The Experimentation Platform was a project headed by Ronny Kohavi from 2006. It went live in 2007 and the team was merged into Bing in 10/2010. Starting March 2014, the mission is back as part of the Artificial Intelligence and Research group: Ronny Kohavi is the General Manager for the Analysis & Experimentation team.
The Analysis & Experimentation Team is Hiring
Developers, Analysts/Data Scientists, and Program Managers.
KDD 2017 Tutorial on A/B Testing at Scale
- Agenda and Slides
- Date/Time/Location: 8/13/2017 1PM-5PM, Suite 302 at Conference hotel in Halifax, CA
Papers (Newest First)
Pavel Dmitriev, Somit Gupta, Dong Woo Kim, and Garnet Vaz, KDD 2017
The Evolution of Continuous Experimentation in Software Product Development (slides)
Aleksander Fabijan, Pavel Dmitriev, Helena Holmström Olsson, Jan Bosch. ICSE 2017
Software development companies are increasingly aiming to become data-driven by trying to continuously experiment with the products used by their customers. Although familiar with the competitive edge that the A/B testing technology delivers, they seldom succeed in evolving and adopting the methodology. In this paper, and based on an exhaustive and collaborative case study research in a large software-intense company with highly developed experimentation culture, we present the evolution process of moving from ad-hoc customer data analysis towards continuous controlled experimentation at scale. Our main contribution is the “Experimentation Evolution Model” in which we detail three phases of evolution: technical, organizational and business evolution. With our contribution, we aim to provide guidance to practitioners on how to develop and scale continuous experimentation in software organizations with the purpose of becoming data-driven at scale.
This is the authors' version of the work. It is posted here for your personal use. Not for redistribution. The definitive version is published in the proceedings of the 39th International Conference on Software Engineering (ICSE '17), Buenos Aires, Argentina. May 20 - 28, 2017. IEEE Press, Piscataway, NJ, USA. See: http://dl.acm.org/citation.cfm?id=3097460
Characterizing Experimentation in Continuous Deployment: a Case Study on Bing (slides)
Katja Kevic, Brendan Murphy, Laurie Williams, Jennifer Beckmann. ICSE 2017
The practice of continuous deployment enables product teams to release content to end users within hours or days, rather than months or years. These faster deployment cycles, along with rich product instrumentation, allows product
teams to capture and analyze feature usage measurements. Product teams define a hypothesis and a set of metrics to assess how a code or feature change will impact the user. Supported by a framework, a team can deploy that change to subsets of users, enabling randomized controlled experiments. Based on the impact of the change, the product team may decide to modify the change, to deploy the change to all users, or to abandon the change. This experimentation process enables product teams to only deploy the changes that positively impact the user experience. The goal of this research is to aid product teams to improve their deployment process through providing an empirical characterization of an experimentation process when applied to a large-scale and mature service. Through an analysis of 21,220 experiments applied in Bing since 2014, we observed the complexity of the experimental process and characterized the full deployment cycle (from code change to deployment to all users). The analysis identified that the experimentation process takes an average of 42 days, including multiple iterations of one or two week experiment runs. Such iterations typically indicate that problems were found that could have hurt the users or business if the feature was just launched, hence the experiment provided real value to the organization. Further, we discovered that code changes for experiments are four times larger than other code changes. We identify that the code associated with 33.4% of the experiments is eventually shipped to all users. These fully-deployed code changes are significantly larger than the code changes for the other experiments, in terms of files (35.7%), changesets (80.4%) and contributors (20.0%).
This is the authors' version of the work. It is posted here for your personal use. Not for redistribution. The definitive version is published in the proceedings of the 39th International Conference on Software Engineering: Software Engineering in Practice Track (ICSE-SEIP '17), Pages 123-132, Buenos Aires, Argentina. May 20 - 28, 2017. IEEE Press, Piscataway, NJ, USA. See http://dl.acm.org/citation.cfm?id=3103129
Trustworthy analysis of online A/B tests: Pitfalls, challenges and solutions
Alex Deng, Jiannan Lu, Jonathan Litz . WSDM 2017
A/B tests (or randomized controlled experiments) play an integral role in the research and development cycles of technology companies. As in classic randomized experiments (e.g., clinical trials), the underlying statistical analysis of A/B tests is based on assuming the randomization unit is independent and identically distributed (i.i.d.). However, the randomization mechanisms utilized in online A/B tests can be quite complex and may render this assumption invalid. Analysis that unjustifiably relies on this assumption can yield untrustworthy results and lead to incorrect conclusions. Motivated by challenging problems arising from actual online experiments, we propose a new method of variance estimation that relies only on practically plausible assumptions, is directly applicable to a wide of range of randomization mechanisms, and can be implemented easily. We examine its performance and illustrate its advantages over two commonly used methods of variance estimation on both simulated and empirical datasets. Our results lead to a deeper understanding of the conditions under which the randomization unit can be treated as i.i.d. In particular, we show that for purposes of variance estimation, the randomization unit can be approximated as i.i.d. when the individual treatment effect variation is small; however, this approximation can lead to variance under-estimation when the individual treatment effect variation is large.
