Vincent Chiu is a Technical Program Manager at Google.

Since 2012, Google has supported the study and publication of broadband pricing for researchers, policymakers and the private sector in order to better understand the affordability landscape and help consumers make smarter choices about broadband access. We released the first dataset in August 2012, and periodically refresh the data (May 2013, March 2014, and February 2015). This data has become an integral part of our understanding of global broadband affordability. Harvard's Berkman Center, Facebook, and others actively use the data to understand the broadband landscape. Today, we’re releasing the latest dataset.

Vincent Chiu is a Technical Program Manager at Google.

Since 2012, Google has supported the study and publication of broadband pricing for researchers, policymakers and the private sector in order to better understand the affordability landscape and help consumers make smarter choices about broadband access. We released the first dataset in August 2012, and periodically refresh the data (May 2013, March 2014, and February 2015). This data has become an integral part of our understanding of global broadband affordability. Harvard's Berkman Center, Facebook, and others actively use the data to understand the broadband landscape. Today, we’re releasing the latest dataset.

For the mobile data set, we increased the number of countries represented from 112 to 157, and the number of carriers from 331 to 402. For the fixed-line dataset, we increased the number of countries from 105 to 159, and number of carriers from 331 to 424. This data covers 99.3% of current Internet users globally on a country level.

The collection methodology is designed to capture the cost of data plans. We collected samples from a broad range of light to heavy data usage plans, and recorded numerous individual plan parameters such as downstream bandwidth, monthly cost, and more. Finally, where possible, we collected plans from multiple carriers in each country to get an accurate picture.

Broadband Data:

  • Price observations for fixed broadband plans can be found here.
  • Mobile broadband prices can be found here.

We provide this information to help people understand the state of internet access and make data-driven decisions. Along with this data collection effort, we have analyzed these pricing data and conducted researches on affordability and Internet penetration. Our early results have indicated several topics deserving further discussions within the ICT data community, including metric normalization, income distribution, and broadband value. We believe:

  • Normalization is essential for any meaningful statistical analysis. The diversity of plan types and the complexity of tariff structures surrounding mobile broadband pricing requires a careful analytical methodology for normalization.
  • Income distribution needs to be considered when assessing the broadband affordability situation. The commonly used GNI per capita metric is based on average national income level, which does not consider inequality of income distribution.
  • High broadband value-to-cost ratio is important to drive Internet adoption: beyond just affordability metrics, we need to take into account of broadband experience to define meaningful value-for-the-money metrics.

We look forward to sharing more findings on these topics in 2016. Please stay tuned for updates on our progress. If you have any feedback on the methodology, contact us at broadband-study@google.com

Joshua New is a policy analyst at the Center for Data Innovation. Reposted from CDI's blog.

Bias in big data. Automated discrimination. Algorithms that erode civil liberties.

These are some of the fears that the White House, the Federal Trade Commission, and other critics have expressed about an increasingly data-driven world. But these critics tend to forget that the world is already full of bias, and discrimination permeates human decision-making.

Joshua New is a policy analyst at the Center for Data Innovation. Reposted from CDI's blog.

Bias in big data. Automated discrimination. Algorithms that erode civil liberties.

These are some of the fears that the White House, the Federal Trade Commission, and other critics have expressed about an increasingly data-driven world. But these critics tend to forget that the world is already full of bias, and discrimination permeates human decision-making.

The truth is that the shift to a more data-driven world represents an unparalleled opportunity to crack down on unfair consumer discrimination by using data analysis to expose biases and reduce human prejudice. This opportunity is aptly demonstrated by the Consumer Financial Protection Bureau’s (CFPB) December 2013 auto loan discrimination suit against Ally Financial, the largest such suit in history, in which data and algorithms played a critical role in identifying and combating racial bias.

CFPB found that, from April 2011 to December 2013, Ally Financial had unfairly set higher interest rates on auto loans for 235,000 minority borrowers and ordered the company to pay out $80 million in damages. But the investigation also posed an interesting challenge: Since creditors are generally prohibited from collecting data on an applicant’s race, there was no hard evidence showing Ally had engaged in discriminatory practices. To piece together what really happened, CFPB used an algorithm to infer a borrower’s race based on other information in his or her loan application. Its analysis identified widespread overcharging of minority borrowers as a result of discriminatory interest rate markups at car dealerships.

