data science

Data Science ethics

Curated on December 13, 2015August 3, 2018 by Stefaan Verhulst

Gov.uk blog: “If Tesco knows day-to-day how poorly the nation is, how can Government access similar insights so it can better plan health services? If Airbnb can give you a tailored service depending on your tastes, how can Government provide people with the right support to help them back into work in a way that is right for them? If companies are routinely using social media data to get feedback from their customers to improve their services, how can Government also use publicly available data to do the same?

Data science allows us to use new types of data and powerful tools to analyse this more quickly and more objectively than any human could. It can put us in the vanguard of policymaking – revealing new insights that leads to better and more tailored interventions. And it can help reduce costs, freeing up resource to spend on more serious cases.

But some of these data uses and machine-learning techniques are new and still relatively untested in Government. Of course, we operate within legal frameworks such as the Data Protection Act and Intellectual Property law. These are flexible but don’t always talk explicitly about the new challenges data science throws up. For example, how are you to explain the decision making process of a deep learning black box algorithm? And if you were able to, how would you do so in plain English and not a row of 0s and 1s?

We want data scientists to feel confident to innovate with data, alongside the policy makers and operational staff who make daily decisions on the data that the analysts provide –. That’s why we are creating an ethical framework which brings together the relevant parts of the law and ethical considerations into a simple document that helps Government officials decide what it can do and what it should do. We have a moral responsibility to maximise the use of data – which is never more apparent than after incidents of abuse or crime are left undetected – as well as to pay heed to the potential risks of these new tools. The guidelines are draft and not formal government policy, but we want to share them more widely in order to help iterate and improve them further….

So what’s in the framework? There is more detail in the fuller document, but it is based around six key principles:

Start with a clear user need and public benefit: this will help you justify the level of data sensitivity and method you use
Use the minimum level of data necessary to fulfill the public benefit: there are many techniques for doing so, such as de-identification, aggregation or querying against data
Build robust data science models: the model is only as good as the data it contains and while machines are less biased than humans they can get it wrong. It’s critical to be clear about the confidence of the model and think through unintended consequences and biases contained within the data
Be alert to public perceptions: put simply, what would a normal person on the street think about the project?
Be as open and accountable as possible: Transparency is the antiseptic for unethical behavior. Aim to be as open as possible (with explanations in plain English), although in certain public protection cases the ability to be transparent will be constrained.
Keep data safe and secure: this is not restricted to data science projects but we know that the public are most concerned about losing control of their data….(More)”

Politics and the New Machine

Curated on November 9, 2015August 3, 2018 by Stefaan Verhulst

Jill Lepore in the NewYorker on “What the turn from polls to data science means for democracy”: “…The modern public-opinion poll has been around since the Great Depression, when the response rate—the number of people who take a survey as a percentage of those who were asked—was more than ninety. The participation rate—the number of people who take a survey as a percentage of the population—is far lower. Election pollsters sample only a minuscule portion of the electorate, not uncommonly something on the order of a couple of thousand people out of the more than two hundred million Americans who are eligible to vote. The promise of this work is that the sample is exquisitely representative. But the lower the response rate the harder and more expensive it becomes to realize that promise, which requires both calling many more people and trying to correct for “non-response bias” by giving greater weight to the answers of people from demographic groups that are less likely to respond. Pollster.com’s Mark Blumenthal has recalled how, in the nineteen-eighties, when the response rate at the firm where he was working had fallen to about sixty per cent, people in his office said, “What will happen when it’s only twenty? We won’t be able to be in business!” A typical response rate is now in the single digits.

Meanwhile, polls are wielding greater influence over American elections than ever….

Still, data science can’t solve the biggest problem with polling, because that problem is neither methodological nor technological. It’s political. Pollsters rose to prominence by claiming that measuring public opinion is good for democracy. But what if it’s bad?