Continuous monitoring of A/B tests without pain: Optional stopping in Bayesian testing
Alex Deng, Jiannan Lu, Shouyuan Chen. DSAA 2016
A/B testing is one of the most successful applications of statistical theory in the Internet age. A crucial problem of Null Hypothesis Statistical Testing (NHST), the backbone of A/B testing methodology, is that experimenters are not allowed to continuously monitor the results and make decisions in real time. Many people see this restriction as a setback against the trend in the technology toward real time data analytics. Recently, Bayesian Hypothesis Testing, which intuitively is more suitable for real time decision making, attracted growing interest as a viable alternative to NHST. While corrections of NHST for the continuous monitoring setting are well established in the existing literature and known in A/B testing community, the debate over the issue of whether continuous monitoring is a proper practice in Bayesian testing exists among both academic researchers and general practitioners. In this paper, we formally prove the validity of Bayesian testing under proper stopping rules, and illustrate the theoretical results with concrete simulation illustrations. We point out common bad practices where stopping rules are not proper, and discuss how priors can be learned objectively. General guidelines for researchers and practitioners are also provided.
Data-Driven Metric Development for Online Controlled Experiments: Seven Lessons Learned
Alex Deng, Xiaolin Shi. KDD 2016.
Online controlled experiments, also called A/B testing, have been established as the mantra for data-driven decision making in many web-facing companies. In recent years, there are emerging research works focusing on building the platform and scaling it up , best practices and lessons learned to obtain trustworthy results [19; 20; 23; 26], and experiment design techniques and various issues related to statistical inference and testing [6; 7; 8]. However, despite playing a central role in online controlled experiments, there is little published work on treating metric development itself as a data-driven process. In this paper, we focus on the topic of how to develop meaningful and useful metrics for online services in their online experiments, and show how data-driven techniques and criteria can be applied in metric development process. In particular, we emphasize two fundamental qualities for the goal metrics (or Overall Evaluation Criteria) of any online service: directionality and sensitivity. We share lessons on why these two qualities are critical, how to measure these two qualities of metrics of interest, how to develop metrics with clear directionality and high sensitivity by using approaches based on user behavior models and data-driven calibration, and how to choose the right goal metrics for the entire online services.
Pitfalls of Long-Term Online Controlled Experiments
Pavel Dmitriev, Brian Frasca, Somit Gupta, Ron Kohavi, and Garnet Vaz. IEEE Big Data 2016
Online controlled experiments (e.g., A/B tests) are now regularly used to guide product development and accelerate innovation in software. Product ideas are evaluated as scientific hypotheses, and tested on web sites, mobile applications, desktop applications, services, and operating system features.
One of the key challenges for organizations that run controlled experiments is to select an Overall Evaluation Criterion (OEC), i.e., the criterion by which to evaluate the different variants. The difficulty is that short-term changes to metrics may not predict the long-term impact of a change. For example, raising prices likely increases short-term revenue but also likely reduces long-term revenue (customer lifetime value) as users abandon. Degrading search results in a Search Engine causes users to search more, thus increasing query share short-term, but increasing abandonment and thus reducing long-term customer lifetime value. Ideally, an OEC is based on metrics in a short-term experiment that are good predictors of long-term value.
To assess long-term impact, one approach is to run long-term controlled experiments and assume that long-term effects are represented by observed metrics. In this paper we share several examples of long-term experiments and the pitfalls associated with running them. We discuss cookie stability, survivorship bias, selection bias, and perceived trends, and share methodologies that can be used to partially address some of these issues.
