BEGIN:VCALENDAR
VERSION:2.0
PRODID:researchseminars.org
CALSCALE:GREGORIAN
X-WR-CALNAME:researchseminars.org
BEGIN:VEVENT
SUMMARY:David Banks (Duke University)
DTSTART:20211004T150000Z
DTEND:20211004T155500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/1
DESCRIPTION:Title: The Statistical Challenges of Computational Advertising\nby David
Banks (Duke University) as part of BIRS workshop: Statistical Methods for
Computational Advertising\n\n\nAbstract\nComputational advertising is a r
elatively young field\, but it touches on almost every aspect of statistic
s. This talk frames the purpose of this workshop\, and details some of th
e ways in which computational advertising intersects with statistics.\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/1/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Tim Hesterberg
DTSTART:20211004T160000Z
DTEND:20211004T165500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/2
DESCRIPTION:Title: Surveys and Big Data for Estimating Brand Lift\nby Tim Hesterberg
as part of BIRS workshop: Statistical Methods for Computational Advertisi
ng\n\n\nAbstract\nGoogle Brand Lift Surveys estimates the effect of displa
y advertising using surveys. Challenges include imperfect A/B experiments\
, response and solicitation bias\, discrepancy between intended and actual
treatment\, comparing treatment group users who took an action with contr
ol users who might have acted\, and estimation for different slices of the
population. We approach these\nissues using a combination of individual-s
tudy analysis and meta-analysis across thousands of studies. This work inv
olves a combination of small and large data - survey responses and logs da
ta\, respectively.\nThere are a number of interesting and even surprising
methodological twists. We use regression to handle imperfect A/B experime
nts and response and solicitation biases\; we find regression to be more s
table than propensity methods. We use a particular form of regularizatio
n that combines advantages of L1 regularization (better predictions) and L
2 (smoothness). We use a variety of slicing methods\, that estimate eithe
r incremental or non-incremental effects of covariates like age and gender
that may be correlated. We bootstrap to obtain standard errors. In contr
ast to many regression settings\, where one may either resample observatio
ns or fix X and\nresample Y\, here only resampling observations is appropr
iate.\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/2/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Art Owen (Stanford University)
DTSTART:20211004T170000Z
DTEND:20211004T175500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/3
DESCRIPTION:Title: Efficiency of Tie-Breaker Designs\nby Art Owen (Stanford Universi
ty) as part of BIRS workshop: Statistical Methods for Computational Advert
ising\n\n\nAbstract\nA company can offer some sort of incentive or gift to
its best customers. Those gifts have a cost and so it is worth investiga
ting their causal impact. Causal impact can be measured by regression dis
continuity analysis. Since the incentive is under the control of the comp
any they can also randomize around the cutoff in what is known as a tie-br
eaker design. Perhaps the top 5% of customers get the gift along with a r
andomly selected half of the next 10% of customers. We show that tie-brea
ker designs improve efficiency (versus regression discontinuity) for linea
r regression models and they have advantages in local linear regression as
well. Surprisingly there is little to gain from using a sliding scale of
gift probabilities instead of the levels 0%\, 50% and 100% that appear in
the tie-breaker design.\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/3/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Mamadou Yauck (Université du Québec à Montréal)
DTSTART:20211004T180000Z
DTEND:20211004T185500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/4
DESCRIPTION:Title: Computational Advertising: A Capture-Recapture Perspective\nby Ma
madou Yauck (Université du Québec à Montréal) as part of BIRS workshop
: Statistical Methods for Computational Advertising\n\n\nAbstract\nThis wo
rk is concerned with the analysis of marketing data on the activation of a
pplications (apps) on mobile devices. Each application has a hashed identi
fication number that is specific to the device on which it has been instal
led. This number can be registered by a platform at each activation of the
application. Activations on the same device are linked together using the
identification number. By focusing on activations that took place at a bu
siness location one can create a capture-recapture data set about devices\
, or more specifically their users\, that "visited" the business: the unit
s are owners of mobile devices\, and the capture occasions are time interv
als such as days. In this talk\, we will present a new algorithm for estim
ating the parameters of a capture-recapture model with a fairly large numb