A “poll” used to mean the top of your head. Ophelia says of Polonius, “His beard as white as snow: All flaxen was his poll.” When voting involved assembling (all in favor of Smith stand here, all in favor of Jones over there), counting votes required counting heads; that is, counting polls. Eventually, a “poll” came to mean the count itself. By the nineteenth century, to vote was to go “to the polls,” where, more and more, voting was done on paper. Ballots were often printed in newspapers: you’d cut one out and bring it with you. With the turn to the secret ballot, beginning in the eighteen-eighties, the government began supplying the ballots, but newspapers kept printing them; they’d use them to conduct their own polls, called “straw polls.” Before the election, you’d cut out your ballot and mail it to the newspaper, which would make a prediction. Political parties conducted straw polls, too. That’s one of the ways the political machine worked….

Ever since Gallup, two things have been called polls: surveys of opinions and forecasts of election results. (Plenty of other surveys, of course, don’t measure opinions but instead concern status and behavior: Do you own a house? Have you seen a doctor in the past month?) It’s not a bad idea to reserve the term “polls” for the kind meant to produce election forecasts. When Gallup started out, he was skeptical about using a survey to forecast an election: “Such a test is by no means perfect, because a preelection survey must not only measure public opinion in respect to candidates but must also predict just what groups of people will actually take the trouble to cast their ballots.” Also, he didn’t think that predicting elections constituted a public good: “While such forecasts provide an interesting and legitimate activity, they probably serve no great social purpose.” Then why do it? Gallup conducted polls only to prove the accuracy of his surveys, there being no other way to demonstrate it. The polls themselves, he thought, were pointless…

If public-opinion polling is the child of a strained marriage between the press and the academy, data science is the child of a rocky marriage between the academy and Silicon Valley. The term “data science” was coined in 1960, one year after the Democratic National Committee hired Simulmatics Corporation, a company founded by Ithiel de Sola Pool, a political scientist from M.I.T., to provide strategic analysis in advance of the upcoming Presidential election. Pool and his team collected punch cards from pollsters who had archived more than sixty polls from the elections of 1952, 1954, 1956, 1958, and 1960, representing more than a hundred thousand interviews, and fed them into a UNIVAC. They then sorted voters into four hundred and eighty possible types (for example, “Eastern, metropolitan, lower-income, white, Catholic, female Democrat”) and sorted issues into fifty-two clusters (for example, foreign aid). Simulmatics’ first task, completed just before the Democratic National Convention, was a study of “the Negro vote in the North.” Its report, which is thought to have influenced the civil-rights paragraphs added to the Party’s platform, concluded that between 1954 and 1956 “a small but significant shift to the Republicans occurred among Northern Negroes, which cost the Democrats about 1 per cent of the total votes in 8 key states.” After the nominating convention, the D.N.C. commissioned Simulmatics to prepare three more reports, including one that involved running simulations about different ways in which Kennedy might discuss his Catholicism….

Data science may well turn out to be as flawed as public-opinion polling. But a stage in the development of any new tool is to imagine that you’ve perfected it, in order to ponder its consequences. I asked Hilton to suppose that there existed a flawless tool for measuring public opinion, accurately and instantly, a tool available to voters and politicians alike. Imagine that you’re a member of Congress, I said, and you’re about to head into the House to vote on an act—let’s call it the Smeadwell-Nutley Act. As you do, you use an app called iThePublic to learn the opinions of your constituents. You oppose Smeadwell-Nutley; your constituents are seventy-nine per cent in favor of it. Your constituents will instantly know how you’ve voted, and many have set up an account with Crowdpac to make automatic campaign donations. If you vote against the proposed legislation, your constituents will stop giving money to your reëlection campaign. If, contrary to your convictions but in line with your iThePublic, you vote for Smeadwell-Nutley, would that be democracy? …(More)”

When Lobbyists Write Legislation, This Data Mining Tool Traces The Paper Trail

Curated on October 28, 2015July 18, 2019 by Stefaan Verhulst

FastCoExist: “Most kids learn the grade school civics lesson about how a bill becomes a law. What those lessons usually neglect to show is how legislation today is often birthed on a lobbyist’s desk.

But even for expert researchers, journalists, and government transparency groups, tracing a bill’s lineage isn’t easy—especially at the state level. Last year alone, there were 70,000 state bills introduced in 50 states. It would take one person five weeks to even read them all. Groups that do track state legislation usually focus narrowly on a single topic, such as abortion, or perhaps a single lobby groups.