While there is clearly value in evaluating long-term trends, experimenters running long-term experiments must be cautious, as results may be due to the above pitfalls more than the true delta between the Treatment and Control. We hope our real examples and analyses will sensitize readers to the issues and encourage the development of new methodologies for this important problem.
by Pavel Dmitriev and Xian Wu. CIKM 2016
You get what you measure, and you can’t manage what you don’t measure. Metrics are a powerful tool used in organizations to set goals, decide which new products and features should be released to customers, which new tests and experiments should be conducted, and how resources should be allocated. To a large extent, metrics drive the direction of an organization, and getting metrics “right” is one of the most important and difficult problems an organization needs to solve. However, creating good metrics that capture long-term company goals is difficult. They try to capture abstract concepts such as success, delight, loyalty, engagement, life-time value, etc. How can one determine that a metric is a good one? Or, that one metric is better than another? In other words, how do we measure the quality of metrics? Can the evaluation process be automated so that anyone with an idea of a new metric can quickly evaluate it? In this paper we describe the metric evaluation system deployed at Bing, where we have been working on designing and improving metrics for over five years. We believe that by applying a data driven approach to metric evaluation we have been able to substantially improve our metrics and, as a result, ship better features and improve search experience for Bing’s users.
The internet connectivity of client software (e.g., apps running on phones and PCs), web sites, and online services provide an unprecedented opportunity to evaluate ideas quickly using controlled experiments, also called A/B tests, split tests, randomized experiments, control/treatment tests, and online field experiments. Unlike most data mining techniques for finding correlational patterns, controlled experiments allow establishing a causal relationship with high probability. Experimenters can utilize the Scientific Method to form a hypothesis of the form "If a specific change is introduced, will it improve key metrics?" and evaluate it with real users.
The theory of a controlled experiment dates back to Sir Ronald A. Fisher’s experiments at the Rothamsted Agricultural Experimental Station in England in the 1920s, and the topic of offline experiments is well developed in Statistics (Box 2005). Online Controlled Experiments started to be used in the late 1990s with the growth of the Internet. Today, many large sites, including Amazon, Bing, Facebook, Google, LinkedIn, and Yahoo! run thousands to tens of thousands of experiments each year testing user interface (UI) changes, enhancements to algorithms (search, ads, personalization, recommendation, etc.), changes to apps, content management system, etc. Online controlled experiments are now considered an indispensable tool, and their use is growing for startups and smaller websites. Controlled experiments are especially useful in combination with Agile software development (Martin 2008, Rubin 2012), Steve Blank’s Customer Development process (Blank 2005), and MVPs (Minimum Viable Products) popularized by Eric Ries’s Lean Startup (Ries 2011).
As A/B testing gains wider adoption in the industry, more people begin to realize the limitations of the traditional frequentist null hypothesis statistical testing (NHST). The large number of search results for the query ``Bayesian A/B testing'' shows just how much the interest in the Bayesian perspective is growing. In recent years there are also voices arguing that Bayesian A/B testing should replace frequentist NHST and is strictly superior in all aspects. Our goal here is to clarify the myth by looking at both advantages and issues of Bayesian methods. In particular, we propose an objective Bayesian A/B testing framework for which we hope to bring the best from Bayesian and frequentist methods together. Unlike traditional methods, this method requires the existence of historical A/B test data to objectively learn a prior. We have successfully applied this method to Bing, using thousands of experiments to establish the priors.
Diluted Treatment Effect Estimation for Trigger Analysis in Online Controlled Experiments
Alex Deng and Victor Hu
Online controlled experiments, also called A/B testing, is playing a central role in many data-driven web-facing companies. It is well known and intuitively obvious to many practitioners that when testing a feature with low coverage, analyzing all data collected without zooming into the part that could be affected by the treatment often leads to under-powered hypothesis testing. A common practice is to use triggered analysis. To estimate the overall treatment effect, certain dilution formula is then applied to translate the estimated effect in triggered analysis back to the original all up population. In this paper, we discuss two different types of trigger analyses. We derive correct dilution formulas and show for a set of widely used metrics, namely ratio metrics, correctly deriving and applying those dilution formulas are not trivial. We observe many practitioners in this industry are often applying approximate formulas or even wrong formulas when doing effect dilution calculation. To deal with that, instead of estimating trigger treatment effect followed by effect translation using dilution formula, we aim at combining these two steps into one streamlined analysis, producing more accurate estimation of overall treatment effect together with even higher statistical power than a triggered analysis. The approach we propose in this paper is intuitive, easy to apply and general enough for all types of triggered analyses and all types of metrics.