Ally Financial buys retail installment contracts from more than 12,000 automobile dealers in the United States, essentially allowing dealers to act as middlemen for auto loans. If a consumer decides to finance his or her new car through a dealership rather than a bank, the dealership submits the consumer’s application to a company like Ally. If approved, the consumer pays back the dealership with interest. The interest rate, of course, matters a great deal. To determine what it will be, Ally calculates a “buy rate”—a minimum interest rate for which it is willing to purchase a retail installment contract, as determined by actuarial models. Ally notifies dealerships of this buy rate, but then also gives them substantial leeway to increase the interest rate to make the contract more profitable. Though consumers are free to negotiate these rates and shop around for the best deal, CFPB’s analysis determined that discretionary dealership pricing had a disparate impact on borrowers who were African American, Hispanic, Asian, or Pacific Islanders. On average, they paid between $200 and $300 more than similarly situated white borrowers.

Since creditors cannot inquire about race or ethnicity, Ally’s algorithmically generated buy rates are objective assessments. But when dealerships increase these rates, their judgments are entirely subjective, relying on humans to make decisions that could very well be influenced by racial bias. If dealerships instead took a similar approach to creditors and automated this decision-making process, there would be no opportunity for human bias to enter the equation. While dealerships could still increase interest rates to capture more profits, they could do so based on algorithmic analysis of predefined criteria about a consumer’s willingness to pay, thereby preventing themselves from offering similar consumers different rates based on their race.

Policymakers should guard against the possibility that automated decision-making could perpetuate bias, but with ever-increasing opportunities to collect and analyze data, the public and private sectors also should follow CFPB’s lead and identify new opportunities where data analytics can help expose and reduce human bias. For example, employers could rely on algorithms to select job applicants for interviews based on their objective qualifications rather than relying on human oversight that can be biased against factors such as whether or not the job applicant has an African American–sounding name. And taxi services could rely on algorithms to match drivers with riders rather than leaving it up to drivers who might be inclined to discriminate against passengers based on their race. If policymakers let fear of computerized decision-making impede wider deployment of fair algorithms, then society will lose a valuable opportunity to build a more just world.

Co-authored by Ashley Varady (Program Manager), Christina Perry (Program Director), Lauren Hall (Director of Evaluation), and Bita Nazarian (Executive Director) of 826 Valencia.

In the world of business, terms like performance indicators and productivity are commonplace. It makes sense to want to evaluate progress and the efficacy of a given program, and change course if adequate growth isn't met. Increasingly, these same buzzwords have entered the conversation in the world of education, whereby teacher quality and school funding are connected to standardized test scores and other forms of evaluation. There has been much debate about what constitutes fair and meaningful measures of performance evaluation, for both teachers and students. With the shift to Common Core Standards, these high stakes assessments are moving from multiple choice tests to "authentic tasks" where students communicate their understanding in writing, making writing an essential skill across all disciplines.

Co-authored by Ashley Varady (Program Manager), Christina Perry (Program Director), Lauren Hall (Director of Evaluation), and Bita Nazarian (Executive Director) of 826 Valencia.

In the world of business, terms like performance indicators and productivity are commonplace. It makes sense to want to evaluate progress and the efficacy of a given program, and change course if adequate growth isn't met. Increasingly, these same buzzwords have entered the conversation in the world of education, whereby teacher quality and school funding are connected to standardized test scores and other forms of evaluation. There has been much debate about what constitutes fair and meaningful measures of performance evaluation, for both teachers and students. With the shift to Common Core Standards, these high stakes assessments are moving from multiple choice tests to "authentic tasks" where students communicate their understanding in writing, making writing an essential skill across all disciplines.

The notion that writing is critical to demonstrating understanding has been core to our work at 826 Valencia since our founding in 2002. We work to cultivate these crucial academic skills through rigorous, "real world" writing tasks that are then published in professional books for a much wider audience. In addition to skill building, though, we also invite students to see writing as a resource and a tool for self-empowerment. Our goal is to transform a young person's relationship with writing—moving from intimidation, dread, and defeat to a source of power, wonder, creativity, and expression. We believe that these attitudinal shifts are the essential foundation for skill development. When students have positive feelings about their own potential, hope becomes the motivator—it links their effort in the classroom to their dreams for their future. When our students achieve significant, authentic successes in their daily life, it inspires them to dream bigger about their future.