Computers can do much better. A prototype tool, presented in September at Bloomberg’sData for Good Exchange 2015 conference, mines the Sunlight Foundation’s database of more than 500,000 bills and 200,000 resolutions for the 50 states from 2007 to 2015. It also compares them to 1,500 pieces of “model legislation” written by a few lobbying groups that made their work available, such as the conservative group ALEC (American Legislative Exchange Council) and the liberal group the State Innovation Exchange(formerly called ALICE).

The results are interesting. In one example of the program in use, the team—all from the Data Science for Social Good fellowship program in Chicago—created a graphic (above) that presents the relative influence of ALEC and ALICE in different states. The thickness of each line in the graphic correlates to the percentage of bills introduced in each state that are modeled on either group’s legislation. So a relatively liberal state like New York is mostly ALICE bills, while a “swing” state like Illinois has a lot from both groups….

Along with researchers from the University of Chicago, Wikimedia Foundation, Microsoft Research, and Northwestern University, Walsh is also co-author of another paperpresented at the Bloomberg conference shows how data science can increase government transparency.

Walsh and these co-authors developed software that automatically identifies earmarks in U.S. Congressional bills, showing how representatives are benefiting their own states with pork barrel projects. They verified that it works by comparing it to the results of a massive effort from the U.S. Office of Management and Budget to analyze earmarks for a few limited years. Their results, extended back to 1995 in a public database, showed that there may be many more earmarks than anyone thought.

“Governments are making more data available. It’s something like a needle in a haystack problem, trying to extract all that information out,” says Walsh. “Both of these projects are really about shining light to these dark places where we don’t know what’s going on.”

The state legislation tracker data is available for download here, and the team is working on an expanded system that automatically downloads new state legislation so it can stay up to date…(More)”

Big data problems we face today can be traced to the social ordering practices of the 19th century.

Curated on October 19, 2015August 3, 2018 by Stefaan Verhulst

Hamish Robertson and Joanne Travaglia in LSE’s The Impact Blog: “This is not the first ‘big data’ era but the second. The first was the explosion in data collection that occurred from the early 19th century – Hacking’s ‘avalanche of numbers’, precisely situated between 1820 and 1840. This was an analogue big data era, different to our current digital one but characterized by some very similar problems and concerns. Contemporary problems of data analysis and control include a variety of accepted factors that make them ‘big’ and these generally include size, complexity and technology issues. We also suggest that digitisation is a central process in this second big data era, one that seems obvious but which has also appears to have reached a new threshold. Until a decade or so ago ‘big data’ looked just like a digital version of conventional analogue records and systems. Ones whose management had become normalised through statistical and mathematical analysis. Now however we see a level of concern and anxiety, similar to the concerns that were faced in the first big data era.

This situation brings with it a socio-political dimension of interest to us, one in which our understanding of people and our actions on individuals, groups and populations are deeply implicated. The collection of social data had a purpose – understanding and controlling the population in a time of significant social change. To achieve this, new kinds of information and new methods for generating knowledge were required. Many ideas, concepts and categories developed during that first data revolution remain intact today, some uncritically accepted more now than when they were first developed. In this piece we draw out some connections between these two data ‘revolutions’ and the implications for the politics of information in contemporary society. It is clear that many of the problems in this first big data age and, more specifically, their solutions persist down to the present big data era….Our question then is how do we go about re-writing the ideological inheritance of that first data revolution? Can we or will we unpack the ideological sequelae of that past revolution during this present one? The initial indicators are not good in that there is a pervasive assumption in this broad interdisciplinary field that reductive categories are both necessary and natural. Our social ordering practices have influenced our social epistemology. We run the risk in the social sciences of perpetuating the ideological victories of the first data revolution as we progress through the second. The need for critical analysis grows apace not just with the production of each new technique or technology but with the uncritical acceptance of the concepts, categories and assumptions that emerged from that first data revolution. That first data revolution proved to be a successful anti-revolutionary response to the numerous threats to social order posed by the incredible changes of the nineteenth century, rather than the Enlightenment emancipation that was promised. (More)”

This is part of a wider series on the Politics of Data. For more on this topic, also see Mark Carrigan’sPhilosophy of Data Science interview series and the Discover Society special issue on the Politics of Data (Science).

