014.
Roger Clarke is Principal of Xamax Consultancy Pty Ltd, Canberra. He is also a Visiting Professor in the Cyberspace Law & Policy Centre at the University of N.S.W., and a Visiting Professor in the Research School of Computer Science at the Australian National University.
The content and infrastructure for these community service pages are provided by Roger Clarke through his consultancy company, Xamax.
From the site's beginnings in August 1994 until February 2009, the infrastructure was provided by the Australian National University. During that time, the site accumulated close to 30 million hits. It passed 50 million in early 2015.
Sponsored by Bunhybee Grasslands, the extended Clarke Family, Knights of the Spatchcock and their drummer
Xamax Consultancy Pty Ltd
ACN: 002 360 456
78 Sidaway St, Chapman ACT 2611 AUSTRALIA
Tel: +61 2 6288 6916
Created: 31 December 2014 - Last Amended: 27 April 2015 by Roger Clarke - Site Last Verified: 15 February 2009
This document is at www.rogerclarke.com/EC/BDSA.html