What should ‘cultural data analytics’ be?
This is a talk I recently had to give on datafication of culture (or actually on what cultural data analytics a’la Tallinn University could be like). Some of my colleagues thought they would like their students to read this, so posting the text here. Here is the video, and below is the transcript.
https://vimeo.com/bfmuniversity/review/367698058/d509cf6412
I propose discussing cultural data analytics via two broad questions, both of which I have filled with provocations that I hope will allow us to discuss
- the implications and the politics of how we define concepts,
- the power of those definitions shape the disciplinary and methodological space we operate in
- and how that in turn suggests a positive and inclusive vision of cultural data analytics.
I have my own answers to these provocations, but I am hoping that you will have yours, and that the CUDAN team, when assembled, will agree on the shared ones.
The two broad questions are seemingly simple:
How do we define cultural data analytics, given the extensive debates that have surrounded all of the words in this formulation? and
What is it that we want cultural data analytics to be do?
Ok, let’s start from the first big question, what do we mean when we say cultural data analytics. And to be systematic, we need to start with what do we mean, when we say culture.
Culture is according to Raymond Williams “one of the two or three most complicated words in the English language” (Williams 1983, 87). Other scholars are of much the same mind. Some even argue that the term is ‘so overused, that it is better to break it down into its component parts and speak of beliefs, ideas, artifacts, languages, symbols, art, or traditions.
In The Long Revolution, Raymond Williams offered three ways of defining culture:
1 the “ideal” definition, referring to the systems of valuation by means of which groups establish hierarchies, and subsequently judge the worth, of people, places, objects, institutions, and ideas;
2 the “documentary” definition, referring to the whole range of artifacts, both material and immaterial, produced by a group of people;
3 the “social” definition, referring to “a particular” or “whole way of life” i.e., the patterns of thought, conduct, and expression, prevalent among members of a collective.
Relying on the last one, which Williams appropriated from anthropology, John Fiske has argued that for cultural studies culture ‘is neither aesthetic nor humanist in emphasis, but political’. Politics in this case is the practice of living together, and we must be better at it, because, at the risk of sounding melodramatic - the alternative to living together is dying separately.
Methodological implications of how we define culture
Marek Tamm (2016) has suggested in his introduction to the book “How to study culture” that culture is not something that is passively available for researchers to come study it, rather it is constructed in the process of defining and making sense of it. Culture is thus created as an object of study and our definition depends on the disciplinary background of the researcher studying it.
A distinction that has had a strong impact on the study of culture is between culture as practice versus culture as a system of symbols and meanings. The first approach focuses on the processes of meaning making, and perhaps coincides with the definition of culture or cultures as particular ways of life. The second focuses on the more or less stabile forms and codes within the body of what can broadly be called “cultural texts”. Of course, ideally, we want to study culture as both – texts and practices. However, I think keeping this distinction in mind has analytical merit for the discussion at hand, because it highlights not only the methodological, but also the critical or the politico-economic implications that accompany both definitions. Let’s look at these
Critical implications of how we define culture
Culture as a way of life or as a set of everything created by everyone happens - to a disturbing extent - on corporately owned platforms, which are - post what we in my field call the API-apocalypse - closed rather than open for researchers, and make unreliable, difficult partners. They are also, as Jose van Dijck and Tarleton Gillespie (also this) have been saying for about a decade, not neutral intermediaries, but performative and constitutive infrastructures. Social media platforms, but also appstores shape the performance of social acts instead of merely facilitating them. This means that relying on data created and classified by these corporate platforms for making research inferences is quite problematic. Richard Rogers has called this an issue of vanity metrics. The data that corporations create reflects their needs and their version of a way of life, a culture, sociality. Their version is made of likes, follows etc, because those help measure impact and worth within the attention economy that social media has become. It is a partial rendering serving capitalist needs, wherein everyone is a laborer, a consumer or a commodity, often all three at once.
The version of culture as an assemblage of cultural texts, could be seen more as an issue of digitalizing heritage. This brings it its own can of worms, because it basically means participating in the datafication and metadatafication of culture, which as we’ll talk about shortly, is not necessarily a uniformly positive goal. Datafication, is usually conceptualized as the transformation of social action and many other previously unquantifiable aspects of the world into quantified data, which allows real-time tracking and predictive analysis (Mayer- Schoenberger & Cukier, 2013). Datafication, as we’ll shortly discuss in more detail, has a politics.