Data Science of the People, for the People, by the People: A Viewpoint on an Emerging Dichotomy

Curated on October 15, 2015August 3, 2018 by Stefaan Verhulst

Paper by Kush R. Varshney: “This paper presents a viewpoint on an emerging dichotomy in data science: applications in which predictions of datadriven algorithms are used to support people in making consequential decisions that can have a profound effect on other people’s lives and applications in which data-driven algorithms act autonomously in settings of low consequence and large scale. An example of the first type of application is prison sentencing and of the second type is selecting news stories to appear on a person’s web portal home page. It is argued that the two types of applications require data, algorithms and models with vastly different properties along several dimensions, including privacy, equitability, robustness, interpretability, causality, and openness. Furthermore, it is argued that the second type of application cannot always be used as a surrogate to develop methods for the first type of application. To contribute to the development of methods for the first type of application, one must really be working on the first type of application….(More)”

Five principles for applying data science for social good

Curated on October 7, 2015July 18, 2019 by Stefaan Verhulst

Jake Porway at O’Reilly: “….Every week, a data or technology company declares that it wants to “do good” and there are countless workshops hosted by major foundations musing on what “big data can do for society.” Add to that a growing number of data-for-good programs from Data Science for Social Good’s fantastic summer program toBayes Impact’s data science fellowships to DrivenData’s data-science-for-good competitions, and you can see how quickly this idea of “data for good” is growing.

Yes, it’s an exciting time to be exploring the ways new datasets, new techniques, and new scientists could be deployed to “make the world a better place.” We’ve already seen deep learning applied to ocean health,satellite imagery used to estimate poverty levels, and cellphone data used to elucidate Nairobi’s hidden public transportation routes. And yet, for all this excitement about the potential of this “data for good movement,” we are still desperately far from creating lasting impact. Many efforts will not only fall short of lasting impact — they will make no change at all….

So how can these well-intentioned efforts reach their full potential for real impact? Embracing the following five principles can drastically accelerate a world in which we truly use data to serve humanity.

1. “Statistics” is so much more than “percentages”

We must convey what constitutes data, what it can be used for, and why it’s valuable.

There was a packed house for the March 2015 release of the No Ceilings Full Participation Report. Hillary Clinton, Melinda Gates, and Chelsea Clinton stood on stage and lauded the report, the culmination of a year-long effort to aggregate and analyze new and existing global data, as the biggest, most comprehensive data collection effort about women and gender ever attempted. One of the most trumpeted parts of the effort was the release of the data in an open and easily accessible way.

I ran home and excitedly pulled up the data from the No Ceilings GitHub, giddy to use it for our DataKind projects. As I downloaded each file, my heart sunk. The 6MB size of the entire global dataset told me what I would find inside before I even opened the first file. Like a familiar ache, the first row of the spreadsheet said it all: “USA, 2009, 84.4%.”

What I’d encountered was a common situation when it comes to data in the social sector: the prevalence of inert, aggregate data. ….

2. Finding problems can be harder than finding solutions

We must scale the process of problem discovery through deeper collaboration between the problem holders, the data holders, and the skills holders.

In the immortal words of Henry Ford, “If I’d asked people what they wanted, they would have said a faster horse.” Right now, the field of data science is in a similar position. Framing data solutions for organizations that don’t realize how much is now possible can be a frustrating search for faster horses. If data cleaning is 80% of the hard work in data science, then problem discovery makes up nearly the remaining 20% when doing data science for good.

The plague here is one of education. …

3. Communication is more important than technology

We must foster environments in which people can speak openly, honestly, and without judgment. We must be constantly curious about each other.

At the conclusion of one of our recent DataKind events, one of our partner nonprofit organizations lined up to hear the results from their volunteer team of data scientists. Everyone was all smiles — the nonprofit leaders had loved the project experience, the data scientists were excited with their results. The presentations began. “We used Amazon RedShift to store the data, which allowed us to quickly build a multinomial regression. The p-value of 0.002 shows …” Eyes glazed over. The nonprofit leaders furrowed their brows in telegraphed concentration. The jargon was standing in the way of understanding the true utility of the project’s findings. It was clear that, like so many other well-intentioned efforts, the project was at risk of gathering dust on a shelf if the team of volunteers couldn’t help the organization understand what they had learned and how it could be integrated into the organization’s ongoing work…..