Seven Rules of Thumb for Web Site Experimenters
Ron Kohavi, Alex Deng, Roger Longbotham, and Ya Xu
Web site owners, from small web sites to the largest properties that include Amazon, Facebook, Google, LinkedIn, Microsoft, and Yahoo, attempt to improve their web sites, optimizing for criteria ranging from repeat usage, time on site, to revenue. Having been involved in running thousands of controlled experiments at Amazon, Booking.com, LinkedIn, and multiple Microsoft properties, we share seven rules of thumb for experimenters, which we have generalized from these experiments and their results. These are principles that we believe have broad applicability in web optimization and analytics outside of controlled experiments, yet they are not provably correct, and in some cases exceptions are known.
To support these rules of thumb, we share multiple real examples, most being shared in a public paper for the first time. Some rules of thumb have previously been stated, such as “speed matters,” but we describe the assumptions in the experimental design and share additional experiments that improved our understanding of where speed matters more: certain areas of the web page are more critical.
This paper serves two goals. First, it can guide experimenters with rules of thumb that can help them optimize their sites. Second, it provides the KDD community with new research challenges on the applicability, exceptions, and extensions to these, one of the goals for KDD’s industrial track.
Statistical Inference in Two-stage Online Controlled Experiments with Treatment Selection and Validation
Alex Deng, Tianxi Li and Yu Guo
Online controlled experiments, also called A/B testing, have been established as the mantra for data-driven decision making in many web-facing companies. A/B Testing support decision making by directly comparing two variants at a time. It can be used for comparison between (1) two candidate treatments and (2) a candidate treatment and an established control. In practice, one typically runs an experiment with multiple treatments together with a control to make decision for both purposes simultaneously. This is known to have two issues. First, having multiple treatments increases false positives due to multiple comparison. Second, the selection process causes an upward bias in estimated effect size of the best observed treatment. To overcome these two issues, a two stage process is recommended, in which we select the best treatment from the first screening stage and then run the same experiment with only the selected best treatment and the control in the validation stage. Traditional application of this two-stage design often focus only on results from the second stage. In this paper, we propose a general methodology for combining the first screening stage data together with validation stage data for more sensitive hypothesis testing and more accurate point estimation of the treatment effect. Our method is widely applicable to existing online controlled experimentation systems.
Online Controlled Experiments at Large Scale
Ron Kohavi, Alex Deng, Brian Frasca, Toby Walker, Ya Xu, Nils Pohlmann
Web-facing companies, including Amazon, eBay, Etsy, Facebook, Google, Groupon, Intuit, LinkedIn, Microsoft, Netflix, Shop Direct, StumbleUpon, Yahoo, and Zynga use online controlled experiments to guide product development and accelerate innovation. At Microsoft’s Bing, the use of controlled experiments has grown exponentially over time, with over 200 concurrent experiments now running on any given day. Running experiments at large scale requires addressing multiple challenges in three areas: cultural/organizational, engineering, and trustworthiness. On the cultural and organizational front, the larger organization needs to learn the reasons for running controlled experiments and the tradeoffs between controlled experiments and other methods of evaluating ideas. We discuss why negative experiments, which degrade the user experience short term, should be run, given the learning value and long-term benefits. On the engineering side, we architected a highly scalable system, able to handle data at massive scale: hundreds of concurrent experiments, each containing millions of users. Classical testing and debugging techniques no longer apply when there are millions of live variants of the site, so alerts are used to identify issues rather than relying on heavy up-front testing. On the trustworthiness front, we have a high occurrence of false positives that we address, and we alert experimenters to statistical interactions between experiments. The Bing Experimentation System is credited with having accelerated innovation and increased annual revenues by hundreds of millions of dollars, by allowing us to find and focus on key ideas evaluated through thousands of controlled experiments. A 1% improvement to revenue equals $10M annually in the US, yet many ideas impact key metrics by 1% and are not well estimated a-priori. The system has also identified many negative features that we avoided deploying, despite key stakeholders’ early excitement, saving us similar large amounts.
