arXiv subjects

Guillaume Saint-Jacques

Publications and source records attributed to Guillaume Saint-Jacques.

9 recordsLinked to original sources

arXiv2022

Choosing an algorithmic fairness metric for an online marketplace: Detecting and quantifying algorithmic bias on LinkedIn

In this paper, we derive an algorithmic fairness metric from the fairness notion of equal opportunity for equally qualified candidates for recommendation algorithms commonly used by two-sided marketplaces. We borrow from the economic literature on discrimination to arrive at a test for detecting bias that is solely attributable to the algorithm, as opposed to other sources such as societal inequality or human bias on the part of platform users. We use the proposed method to measure and quantify algorithmic bias with respect to gender of two algorithms used by LinkedIn, a popular online platform used by job seekers and employers. Moreover, we introduce a framework and the rationale for distinguishing algorithmic bias from human bias, both of which can potentially exist on a two-sided platform where algorithms make recommendations to human users. Finally, we discuss the shortcomings of a few other common algorithmic fairness metrics and why they do not capture the fairness notion of equal opportunity for equally qualified candidates.

econ.GN↗

arXiv2022

Representation-Aware Experimentation: Group Inequality Analysis for A/B Testing and Alerting

As companies adopt increasingly experimentation-driven cultures, it is crucial to develop methods for understanding any potential unintended consequences of those experiments. We might have specific questions about those consequences (did a change increase or decrease gender representation equality among content creators?); we might also wonder whether if we have not yet considered the right question (that is, we don't know what we don't know). Hence we address the problem of unintended consequences in experimentation from two perspectives: namely, pre-specified vs. data-driven selection, of dimensions of interest. For a specified dimension, we introduce a statistic to measure deviation from equal representation (DER statistic), give its asymptotic distribution, and evaluate finite-sample performance. We explain how to use this statistic to search across large-scale experimentation systems to alert us to any extreme unintended consequences on group representation. We complement this methodology by discussing a search for heterogeneous treatment effects along a set of dimensions with causal trees, modified slightly for practicalities in our ecosystem, and used here as a way to dive deeper into experiments flagged by the DER statistic alerts. We introduce a method for simulating data that closely mimics observed data at LinkedIn, and evaluate the performance of DER statistics in simulations. Last, we give a case study from LinkedIn, and show how these methodologies empowered us to discover surprising and important insights about group representation. Code for replication is available in an appendix.

stat.AP↗

arXiv2021

Measuring Equity: Funnel Representation Measurement

We present a methodology to measure the gender representation for online product funnels. It is a part of the overall equity framework to better understand our products through funnel analysis. By leveraging the coarsened exact matching method from causal inference literature, we show that the funnel survival ratio metric we design can detect the representation differences inherent in our products. Understanding how big the representation differences are between different member groups, as well as understanding what explains them, is critical for fostering more equitable outcomes.

stat.AP↗

arXiv2021

An Outcome Test of Discrimination for Ranked Lists

This paper extends Becker (1957)'s outcome test of discrimination to settings where a (human or algorithmic) decision-maker produces a ranked list of candidates. Ranked lists are particularly relevant in the context of online platforms that produce search results or feeds, and also arise when human decisionmakers express ordinal preferences over a list of candidates. We show that non-discrimination implies a system of moment inequalities, which intuitively impose that one cannot permute the position of a lower-ranked candidate from one group with a higher-ranked candidate from a second group and systematically improve the objective. Moreover, we show that that these moment inequalities are the only testable implications of non-discrimination when the auditor observes only outcomes and group membership by rank. We show how to statistically test the implied inequalities, and validate our approach in an application using data from LinkedIn.

econ.EM↗

arXiv2020

Fairness through Experimentation: Inequality in A/B testing as an approach to responsible design

As technology continues to advance, there is increasing concern about individuals being left behind. Many businesses are striving to adopt responsible design practices and avoid any unintended consequences of their products and services, ranging from privacy vulnerabilities to algorithmic bias. We propose a novel approach to fairness and inclusiveness based on experimentation. We use experimentation because we want to assess not only the intrinsic properties of products and algorithms but also their impact on people. We do this by introducing an inequality approach to A/B testing, leveraging the Atkinson index from the economics literature. We show how to perform causal inference over this inequality measure. We also introduce the concept of site-wide inequality impact, which captures the inclusiveness impact of targeting specific subpopulations for experiments, and show how to conduct statistical inference on this impact. We provide real examples from LinkedIn, as well as an open-source, highly scalable implementation of the computation of the Atkinson index and its variance in Spark/Scala. We also provide over a year's worth of learnings -- gathered by deploying our method at scale and analyzing thousands of experiments -- on which areas and which kinds of product innovations seem to inherently foster fairness through inclusiveness.