4. We need diverse viewpoints

To tackle sector-wide challenges, we need a range of voices involved.

One of the most challenging aspects to making change at the sector level is the range of diverse viewpoints necessary to understand a problem in its entirety. In the business world, profit, revenue, or output can be valid metrics of success. Rarely, if ever, are metrics for social change so cleanly defined….

Challenging this paradigm requires diverse, or “collective impact,” approaches to problem solving. The idea has been around for a while (h/t Chris Diehl), but has not yet been widely implemented due to the challenges in successful collective impact. Moreover, while there are many diverse collectives committed to social change, few have the voice of expert data scientists involved. DataKind is piloting a collective impact model called DataKind Labs, that seeks to bring together diverse problem holders, data holders, and data science experts to co-create solutions that can be applied across an entire sector-wide challenge. We just launchedour first project with Microsoft to increase traffic safety and are hopeful that this effort will demonstrate how vital a role data science can play in a collective impact approach.

5. We must design for people

Data is not truth, and tech is not an answer in-and-of-itself. Without designing for the humans on the other end, our work is in vain.

So many of the data projects making headlines — a new app for finding public services, a new probabilistic model for predicting weather patterns for subsistence farmers, a visualization of government spending — are great and interesting accomplishments, but don’t seem to have an end user in mind. The current approach appears to be “get the tech geeks to hack on this problem, and we’ll have cool new solutions!” I’ve opined that, though there are many benefits to hackathons, you can’t just hack your way to social change….(More)”

Addressing Inequality and the ‘Data Divide’

Curated on September 23, 2015August 3, 2018 by Stefaan Verhulst

Daniel Castro at the US Chamber of Commerce Foundation: “In the coming years, communities across the nation will increasingly rely on data to improve quality of life for their residents, such as by improving educational outcomes, reducing healthcare costs, and increasing access to financial services. However, these opportunities require that individuals have access to high-quality data about themselves and their communities. Should certain individuals or communities not routinely have data about them collected, distributed, or used, they may suffer social and economic consequences. Just as the digital divide has held back many communities from reaping the benefits of the modern digital era, a looming “data divide” threatens to stall the benefits of data-driven innovation for a wide swathe of America. Given this risk, policymakers should make a concerted effort to combat data poverty.

Data already plays a crucial role in guiding decision making, and it will only become more important over time. In the private sector, businesses use data for everything from predicting inventory demand to responding to customer feedback to determining where to open new stores. For example, an emerging group of financial service providers use non-traditional data sources, such as an individual’s social network, to assess credit risk and make lending decisions. And health insurers and pharmacies are offering discounts to customers who use fitness trackers to monitor and share data about their health. In the public sector, data is at the heart of important efforts like improving patient safety, cutting government waste, and helping children succeed in school. For example, public health officials in states like Indiana and Maryland have turned to data science in an effort to reduce infant mortality rates.

Many of these exciting advancements are made possible by a new generation of technologies that make it easier to collect, share, and disseminate data. In particular, the Internet of Everything is creating a plethora of always-on devices that record and transmit a wealth of information about our world and the people and objects in it. Individuals are using social media to create a rich tapestry of interactions tied to particular times and places. In addition, government investments in critical data systems, such as statewide databases to track healthcare spending and student performance over time, are integral to efforts to harness data for social good….(More)”

How Africa can benefit from the data revolution

Curated on August 26, 2015October 9, 2018 by Stefaan Verhulst

Neil Lawrence in The Guardian: “….The modern information infrastructure is about movement of data. From data we derive information and knowledge, and that knowledge can be propagated rapidly across the country and throughout the world. Facebook and Google have both made massive investments in machine learning, the mainstay technology for converting data into knowledge. But the potential for these technologies in Africa is much larger: instead of simply advertising products to people, we can imagine modern distributed health systems, distributed markets, knowledge systems for disease intervention. The modern infrastructure should be data driven and deployed across the mobile network. A single good idea can then be rapidly implemented and distributed via the mobile phone app ecosystems.