Improving the Sensitivity of Online Controlled Experiments by Utilizing Pre-Experiment Data
Alex Deng, Ya Xu, Ron Kohavi, Toby Walker
Online controlled experiments are at the heart of making data-driven decisions at a diverse set of companies, including Amazon, eBay, Facebook, Google, Microsoft, Yahoo, and Zynga. Small diﬀerences in key metrics, on the order of fractions of a percent, may have very signiﬁcant business implications. At Bing it is not uncommon to see experiments that impact annual revenue by millions of dollars, even tens of millions of dollars, either positively or negatively. With thousands of experiments being run annually, improving the sensitivity of experiments allows for more precise assessment of value, or equivalently running the experiments on smaller populations (supporting more experiments) or for shorter durations (improving the feedback cycle and agility). We propose an approach (CUPED) that utilizes data from the pre-experiment period to reduce metric variability and hence achieve better sensitivity. This technique is applicable to a wide variety of key business metrics, and it is practical and easy to implement. The results on Bing’s experimentation system are very successful: we can reduce variance by about 50%, eﬀectively achieving the same statistical power with only half of the users, or half the duration
Trustworthy Online Controlled Experiments: Five Puzzling Outcomes Explained
Ron Kohavi, Alex Deng, Brian Frasca, Roger Longbotham, Toby Walker, Ya Xu
Online controlled experiments are often utilized to make data-driven decisions at Amazon, Microsoft, eBay, Facebook, Google, Yahoo, Zynga, and at many other companies. While the theory of a controlled experiment is simple, and dates back to Sir Ronald A. Fisher’s experiments at the Rothamsted Agricultural Experimental Station in England in the 1920s, the deployment and mining of online controlled experiments at scale—thousands of experiments now—has taught us many lessons. These exemplify the proverb that the difference between theory and practice is greater in practice than in theory. We present our learnings as they happened: puzzling outcomes of controlled experiments that we analyzed deeply to understand and explain. Each of these took multiple-person weeks to months to properly analyze and get to the often surprising root cause. The root causes behind these puzzling results are not isolated incidents; these issues generalized to multiple experiments. The heightened awareness should help readers increase the trustworthiness of the results coming out of controlled experiments. At Microsoft’s Bing, it is not uncommon to see experiments that impact annual revenue by millions of dollars, thus getting trustworthy results is critical and investing in understanding anomalies has tremendous payoff: reversing a single incorrect decision based on the results of an experiment can fund a whole team of analysts. The topics we cover include: the OEC (Overall Evaluation Criterion), click tracking, effect trends, experiment length and power, and carryover effects.
Unexpected Results in Online Controlled Experiments, SIGKDD Dec 2010
Ron Kohavi and Roger Longbotham
Abstract: Controlled experiments, also called randomized experiments and A/B tests, have had a profound influence on multiple fields, including medicine, agriculture, manufacturing, and advertising. Offline controlled experiments have been well studied and documented since Sir Ronald A. Fisher led the development of statistical experimental design while working at the Rothamsted Agricultural Experimental Station in England in the 1920s. With the growth of the world-wide-web and web services, online controlled experiments are being used frequently, utilizing software capabilities like ramp-up (exposure control) and running experiments on large server farms with millions of users. We share several real examples of unexpected results and lessons learned.
Tracking Users' Clicks and Submits: Tradeoffs between User Experience and Data Loss, Oct 2010
Ron Kohavi, David Messner,Seth Eliot, Juan Lavista Ferres, Randy Henne, Vignesh Kannappan, and Justin Wang
Abstract: Tracking users’ online clicks and form submits (e.g., searches) is critical for web analytics, controlled experiments, and business intelligence.Most sites use web beacons to track user actions, but waiting for the beacon to return on clicks and submits slows the next action (e.g., showing search results or the destination page).One possibility is to use a short timeout and common wisdom is that the more time given to the tracking mechanism (suspending the user action), the lower the data loss.Research from Amazon, Google, and Microsoft showed that small delays of a few hundreds of milliseconds have dramatic negative impact on revenue and user experience (Kohavi, et al., 2009 p. 173), yet we found that many websites allow long delays in order to collect click.For example, until March 2010, multiple Microsoft sites waited for click beacons to return with a 2-second timeout, introducing a delay of about 400msec on user clicks. To the best of our knowledge, this is the first published empirical study of the subject under a controlled environment. While we confirm the common wisdom about the tradeoff in general, a surprising result is that the tradeoff does not exist for the most common browser family, Microsoft Internet Explorer (IE), where no delay suffices. This finding has significant implications for tracking users since no waits is required to prevent data loss for IE browsers and it could significantly improve revenue and user experience.The recommendations here have been implemented by the MSN US home page and Hotmail.