Basically, how we define culture implicates whether we want to use existing data or create data, which invites a rather different set of methods, and has a rather different set of risks, implications and ethics.
Ok, this brings us to our second word in search of a definition. What is data?
Data is a concept that is most tightly linked to empiricism and positivism. We answer empirical questions by obtaining direct, observable information from the world. That direct observable information, often conceptualized as discrete units of information, is what is called data. Once we verify data, we get, from the positivist perspective - facts.
However, this only seems straightforward. Just like the definition of culture emerges out of and depends on the process of defining it, so is data made and not found. As Geoffry Bowker has famously said: “Raw data is both an oxymoron and a bad idea; to the contrary, data should be cooked with care (2005, p. 183-184). Lisa Gitelman and Virginia Jackson (2013) propose that the seductive power of the term raw data lies in it echoing a presumption that data come before fact, which suggests that data are the starting point for what we know, and that hence data must be transparent or objective.
Data as a thing, data as ideology
My friend and mentor prof. Annette Markham has argued (2016) that in academic discourse data operates on at least two levels – as a thing and as an ideology, both of which obscure the fact that data is not where meaning resides.
She argues that speaking of data as a thing is an ideological stance, which leads us to focus our attention on the wrong part of the process, we focus on what remains after we tidy, clean, condense and simplify and invites us to focus on pieces of text, or outcomes of interaction, distracting us from the point that this is not where meaning resides. Meaning, arguably, resides in the interaction not the outcome of the interaction, it resides in the making and consuming of the text, not in the text itself.
This doesn’t mean that data is useless, or we should not try to make data, it rather means that just as we need to be clear on what culture means for CUDAN, we need to be clear on how we cook data in this project. What tools do we make or use, and how do these tools function as frames or filters. Because tools carry the epistemic traditions they derive from (cf. this by Eef Masson, 2017) and most, if not all of the analytics tools used to study culture today were built by empiricists and positivists. To make this point clearer, I invite us to think about data through metaphors
You have probably all seen and heard a version of “data is the new …”
Cornelius Puschmann and Jean Burgess suggest that there are metaphors of data as a natural force to be controlled (so here there are a lot of oil and water metaphors) and metaphors of data as a resource to be consumed (where they place food and fuel metaphors).
Metaphor scholars (Lakoff and Johnson 1980) have been saying for decades that metaphors function conceptually to not only reflect but to construct our experience of reality. If we say “data is the new oil” the comparison of terms builds or promotes a particular meaning. The term being defined (data) is connected to the supposedly more known term (oil). So if we think of petroleum oil then we think of it having to be drilled, which is dangerous to do, we think that world economies depend on it, that finding it unexpectedly will make you very rich, that you can make anything from it. If the comparison sticks, and everyone starts calling data the new oil, as they kind of have, it will work under the surface not only to reflect, but to influence how we think about data.
As Luke Stark and Anna Lauren Hoffman recently argued, the data metaphors, in particular the oil metaphor, invites specific data practices and specific approaches to data ethics. Liquid metaphors of data lakes, data oceans, data floods and data tsunamis tend to “forestall ethical or regulatory interventions by positioning data as uncontrollable” (Lupton 2013). But Stark and Hoffman also propose that we can look to common data metaphors to solve some of the regulatory and ethical problems we’re having with internet intermediaries abusing our data. If data is a natural resource, then perhaps we need to borrow from the ethical codes of forestry and think of data stewardship. If we think of personal data as of personal digital remains, maybe we need to borrow from morticians or doctors, and think of data care or data fiduciaries - fiduciary duty is the legal obligation of one party to act in the best interest of another.
Again, for the talk at hand, I want to ask – what kind of a data metaphor do we at Tallinn University want to operate with? Should we be satisfied with pre-existing metaphors, and live with what they illuminate and obscure about the world?
Methodological implications of how we define data
Something we hear repeated so often, is that the volume, velocity and variability of ‘big’ data has transformed how social research is conducted. More interestingly, it impacts what we think we are doing when we conduct said research. This, I think is the biggest methodological implication of how we define data. Do we think we’re cooking it? And what do we think it means if we’re cooking it? Some of my colleagues have noted a dangerous erosion of the role and meaning of interpretation in “data-driven” research (Markham 2016). If we agree that there is an erosion, and if we agree that reducing phenomena to data involves classification, which in turn obscures ambiguity and contradiction (Gitleman and Jackson, 2013), and if we think ambiguity and contradiction are important when speaking of and for cultures, then we need to think of how to bring them back. One option is to try to imagine an interpretivist data analytics.