And so what is the impact of our work? How do you measure the shifts in confidence a student experiences when they subject their writing to multiple revisions and watch their story come to life on paper? What assessment can tease out the sense of pride a student feels when her writing is published and shared beyond her family and her classroom? Or when a student who hates to write begins to see himself as a writer when he learns he has an ear for the rhythm of language? Or that his story means something to someone outside his community?

We ask students what matters to them, who they are, who they want to become, what they have to say, and why. Listening is at the core of this process. We let them decide how they want to tell their story and we show them that their story matters by giving them a forum—publishing and printing their words. We cultivate a sense of wonder, hold a value of creativity, and we have fun. And we see big shifts in our students' lives. We see grades improve and we also see beaming smiles—indicators of pride, confidence, and hope. And all the while, students are improving their skills.

As a result, our evaluation portfolio seeks to measure changes along the spectrum from skills to affinity. We pair writing assessment data with less tangible data in order to paint a complete picture of the impact of 826 Valencia. To this end, we use a variety of tools, including district-wide writing assessments and Fountas and Pinnell Reading Assessments, attitudinal surveys, conversation, reflection, and feedback—all in an effort to better understand our students' learning and demonstrate the efficacy of our model. And we see that rigorous and fun writing positively impacts students for the long haul.

Here's an example of this combined approach at evaluation last year: Students who participated in the 826 program at Buena Vista Horace Mann K-8 demonstrated accelerated reading level growth, with 74% of third and fourth grade students growing more than a year, in the course of just five months. This is especially important for English Language Learners, who are often entering school below grade level in reading and writing. In the first half of the 2014-2015 school year, over 35% of 826 students have already made over a year of growth, which includes 42% of fifth graders, who are participating in the 826 program for the second consecutive year.

We also move beyond these hard stats to hear our students' reflections. From a student in the program, Jesús Islas Garcia, who began the year hating to write: "When I'm published I feel like I'm a movie star or a millionaire. When I write about my life or funny stories or serious and sad stories, other people can laugh, cry or be sad…and when I'm an adult I will be a famous writer in the city of San Francisco."

A third grade student in the program wrote when asked to explain why the sky is blue:

The sky is made with secret ingredients. It had to be made so everyone could breathe. After it was made, people threw blue glitter at it. A crocodile was selling blue glitter every night. He was selling glitter so everyone could have it. He wanted everyone to have some glitter to throw into the sky. The sky was turquoise or blue or sometimes even purple when people threw glitter at it.
This brief excerpt is infused with creativity—the student is deftly imaginative and lyrical.

Shortly after this program wrapped, we held a two-week writing camp for high schoolers called the Young Authors' Workshop, and we got this written feedback: "I learned that the stories we write are not just words. They are the windows to different worlds, to the writer's soul, to the writer's mind. Writing can take you anywhere."

So how do you measure that "movie star" feeling? Or the new realization in the summer camper that writing opens every door? We consider our students' words, stories, and reflections to be key performance indicators used to drive strategy and decision-making, with the same weight we give quantitative assessments. After all, everyone needs a healthy dose of glitter in their lives.

Hannah Walker is Director of Government Relations at the Food Marketing Institute (FMI). Reposted with permission from FMI's blog.

We have all heard the concerning statistics regarding the amount of food going to waste in the United States while many go without. As a founding member of the Food Waste Reeducation Alliance (FWRA), FMI has been addressing the challenge of reducing food waste in both the United States and globally. FMI recently participated in an announcement with the U.S. Department of Agriculture and the Environmental Protection Agency to emphasize our commitment to the issue and highlight the importance of collaboration between government and the private sector.

Hannah Walker is Director of Government Relations at the Food Marketing Institute (FMI). Reposted with permission from FMI's blog.

We have all heard the concerning statistics regarding the amount of food going to waste in the United States while many go without. As a founding member of the Food Waste Reeducation Alliance (FWRA), FMI has been addressing the challenge of reducing food waste in both the United States and globally. FMI recently participated in an announcement with the U.S. Department of Agriculture and the Environmental Protection Agency to emphasize our commitment to the issue and highlight the importance of collaboration between government and the private sector.