Online Experiments: Practical Lessons, IEEE Computer Sept 2010
Ron Kohavi, Roger Longbotham, and Toby Walker
Abstract: From ancient times through the 19th century, physicians used bloodletting to treat acne, cancer, diabetes, jaundice, plague, and hundreds of other diseases and ailments (D. Wooton, Doctors Doing Harm since Hippocrates, Oxford Univ. Press, 2006). It was judged most effective to bleed patients while they were sitting upright or standing erect, and blood was often removed until the patient fainted. On 12 December 1799, 67-year-old President George Washington rode his horse in heavy snowfall to inspect his plantation at Mount Vernon. A day later, he was in respiratory distress and his doctors extracted nearly half of his blood over 10 hours, causing anemia and hypotension; he died that night.
Today, we know that bloodletting is unhelpful because in 1828 a Parisian doctor named Pierre Louis did a controlled experiment. He treated 78 people suffering from pneumonia with early and frequent bloodletting or less aggressive measures and found that bloodletting did not help survival rates or recovery times.
Having roots in agriculture and medicine, controlled experiments have spread into the online world of websites and services. In an earlier Web Technologies article (R. Kohavi and R. Longobotham, “Online Experiments: Lessons Learned,” Computer, Sept. 2007, pp. 85-87) and a related survey (R. Kohavi et al., “Controlled Experiments on the Web: Survey and Practical Guide,” Data Mining and Knowledge Discovery, Feb. 2009, pp. 140-181), Microsoft’s Experimentation Platform team introduced basic practices of good online experimentation.
Three years later and having run hundreds of experiments on more than 20 websites, including some of the world’s largest, like msn.com and bing.com, we have learned some important practical lessons about the limitations of standard statistical formulas and about data traps. These lessons, even for seemingly simple univariate experiments, aren’t taught in Statistics 101. After reading this article we hope you will have better negative introspection: to know what you don’t know.
Online Experimentation at Microsoft, Sept 2009
Ron Kohavi, Thomas Crook, and Roger Longbotham
The paper won 3rd place at the Third Workshop on Data Mining Case Studies and Practice Prize.
Abstract: Knowledge Discovery and Data Mining techniques are now commonly used to find novel, potentially useful, patterns in data (Fayyad, et al., 1996; Chapman, et al., 2000). Most KDD applications involve post-hoc analysis of data and are therefore mostly limited to the identification of correlations. Recent seminal work on Quasi-Experimental Designs (Jensen, et al., 2008) attempts to identify causal relationships. Controlled experiments are a standard technique used in multiple fields. Through randomization and proper design, experiments allow establishing causality scientifically, which is why they are the gold standard in drug tests. In software development, multiple techniques are used to define product requirements; controlled experiments provide a way to assess the impact of new features on customer behavior. The Data Mining Case Studies workshop calls for describing completed implementations related to data mining. Over the last three years, we built an experimentation platform system (ExP) at Microsoft, capable of running and analyzing controlled experiments on web sites and services. The goal is to accelerate innovation through trustworthy experimentation and to enable a more scientific approach to planning and prioritization of features and designs (Foley, 2008).Along the way, we ran many experiments on over a dozen Microsoft properties and had to tackle both technical and cultural challenges. We previously surveyed the literature on controlled experiments and shared technical challenges (Kohavi, et al., 2009). This paper focuses on problems not commonly addressed in technical papers: cultural challenges, lessons, and the ROI of running controlled experiments.
Longer version of the above paper is at Online Experimentation at Microsoft (sanitized Thinkweek 2009)http://bit.ly/expPracticalLessons
By Ron Kohavi, Thomas Crook, Roger Longbotham, Brian Frasca, Randy Henne, Juan Lavista Ferres, Tamir Melamed
The ThinkWeek paper was recognized as a top-30 ThinkWeek at Microsoft.