cs.SI↗

arXiv2019

Using Ego-Clusters to Measure Network Effects at LinkedIn

A network effect is said to take place when a new feature not only impacts the people who receive it, but also other users of the platform, like their connections or the people who follow them. This very common phenomenon violates the fundamental assumption underpinning nearly all enterprise experimentation systems, the stable unit treatment value assumption (SUTVA). When this assumption is broken, a typical experimentation platform, which relies on Bernoulli randomization for assignment and two-sample t-test for assessment of significance, will not only fail to account for the network effect, but potentially give highly biased results. This paper outlines a simple and scalable solution to measuring network effects, using ego-network randomization, where a cluster is comprised of an "ego" (a focal individual), and her "alters" (the individuals she is immediately connected to). Our approach aims at maintaining representativity of clusters, avoiding strong modeling assumption, and significantly increasing power compared to traditional cluster-based randomization. In particular, it does not require product-specific experiment design, or high levels of investment from engineering teams, and does not require any changes to experimentation and analysis platforms, as it only requires assigning treatment an individual level. Each user either has the feature or does not, and no complex manipulation of interactions between users is needed. It focuses on measuring the one-out network effect (i.e the effect of my immediate connection's treatment on me), and gives reasonable estimates at a very low setup cost, allowing us to run such experiments dozens of times a year.

cs.SI↗

arXiv2019

A Method for Measuring Network Effects of One-to-One Communication Features in Online A/B Tests

A/B testing is an important decision making tool in product development because can provide an accurate estimate of the average treatment effect of a new features, which allows developers to understand how the business impact of new changes to products or algorithms. However, an important assumption of A/B testing, Stable Unit Treatment Value Assumption (SUTVA), is not always a valid assumption to make, especially for products that facilitate interactions between individuals. In contexts like one-to-one messaging we should expect network interference; if an experimental manipulation is effective, behavior of the treatment group is likely to influence members in the control group by sending them messages, violating this assumption. In this paper, we propose a novel method that can be used to account for network effects when A/B testing changes to one-to-one interactions. Our method is an edge-based analysis that can be applied to standard Bernoulli randomized experiments to retrieve an average treatment effect that is not influenced by network interference. We develop a theoretical model, and methods for computing point estimates and variances of effects of interest via network-consistent permutation testing. We then apply our technique to real data from experiments conducted on the messaging product at LinkedIn. We find empirical support for our model, and evidence that the standard method of analysis for A/B tests underestimates the impact of new features in one-to-one messaging contexts.

stat.AP↗

arXiv2019

Estimating Network Effects Using Naturally Occurring Peer Notification Queue Counterfactuals

Randomized experiments, or A/B tests are used to estimate the causal impact of a feature on the behavior of users by creating two parallel universes in which members are simultaneously assigned to treatment and control. However, in social network settings, members interact, such that the impact of a feature is not always contained within the treatment group. Researchers have developed a number of experimental designs to estimate network effects in social settings. Alternatively, naturally occurring exogenous variation, or 'natural experiments,' allow researchers to recover causal estimates of peer effects from observational data in the absence of experimental manipulation. Natural experiments trade off the engineering costs and some of the ethical concerns associated with network randomization with the search costs of finding situations with natural exogenous variation. To mitigate the search costs associated with discovering natural counterfactuals, we identify a common engineering requirement used to scale massive online systems, in which natural exogenous variation is likely to exist: notification queueing. We identify two natural experiments on the LinkedIn platform based on the order of notification queues to estimate the causal impact of a received message on the engagement of a recipient. We show that receiving a message from another member significantly increases a member's engagement, but that some popular observational specifications, such as fixed-effects estimators, overestimate this effect by as much as 2.7x. We then apply the estimated network effect coefficients to a large body of past experiments to quantify the extent to which it changes our interpretation of experimental results. The study points to the benefits of using messaging queues to discover naturally occurring counterfactuals for the estimation of causal effects without experimenter intervention.

cs.SI↗

arXiv2019

Testing for arbitrary interference on experimentation platforms

Experimentation platforms are essential to modern large technology companies, as they are used to carry out many randomized experiments daily. The classic assumption of no interference among users, under which the outcome of one user does not depend on the treatment assigned to other users, is rarely tenable on such platforms. Here, we introduce an experimental design strategy for testing whether this assumption holds. Our approach is in the spirit of the Durbin-Wu-Hausman test for endogeneity in econometrics, where multiple estimators return the same estimate if and only if the null hypothesis holds. The design that we introduce makes no assumptions on the interference model between units, nor on the network among the units, and has a sharp bound on the variance and an implied analytical bound on the type I error rate. We discuss how to apply the proposed design strategy to large experimentation platforms, and we illustrate it in the context of an experiment on the LinkedIn platform.

stat.ME↗