I wanted to highlight an interesting and innovative case of using data to address this pressing public concern and stewardship issue. Feeding America has partnered with dozens of grocers across the country seeking creative ways to solve both the food waste and hunger problems. For years, grocers had limited options when perishables were reaching their sell-by-date; the primary one being send it to the landfill. Significant changes came in 2006 when many of FMI’s members began teaming up with Feeding America to better identify perishable food and donate it rather than discard it. This improved collaboration has proven incredibly successful; grocers have donated over 1.4 billion pounds of food between July 1, 2014 and June 30, 2015, a truly amazing increase from the 140 million first donated when the program started in 2006.

In a recent conversation with Feeding America, I learned that they have found a great willingness from our retail members from large national chains down to smaller operators—to donate perishables that will stock the shelves of the local food bank as opposed to adding to their local landfill. In one short decade, the partnership between the grocery industry and Feeding America has made perishables, such as meat, dairy, and produce much more common items on food bank shelves.

This smart and seemingly simple solution is backed by the use of data, innovation and analytics to measure what and how much food is received and where to send it so that it reaches those in the greatest need. By matching meal gap data with available resources, our local food banks are able to serve those who are in the greatest need while reducing our national food waste at the same time.

While 1.4 billion pounds is an incredible improvement from the reported 140 million reported just nine short years ago, there is always more that can be done. Feeding America and grocery partners are currently targeting an additional 300 million pounds of food they believe they can get by further optimizing the data and collection process.

There will never be one solution to solve the challenges of food waste and hunger in the United States and abroad; however, creative ideas like this partnership backed with strong data and creative innovation are making great strides toward both goals.

Andreas Terzis is a Software Engineer at Google. This post originally appeared on the Google Research Blog.

As Internet traffic has grown and changed, Google and other content and application providers have worked cooperatively with Internet service providers (ISPs) so that services can be delivered quickly, efficiently and cost-effectively. For example, rather than content having to traverse a long distance and many different networks to reach an Internet access provider's network, a content provider might store (cache) the data close by and interconnect ("peer ) directly with the access provider. Google has invested billions of dollars in the network and infrastructure necessary to bring our services as close to your Internet access provider's front door as possible, for free—which both reduces ISPs' costs and improves the user experience.

Andreas Terzis is a Software Engineer at Google. This post originally appeared on the Google Research Blog.

As Internet traffic has grown and changed, Google and other content and application providers have worked cooperatively with Internet service providers (ISPs) so that services can be delivered quickly, efficiently and cost-effectively. For example, rather than content having to traverse a long distance and many different networks to reach an Internet access provider's network, a content provider might store (cache) the data close by and interconnect ("peer ) directly with the access provider. Google has invested billions of dollars in the network and infrastructure necessary to bring our services as close to your Internet access provider's front door as possible, for free—which both reduces ISPs' costs and improves the user experience.

Content and application providers can also tune their services for congested and/or lower bandwidth environments. For instance, YouTube detects how smoothly a video is playing and adjusts the quality to account for temporary fluctuations in bandwidth or congestion. In the Google Video Quality Report, we transparently reveal the speeds YouTube is experiencing on different networks.

As more of Internet traffic becomes encrypted, some network operators have expressed concern about the effect encryption might have on their ability to manage their networks. We don't think there has to be a trade-off here—there are ways to do effective network management of encrypted traffic today, and, through further cooperation between content and application providers and ISPs, we believe this could be made easier while still respecting encryption.

To spur discussion and collaboration on this front, we recently submitted a paper to a workshop organized by the Internet Architecture Board outlining some ideas. We advocate for a model where ISPs selectively share network state to content and applications providers, enabling them to adapt to available network resources.

For example, we recently proposed to the Internet Engineering Task Force the concept of Throughput Guidance (TG), whereby mobile network operators could share information about the throughput of a radio downlink. Preliminary field tests in a production LTE network showed that TG reduces YouTube join latency, defined as the amount of time until the video starts playing, by 8% on average, rebuffering time by 20% on average, and rebuffer count by 2% on average. In addition to improving quality of experience for users, this mechanism improves the utilization of providers’ networks. Encryption of traffic would have no impact on the efficacy of this approach; it works equally well with encrypted and unencrypted traffic.

Throughput Guidance is one possible solution and many questions remain unanswered. It's still relatively early days in our exploration of this and the other measures in our short paper, and we're looking forward to getting feedback and collaborating with network operators and others.

Earlier this week, GOOD Magazine published an interesting piece by Mark Hay on suicide prevention titled "Can Big Data Help Us Fight Rising Suicide Rates?" The part of the article that talks about data-driven prevention starts about halfway through. What follows is an excerpt from that section.