Controlled experiments on the web: survey and practical guide (2009)
Ron Kohavi, Roger Longbotham, Dan Sommerfield, Randy Henne
Abstract: The web provides an unprecedented opportunity to evaluate ideas quickly using controlled experiments, also called randomized experiments, A/B tests (and their generalizations), split tests, Control/Treatment tests, MultiVariable Tests (MVT) and parallel flights. Controlled experiments embody the best scientific design for establishing a causal relationship between changes and their influence on user-observable behavior. We provide a practical guide to conducting online experiments, where end-users can help guide the development of features. Our experience indicates that significant learning and return-on-investment (ROI) are seen when development teams listen to their customers, not to the Highest Paid Person’s Opinion (HiPPO). We provide several examples of controlled experiments with surprising results. We review the important ingredients of running controlled experiments, and discuss their limitations (both technical and organizational). We focus on several areas that are critical to experimentation, including statistical power, sample size, and techniques for variance reduction. We describe common architectures for experimentation systems and analyze their advantages and disadvantages. We evaluate randomization and hashing techniques, which we show are not as simple in practice as is often assumed. Controlled experiments typically generate large amounts of data, which can be analyzed using data mining techniques to gain deeper understanding of the factors influencing the outcome of interest, leading to new hypotheses and creating a virtuous cycle of improvements. Organizations that embrace controlled experiments with clear evaluation criteria can evolve their systems with automated optimizations and real-time analyses. Based on our extensive practical experience with multiple systems and organizations, we share key lessons that will help practitioners in running trustworthy controlled experiments.
Seven Pitfalls to Avoid when Running Controlled Experiments on the Web, KDD 2009
Thomas Crook, Brian Frasca, Ron Kohavi, and Roger Longbotham
Abstract: Controlled experiments, also called randomized experiments and A/B tests, have had a profound influence on multiple fields, including medicine, agriculture, manufacturing, and advertising. While the theoretical aspects of offline controlled experiments have been well studied and documented, the practical aspects of running them in online settings, such as web sites and services, are still being developed. As the usage of controlled experiments grows in these online settings, it is becoming more important to understand the opportunities and pitfalls one might face when using them in practice. A survey of online controlled experiments and lessons learned were previously documented in Controlled Experiments on the Web: Survey and Practical Guide(Kohavi, et al., 2009). In this follow-on paper, we focus on pitfalls we have seen after running numerous experiments at Microsoft.The pitfalls include a wide range of topics, such as assuming that common statistical formulas used to calculate standard deviation and statistical power can be applied and ignoring robots in analysis (a problem unique to online settings). Online experiments allow for techniques like gradual ramp-up of treatments to avoid the possibility of exposing many customers to a bad (e.g., buggy) Treatment. With that ability, we discovered that it’s easy to incorrectly identify the winning Treatment because of Simpson’s paradox.
Online Experiments, Lessons Learned, IEEE 2007
Ronny Kohavi and Roger Longbotham
Practical Guide to Controlled Experiments on the Web: Listen to Your Customers not to the HiPPO, KDD 2007
Ron Kohavi, Randy Henne, Dan Sommerfield
The web provides an unprecedented opportunity to evaluate ideas quickly using controlled experiments, also called randomized experiments (single-factor or factorial designs), A/B tests (and their generalizations), split tests, Control/Treatment tests, and parallel flights. Controlled experiments embody the best scientific design for establishing a causal relationship between changes and their influence on user-observable behavior. We provide a practical guide to conducting online experiments, where end-users can help guide the development of features. Our experience indicates that significant learning and return-on-investment (ROI) are seen when development teams listen to their customers, not to the Highest Paid Person’s Opinion (HiPPO). We provide several examples of controlled experiments with surprising results. We review the important ingredients of running controlled experiments, and discuss their limitations (both technical and organizational). We focus on several areas that are critical to experimentation, including statistical power, sample size, and techniques for variance reduction. We describe common architectures for experimentation systems and analyze their advantages and disadvantages. We evaluate randomization and hashing techniques, which we show are not as simple in practice as is often assumed. Controlled experiments typically generate large amounts of data, which can be analyzed using data mining techniques to gain deeper understanding of the factors influencing the outcome of interest, leading to new hypotheses and creating a virtuous cycle of improvements. Organizations that embrace controlled experiments with clear evaluation criteria can evolve their systems with automated optimizations and real-time analyses. Based on our extensive practical experience with multiple systems and organizations, we share key lessons that will help practitioners in running trustworthy controlled experiments.
For feedback, send e-mail to ronnyk at live dot you know what.