Interpretivism rejects the view that meaning resides within the world independently of how people and groups interpret it. Typically, interpretivists advocate for context, which often means asking people things. This might not always be possible or wise with the types of projects we are imagining for CUDAN. However, the idea of context sensitivity or thick descriptions has been utilized in the more recent discussions on whether we can and should thicken our data.
Latzko-Tith, Bonneau and Millette (2017) say that thickening data means supplementing data with richly textured information, in other words, adding layers to them. Thick data is coated with several layers of rich metadata and paradata, so it is like an onion. Instead of points, thick data are whole little structured worlds. But we can think of thickening data also in terms of being more creative with what counts as data or what kinds of data we have, want, what we discard when we clean it, do we clean all of it, etc. My own experience working with data scraped from the Instagram API and Twitter API have highlighted this on a very personal level. Thickening or layering 90 000 image posts or 25 000 tweets with anything other than the metadata that the platform provides may seem impossible. But computational tools can also show you that the 90 000 images are from 180 accounts, or that in the 25 000 tweets include only 520 heterogenous ones that have been retweeted even once, which makes space for layering based on the computational power of the human brain. Basically, the argument is that layering embodies what interpretivism has learned from hermeneutics, the circular way of working a chunk of data and its context.
The critical (political-economic) implications of how we define data
Now, depending on how CUDAN decides to define data and go about cooking it will situate it at more or less problematic end of the spectrum of what can be called the political-economy of datafication. One of the best questions I heard two weeks ago at the AoIR conference in a methods session, was: “What evil things could be done with these new insights you have generated?” So, I think it is important that we too contemplate what evil things can be done with CUDAN, and what version of the datafication of culture and life we want to contribute to.
Many professionals and scholars see datafication as a revolutionary research opportunity to investigate human conduct. But, datafication is also heavily critiqued. A very poignant recent critique comes from Jathan Sadowski (2019), who recently published an elegant analysis of data as capital (as opposed to data as a commodity, which other work has done).
Sadowski argues that like social and cultural capital, data capital is convertible, in certain conditions, to economic capital. It adds new sources of value and new tools for accumulation. It also currently guarantees that those who already have a lot of this capital, like GAFA (Google, Apple, Facebook, Amazon) or BAT (Baidu, Alibaba, Tencent), will accumulate more, and those who don’t have it, are unlikely to amass any significant amounts of it.
Looking at data as capital allows him to notice that the data imperative, or the drive to accumulate all and any data from all sources, by all means possible, now propels how business is done and how governance is enacted. This means a total datafication of everything, by subjecting previously non-commodified and non-monetized parts of life to the logic of datafication and colonizing new spheres of life or new places in the world, so they can become sites of data extraction. So decisions like buying a company or launching a service are increasingly made for data potential, not because of revenue. Google gives primary school studenst free laptops or invests in healthcare or hosts all of Tallinn University’s emails and documents not because it cares, but because it is already or will very soon profit from all of that data. Extraction of data – and Sadowski is specific about calling it extraction and not collection or even mining, because calling it extraction highlights the exploitative nature of dataveillance, where data is taken without meaningful consent or fair compensation - is a core component of political economy in the 21st century.
Ok so this brings us to the end of the prompts and provocations around definitions, and implores us to ask what we mean when we say ‘cultural data’ and through addressing the methodological implications of defining cultural data – what we mean by cultural data analytics
That the computational processes of sorting and classifying people, places, objects and ideas have profoundly altered the way ‘culture’, as a category of experience, is practiced, experienced and understood, is something that many authors have addressed (Striphas 2015, Andrejevic, Hearn and Kennedy 2015). So, the question is - is there any other way to define cultural data than as the process and outcomes of the datafication of culture. And if there is none, then the question becomes, is there a way to shape how datafication of culture happens or to imagine alternative ways of datafying culture, because what we have now, is consolidation of the work of culture into the hands of a few powerful corporations, which, if we believe Ted Striphas, will lead to “the gradual abandonment of culture’s publicness”
Methodological and critical implications of how we define cultural data
If cultural data is the process and the outcomes of the datafication of culture, which is currently to a large extent governed by corporations for corporate interests, then this invites another question for CUDAN -
Do we need to come up with so called alt metrics for understanding culture? And what would those be?