Earlier this week, GOOD Magazine published an interesting piece by Mark Hay on suicide prevention titled "Can Big Data Help Us Fight Rising Suicide Rates?" The part of the article that talks about data-driven prevention starts about halfway through. What follows is an excerpt from that section.

Yet there is one frontier in suicide prevention that seems especially promising, though in a way, it maybe a bit removed from the problem’s human element: big data predictions and intervention targeting.

We know that some populations are more likely than others to commit suicide. Men in the United States account for 79 percent of all suicides. People in their 20s are at higher risk than others. And whites and Native Americans tend to have higher suicide rates than other ethnicities. Yet we don’t have the greatest ability to grasp trends and other niche factors to build up actionable, targetable profiles of communities where we should focus our efforts. We’re stuck trying to expand a suicide prevention dragnet, as opposed to getting individuals at risk the precise information they need (even if they don’t tip off major signs to their friends and family).

That’s a big part of why last year, groups like the National Action Alliance for Suicide Prevention’s Research Prioritization Task Force listed better surveillance, data collection, and research on existing data as priorities for work in the field over the next decade. It’s also why multiple organizations are now developing algorithms to sort through diverse datasets, trying to identify behaviors, social media posting trends, language, lifestyle changes, or any other proxy that can help us predict suicidal tendencies. By doing this, the theory goes, we can target and deliver exactly the right information.

One of the greatest proponents of this data-heavy approach to suicide prevention is the United States Army, which suffers from a suicide rate many times higher than the general population. In 2012, they had more suicide deaths than casualties in Afghanistan. Yet with millions of soldiers stationed around the globe and limited suicide prevention resources, it’s been difficult to simply rely on expanding the dragnet. Instead, last December the Army announced that they’d developed an algorithm that distills the details of a soldier’s personal information into a set of 400 characteristics that mix and match to show whether an individual is likely in need of intervention. Their analysis isn’t perfect yet, but they’ve been able to identify a cluster of characteristics within 5 percent of military personnel who accounted for 52 percent of suicides, showing that they’re on the right track to better targeting and allocating prevention resources.

Yet perhaps the greatest distillation of this data-driven approach (combined with the expansive, barrier-reducing impulse of mainstream efforts) is the Crisis Text Line. Created in 2013 by organizers from DoSomething.org, the text line allows those too scared, embarrassed, or uncomfortable to vocalize their problems to friends, or over a hotline, to simply trace a pattern on a cell phone keypad (741741) and then type their problems in a text message. As of 2015, algorithmic learning allows the Crisis Text Line to search for keywords, based on over 8 million previous texts and data gathered from hundreds of suicide prevention workers, to identify who’s at serious risk and assign counselors to respond. But more than that, the data in texts can trip off time and vocabulary sensors, matching counselors with expertise in certain areas to respond to specific texters, or bringing up precisely tailored resources. For example, the system knows that self-harm peaks at 4 a.m. and that people typing “Mormon” are usually dealing with issues related to LGBTQ identity, discrimination, and isolation. Low-impact and low-cost with high potential for delivering the best information possible to those in need, it’s one of the cleverer young programs out there pushing the suicide prevention gains made over the last century.

It’ll be a few years before we can understand the impact of data analysis and targeting on suicide prevention efforts, especially relative to general attempts to expand existing programs. And given the limited success of a half-century of serious gains in understanding and resource provision, we’d be wise not to get our hopes up too much. But it’s not unreasonable to suspect that a combination of diversifying means of access, lowering barriers of communication, and better identifying those at risk could help us bring programs to populations that have not yet received them (or that we could not support quickly enough before). At the very least, crunching existing data may help us to discover why suicide rates have increased in recent years and to understand the mechanisms of this widespread social issue. We have solid, logical reason to support the development of programs like the Army’s algorithms and the Crisis Text Line, and to push for further similar initiatives. But really we have reason to support any kind of suicide prevention innovation, even if it feels less robust or promising than the recent data-driven efforts. If you've ever witnessed the pain that those moving towards suicide feel, or the wide-reaching fallout after someone takes his or her life, you'll understand the visceral, human need to let a thousand flowers bloom, desperately hoping that one of them sticks. Hopefully, if data mining and targeting works well, that'll only inspire further innovation, slowly putting a greater and greater dent in the phenomenon of suicide.