The information infrastructure does not require large scale thinking and investment to deliver. In fact, it requires just the reverse. It requires agility and innovation. Larger companies cannot react quickly enough to exploit technological advances. Small companies with a good idea can grow quickly. From IBM to Microsoft, Google and now Facebook. All these companies now agree on one thing: data is where the value lies. Modern internet companies are data-driven from the ground up. Could the same thing happen in Africa’s economies? Can entire countries reformulate their infrastructures to be data-driven from the ground up?

Maybe, or maybe not, but it isn’t necessary to have a grand plan to give it a go. It is already natural to use data and communication to solve real world problems. In Silicon Valley these are the challenges of getting a taxi or reserving a restaurant. In Africa they are often more fundamental. John Quinn has been in Kampala, Uganda at Makerere University for eight years now targeting these challenges. In June this year, John and other researchers from across the region came together for Africa’s first workshop on data science at Dedan Kimathi University of Technology. The objective was to spread knowledge of technologies, ideas and solutions. For the modern information infrastructure to be successful software solutions need to be locally generated. African apps to solve African problems. With this in mind the workshop began with a three day summer school on data science which was then followed by two days of talks on challenges in African data science.

The ideas and solutions presented were cutting edge. The Umati project uses social media to understand the use of ethnic hate speech in Kenya (Sidney Ochieng, iHub, Nairobi). The use of social media for monitoring the evolution and effects of Ebola in west Africa (Nuri Pashwani, IBM Research Africa). The Kudusystem for market making in Ugandan farm produce distribution via SMS messages (Kenneth Bwire, Makerere University, Kampala). Telecommunications data for inferring the source and spread of a typhoid outbreak in Kampala (UN Pulse Lab, Kampala). The Punya system for prototyping and deployment of mobile phone apps to deal with emerging crises or market opportunities (Julius Adebayor, MIT) and large scale systems for collating and sharing data resources Open Data Kenya and UN OCHA Human Data Exchange….(More)”

5 Tips for Designing a Data for Good Initiative

Curated on August 18, 2015August 15, 2018 by Stefaan Verhulst

Mitul Desai at Mastercard Center for Inclusive Growth: “The transformative impact of data on development projects, captured in the hashtag #DATARevolution, offers the social and private sectors alike a rallying point to enlist data in the service of high-impact development initiatives.

To help organizations design initiatives that are authentic to their identity and capabilities, we’re sharing what’s necessary to navigate the deeply interconnected organizational, technical and ethical aspects of creating a Data for Good initiative.

1) Define the need

At the center of a Data for Good initiative are the individual beneficiaries you are seeking to serve. This is foundation on which the “Good” of Data for Good rests.

Understanding the data and expertise needed to better serve such individuals will bring into focus the areas where your organization can contribute and the partners you might engage. As we’ve covered in past posts, collaboration between agents who bring different layers of expertise to Data for Good projects is a powerful formula for change….

2) Understand what data can make a difference

Think about what kind of data can tell a story that’s relevant to your mission. Claudia Perlich of Dstillery says: “The question is first and foremost, what decision do I have to make and which data can tell me something about that decision.” This great introduction to what different kinds of data are relevant in different settings can give you concrete examples.

3) Get the right tools for the job

By one estimate, some 90% of business-relevant data are unstructured or semi-structured (think texts, tweets, images, audio) as opposed to structured data like numbers that easily fit into the lines of a spreadsheet. Perlich notes that while it’s more challenging to mine this unstructured data, they can yield especially powerful insights with the right tools—which thankfully aren’t that hard to identify…..

4) Build a case that moves your organization

“While our programs are designed to serve organizations no matter what their capacity, we do find that an organization’s clarity around mission and commitment to using data to drive decision-making are two factors that can make or break a project,” says Jake Porway, founder and executive director of DataKind, a New York-based data science nonprofit that helps organizations develop Data for Good initiatives…..

5) Make technology serve people-centric ethics

The two most critical ethical factors to consider are informed consent and privacy—both require engaging the community you wish to serve as individual actors….

“Employ data-privacy walls, mask the data from the point of collection and encrypt the data you store. Ensure that appropriate technical and organizational safeguards are in place to verify that the data can’t be used to identify individuals or target demographics in a way that could harm them,” recommends Quid’s Pedraza. To understand the technology of data encryption and masking, check out this post. (More)”