I mentioned Richard Rogers (2018) work in the beginning of this talk. He proposes metrics that do not build on social media as a vanity space, but as one for social issue work. He calls them critical analytics. We can basically treat the past 15 years of social media as a case study for why we can’t rely on the metadata and the datafication models that corporations have created for their own needs, because analyzing those creates a particularly tilted view of the studied phenomena and makes CUDAN contribute to instead of subvert what is arguably currently wrong with the datafication of culture.
Epistemology of cultural data analytics
This definition work quite neatly introduces a bigger issue, which is what kinds of ontological, epistemological and axiological premises do we want cultural data analytics to have? We’ve talked earlier about bringing a certain interpretivist sensibility to data analytics, at least to our methods of cooking data, but I’m not sure we necessarily want to situate CUDAN fully in interpretivism. We also don’t want to situate it in what Christian Fuchs (2017) calls digital positivism, which he says does not connect “statistical and computational research results to a broader analysis of human meanings, interpretations, experiences, attitudes, moral values, ethical dilemmas, uses, contradictions and macro-sociological implications. And which he says means that it is just what Paul Lazarsfeld called administrative research predominantly concerned with how to make technologies and administration more efficient and effective.”
Instead, I would suggest, and Fuchs suggests, and frankly most authors who have studied social media for many years are suggesting a critical theoretical alternative. What does that mean?
What is it that we want to accomplish?
Ok this finally brings us to the second big question, which is, what do we want cultural data analytics to do? If we want to build critical cultural data analytics, then whatever else we want it to do, we will want it to challenge dominant assumptions and, ideally, change the world towards a better place. No pressure, right?
Looking across various academic, corporate and strategy documents big data analytics and cultural data is imagined to promise the following:
data analytics in general seems to promise to:
help us gain unprecedented insight into stuff –like public opinion, behavior patterns and relationships.
build a more ‘productive and intuitive’ user/consumer experience.
overall, there are a lot of vague but optimistic promises that we can do research that doesn’t exist yet, ask questions that do not exist yet, open up new avenues for inquiry
Within the realm of cultural analytics and digital humanities more broadly, the promises seem to be that we can:
digitally preserve and share cultural heritage. Which:
allows new discoveries that will transform our understanding of our cultures, identities, heritage and history.
make sure these cultures do not disappear;
make sure the heritage industry is relevant in the digital age
allow cultural differences and commonalities to be explored.
shed light on human history and the relationships between cultural and geographic areas.
Help us understand the dissemination of ideas and cultural phenomena and,
in relevant cases (such as in art fairs, universal exhibitions, or Olympic games), improve the management of current events.
introduce data-driven decision-making in the cultural sector (how to do this without adding to the accumulation of privilege and disadvantages, inequality, discrimination etc
provide arguments for the provision and allocation of public funding and measurement of its impact
Frankly, most of these do not sound like critical ambition. Some of these sound outright administrative, many descriptive, some interpretative.
So, again, the question for CUDAN is – which goals do we want to set for our version of cultural data analytics.
Do we want to say that cultural data analytics will help us understand culture better?
Does that mean that we think that the ways in which we understand it now are not good enough? And I am looking at Marek Tamm who has recently edited a whole volume on this. So, you know, provocatively I ask, what’s wrong with those ways of understanding culture? Did you know that training creating just one AI model for natural-language processing can emit as much as 600,000 pounds of carbon dioxide? (Strubell, Ganesh and McCallum 2019
via this). That’s about the same amount produced by 125 roundtrip flights between New York and Beijing. How can we make sure that cultural data analytics is better enough than the more eco-friendly alternatives to be worth it?
Do we want to say that cultural data analytics will be more efficient in understanding culture? That it will create more actionable insights both for researchers and for policy makers? That it will release us from the chains of stepping on the same rake and making the same mistakes?
That, in and of itself, is a great goal. Sheila Jasanoff has suggested that actionable data can problematize the taken-for-granted order of society by pointing to questions or imbalances that can be corrected or rectified, or simply better understood, through systematic compilations of occurrences, frequencies, distributions, or correlations. She speaks specifically of the power of the compilations of climate data, but surely this could be a great asset in the cultural sector as well.