Moritz Hardt is a Research Scientist at Google. This post was originally published on the Google Research Blog.

Machine learning and statistical analysis play an important role at the forefront of scientific and technological progress. But with all data analysis, there is a danger that findings observed in a particular sample do not generalize to the underlying population from which the data were drawn. A popular XKCD cartoon illustrates that if you test sufficiently many different colors of jelly beans for correlation with acne, you will eventually find one color that correlates with acne at a p-value below the infamous 0.05 significance level.

Moritz Hardt is a Research Scientist at Google. This post was originally published on the Google Research Blog.

Machine learning and statistical analysis play an important role at the forefront of scientific and technological progress. But with all data analysis, there is a danger that findings observed in a particular sample do not generalize to the underlying population from which the data were drawn. A popular XKCD cartoon illustrates that if you test sufficiently many different colors of jelly beans for correlation with acne, you will eventually find one color that correlates with acne at a p-value below the infamous 0.05 significance level.

Image credit: XKCD

Unfortunately, the problem of false discovery is even more delicate than the cartoon suggests. Correcting reported p-values for a fixed number of multiple tests is a fairly well understood topic in statistics. A simple approach is to multiply each p-value by the number of tests, but there are more sophisticated tools. However, almost all existing approaches to ensuring the validity of statistical inferences assume that the analyst performs a fixed procedure chosen before the data are examined. For example, "test all 20 flavors of jelly beans." In practice, however, the analyst is informed by data exploration, as well as the results of previous analyses. How did the scientist choose to study acne and jelly beans in the first place? Often such choices are influenced by previous interactions with the same data. This adaptive behavior of the analyst leads to an increased risk of spurious discoveries that are neither prevented nor detected by standard approaches. Each adaptive choice the analyst makes multiplies the number of possible analyses that could possibly follow; it is often difficult or impossible to describe and analyze the exact experimental setup ahead of time.

In The Reusable Holdout: Preserving Validity in Adaptive Data Analysis, a joint work with Cynthia Dwork (Microsoft Research), Vitaly Feldman (IBM Almaden Research Center), Toniann Pitassi (University of Toronto), Omer Reingold (Samsung Research America) and Aaron Roth (University of Pennsylvania), to appear in Science tomorrow, we present a new methodology for navigating the challenges of adaptivity. A central application of our general approach is the reusable holdout mechanism that allows the analyst to safely validate the results of many adaptively chosen analyses without the need to collect costly fresh data each time.

The curse of adaptivity

A beautiful example of how false discovery arises as a result of adaptivity is Freedman's paradox. Suppose that we want to build a model that explains "systolic blood pressure" in terms of hundreds of variables quantifying the intake of various kinds of food. In order to reduce the number of variables and simplify our task, we first select some promising looking variables, for example, those that have a positive correlation with the response variable (systolic blood pressure). We then fit a linear regression model on the selected variables. To measure the goodness of our model fit, we crank out a standard F-test from our favorite statistics textbook and report the resulting p-value.

Inference after selection: We first select a subset of the variables based on a data-dependent criterion and then fit a linear model on the selected variables.

Freedman showed that the reported p-value is highly misleading—even if the data were completely random with no correlation whatsoever between the response variable and the data points, we'd likely observe a significant p-value! The bias stems from the fact that we selected a subset of the variables adaptively based on the data, but we never account for this fact. There is a huge number of possible subsets of variables that we selected from. The mere fact that we chose one test over the other by peeking at the data creates a selection bias that invalidates the assumptions underlying the F-test.

Freedman's paradox bears an important lesson. Significance levels of standard procedures do not capture the vast number of analyses one can choose to carry out or to omit. For this reason, adaptivity is one of the primary explanations of why research findings are frequently false as was argued by Gelman and Loken who aptly refer to adaptivity as "garden of the forking paths."

Machine learning competitions and holdout sets

Adaptivity is not just an issue with p-values in the empirical sciences. It affects other domains of data science just as well. Machine learning competitions are a perfect example. Competitions have become an extremely popular format for solving prediction and classification problems of all sorts.