Then again, here too, we can ask what that costs. Another example - AMS, Austria’s employment agency, is about to roll out a sorting algorithm built to increase efficiency. They ran statistical regressions to find out which factors were best at predicting an individual’s chances of finding a job. So they can stop giving support to those who are less likely to find a job. Like women and disabled people. The algorithm increases efficiency and offers highly actionable insights, as it ensures that the agency does not waste resources on giving support to people who will not, in the end, benefit from it. How can we make sure we don’t build this type of efficiency?
What should cultural data analytics be?
Ok so, lets reiterate. I presume that everyone’s answer to what cultural data analytics should be is different, and that is the point of asking these questions and raising these provocations, but let me clarify my take on it and offer some quick examples.
1. I think that while in abstract it makes sense to think of culture as both a practice and a set of texts, it is always also political in emphasis. I also think CUDAN would possibly benefit from a narrower definition of culture, or at least assigning different narrower definitions of culture to specific subprojects. What I’m trying to say is that it is not enough, and perhaps it is even a bad idea to try to combine what is usually called social analytics, i.e. analytics of the trace data cooked on and by GAFA (Google, Amazon, Facbook, Apple) or BAT (Baidu, Alibaba, Tencent) platforms and what is usually called digital humanities, i.e. analytics of digitalized cultural heritage data, and call it cultural data analytics. I don’t think that this is the innovation we’re looking for. My work in social media allows me to see the problematic aspects making inferences of platform data, but it also makes me weary at the ambition to turn cultural heritage data into platform ready data. I think combining these two will keep us stuck in the social media logic that has or will soon colonize all our data, so true innovation lies in coming up with alternatives. This is, of course, easier said than done. If we do want to engage with the “existing” data people generate on the platforms, then I do think that instead of using their data as evidence of practices or ways of life, we should critically analyze infrastructures.
Let me offer an example - Nic Carah and Dan Angus (2018) at University of Queensland are, working on a project that they call “critical simulations”. They engineer and scrutinize how Instagram’s algorithms process, classify and make judgements about cultural life. So they are trying to build the infrastructure to critically analyze it.
2. I think CUDAN needs to be adamant that it is cooking data, and careful in who elses cooking it consumes, as well as who it cooks for, and whom it cooks for for free (and this invites an open data discussion, which I didn’t have tome to go into, but we can in the Q & A). This means that we should set aside resources towards critical tinkering with existing tools, invention of new ones, a reimagination of metrics.
Let me offer another example. Trevor Paglen and Kate Crawford recently organized an artistic intervention called ImageNetRoulette (look here). Image Net is a huge database of photographs that is broadly used to train AI systems in how to recognize, categorize and classify. It is one of the more widely used training sets for machine reading. Among the 14 million images ImageNet was trained on, there were images of people that were sorted manually by humans like Amazon Turkers. They categorized what they saw based on their own biases, and their biases ended up in the algorithm. So while it is easy to imagine a cultural data analytics project that just uses an existing tool to generate some sort of a semi metaphorical rendering of what people represent on social media, ImageNetRoulette was conceived to expose the biases and politics behind the datasets and thus the AI that classifies humans. The project was hugely popular and made its point elegantly. People were labeled in racist, sexist, misogynist and otherwise judgmental terms. And it has already had an impact, the researchers behind ImageNet promised to delete more than half of the 1.2 million “people” images from the dataset.
3. I think CUDAN needs to be ambitious, but profoundly critical in setting goals, to avoid digital positivism, administrative research at the service of efficiency, as well as artsy vanity projects with limited social impact. I think it needs to commit to impact and data justice.
Pitting research questions or interests against each other is problematic, and I am not trying to suggest that everyone needs to study populism, climate change or alternatives to the particular version of capitalism we have, but maybe we should. At least in some way. Is there way to make a project about Estonian heritage cultures to be about the current debates surrounding Estonian forests. Is there a way to simulate and critique and then productively build alternatives to existing infratructures or data logics? In social media research datafication, appification and platformization have become almost curse words, yet in what I have read about the digitalization of cultural heritage, we seem to be hardly able to wait before everything is an app.
I feel like CUDAN has a decision to make. What kind of a project does it want to be. Critical? Descriptive? Computational? Administrative? I don’t think it can be all in equal measures. But I am very excited about the idea of a truly critical, contextual and ethical version of data analytics.
Does it exist? No. Can it be built? I believe so. Maybe this will be CUDANs gift to the world.