Each team in the competition has full access to a publicly available training set which they use to build a predictive model for a certain task such as image classification. Competitors can repeatedly submit a model and see how the model performs on a fixed holdout data set not available to them. The central component of any competition is the public leaderboard which ranks all teams according to the prediction accuracy of their best model so far on the holdout. Every time a team makes a submission they observe the score of their model on the same holdout data. This methodology is inspired by the classic holdout method for validating the performance of a predictive model.

Ideally, the holdout score gives an accurate estimate of the true performance of the model on the underlying distribution from which the data were drawn. However, this is only the case when the model is independent of the holdout data! In contrast, in a competition the model generally incorporates previously observed feedback from the holdout set. Competitors work adaptively and iteratively with the feedback they receive. An improved score for one submission might convince the team to tweak their current approach, while a lower score might cause them to try out a different strategy. But the moment a team modifies their model based on a previously observed holdout score, they create a dependency between the model and the holdout data that invalidates the assumption of the classic holdout method. As a result, competitors may begin to overfit to the holdout data that supports the leaderboard. This means that their score on the public leaderboard continues to improve, while the true performance of the model does not. In fact, unreliable leaderboards are a widely observed phenomenon in machine learning competitions.

Reusable holdout sets

A standard proposal for coping with adaptivity is simply to discourage it. In the empirical sciences, this proposal is known as pre-registration and requires the researcher to specify the exact experimental setup ahead of time. While possible in some simple cases, it is in general too restrictive as it runs counter to today's complex data analysis workflows.

Rather than limiting the analyst, our approach provides means of reliably verifying the results of an arbitrary adaptive data analysis. The key tool for doing so is what we call the reusable holdout method. As with the classic holdout method discussed above, the analyst is given unfettered access to the training data. What changes is that there is a new algorithm in charge of evaluating statistics on the holdout set. This algorithm ensures that the holdout set maintains the essential guarantees of fresh data over the course of many estimation steps.

The limit of the method is determined by the size of the holdout set—the number of times that the holdout set may be used grows roughly as the square of the number of collected data points in the holdout, as our theory shows.

Armed with the reusable holdout, the analyst is free to explore the training data and verify tentative conclusions on the holdout set. It is now entirely safe to use any information provided by the holdout algorithm in the choice of new analyses to carry out, or the tweaking of existing models and parameters.

A general methodology

The reusable holdout is only one instance of a broader methodology that is, perhaps surprisingly, based on differential privacy—a notion of privacy preservation in data analysis. At its core, differential privacy is a notion of stability requiring that any single sample should not influence the outcome of the analysis significantly.

Example of a stable learning algorithm: Deletion of any single data point does not affect the accuracy of the classifier much.

A beautiful line of work in machine learning shows that various notions of stability imply generalization. That is any sample estimate computed by a stable algorithm (such as the prediction accuracy of a model on a sample) must be close to what we would observe on fresh data.

What sets differential privacy apart from other stability notions is that it is preserved by adaptive composition. Combining multiple algorithms that each preserve differential privacy yields a new algorithm that also satisfies differential privacy albeit at some quantitative loss in the stability guarantee. This is true even if the output of one algorithm influences the choice of the next. This strong adaptive composition property is what makes differential privacy an excellent stability notion for adaptive data analysis.

In a nutshell, the reusable holdout mechanism is simply this: access the holdout set only through a suitable differentially private algorithm. It is important to note, however, that the user does not need to understand differential privacy to use our method. The user interface of the reusable holdout is the same as that of the widely used classical method.

Reliable benchmarks

A closely related work with Avrim Blum dives deeper into the problem of maintaining a reliable leaderboard in machine learning competitions (see this blog post for more background). While the reusable holdout could directly be used for this purpose, it turns out that a variant of the reusable holdout, we call the Ladder algorithm, provides even better accuracy.

This method is not just useful for machine learning competitions, since there are many problems that are roughly equivalent to that of maintaining an accurate leaderboard in a competition. Consider, for example, a performance benchmark that a company uses to test improvements to a system internally before deploying them in a production system. As the benchmark data set is used repeatedly and adaptively for tasks such as model selection, hyper-parameter search and testing, there is a danger that eventually the benchmark becomes unreliable.

Conclusion

Modern data analysis is inherently an adaptive process. Attempts to limit what data scientists will do in practice are ill-fated. Instead we should create tools that respect the usual workflow of data science while at the same time increasing the reliability of data driven insights. It is our goal to continue exploring techniques that can help to create more reliable validation techniques and benchmarks that track true performance more accurately than existing methods.