er of capture occasions and a simple parametric bootstrap variance estimat
or.\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/4/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Ben Skraina (eBay)
DTSTART:20211004T200000Z
DTEND:20211004T202500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/5
DESCRIPTION:Title: Triumph and Tragedy in A/B tests: War Stories from Amazon\, eBay\, an
d Startups\nby Ben Skraina (eBay) as part of BIRS workshop: Statistica
l Methods for Computational Advertising\n\nAbstract: TBA\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/5/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Anru Zhang (Duke University)
DTSTART:20211004T203000Z
DTEND:20211004T205500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/6
DESCRIPTION:Title: High-order Clustering with Application in Click-through Prediction\nby Anru Zhang (Duke University) as part of BIRS workshop: Statistical M
ethods for Computational Advertising\n\n\nAbstract\nIn e-commerce\, predic
ting click-through for user-item pairs in a time-specific way plays an imp
ortant role in the online recommendation system. The click-through data ca
n be organized as an order-3 tensor\, where each entry is indexed by (user
s\, items\, time) and represents whether there is user-item interaction in
a time period. The users/items often exhibit clustering structures due to
similar preferences/attributes. It is important to do high-order clusteri
ng\, i.e.\, to exploit such high-order clustering structures. The high-ord
er clustering problem also arises from applications in genomics and social
network studies. The non-convex and discontinuous nature of the high-orde
r clustering problem pose significant challenges in both statistics and co
mputation.\n\nIn this talk\, we introduce a tensor block model and the com
putationally efficient methods\, high-order Lloyd algorithm (HLloyd)\, and
high-order spectral clustering (HSC)\, for high-order clustering. The loc
al convergence of the proposed procedure is established under a mild sub-G
aussian noise assumption. In particular\, for the Gaussian tensor block mo
del\, we give a complete characterization of the statistical-computational
trade-off for achieving high-order exact clustering based on three differ
ent signal-to-noise ratio regimes. We show the merits of the proposed proc
edures on the real online-click through data.\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/6/
END:VEVENT
BEGIN:VEVENT
SUMMARY:S. Yaser Samadi (Southern Illinois University)
DTSTART:20211004T210000Z
DTEND:20211004T212500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/7
DESCRIPTION:Title: Dimension Reduction for Vector Autoregressive Models\nby S. Yaser
Samadi (Southern Illinois University) as part of BIRS workshop: Statistic
al Methods for Computational Advertising\n\n\nAbstract\nThe classical vect
or autoregressive (VAR) models have been widely used to model multivariate
time series data\, because of their flexibility and ease of use. However\
, the VAR model suffers from overparameterization particularly when the n
umber of lags and number of time series get large. There are several stat
istical methods of achieving dimension reduction of the parameter space in
VAR models. In this talk\, we introduce the reduced-rank VAR model (Velu
et al.\, 1986\; Reinsel and Velu\, 2013) which restricts the rank of the p
arameter matrix in one direction\, and the envelope VAR model (Wang and Di
ng\, 2018) which is another solution to overcome the overparameterization
problem. Then\, we propose a new parsimonious VAR model by incorporating t
he idea of envelope models into the reduced-rank VAR. We show the strength
and efficacy of the proposed model by some simulation studies and an econ
omic dataset.\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/7/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Patrick LeBlanc (Duke University)
DTSTART:20211005T150000Z
DTEND:20211005T155500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/8
DESCRIPTION:Title: An Overview of Recommender System Theory\nby Patrick LeBlanc (Duk
e University) as part of BIRS workshop: Statistical Methods for Computatio
nal Advertising\n\n\nAbstract\nThis talk is a literature survey of approac
hes that have been taken to various kinds of recommender systems. I discu
ss both active and passive systems.\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/8/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Deborshee Sen (University of Bath)
DTSTART:20211005T160000Z
DTEND:20211005T165500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/9
DESCRIPTION:Title: Cross-Domain Recommender Systems\nby Deborshee Sen (University of
Bath) as part of BIRS workshop: Statistical Methods for Computational Adv
ertising\n\nAbstract: TBA\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/9/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Grace Yi (University of Western Ontario)
DTSTART:20211005T170000Z
DTEND:20211005T175500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/10
DESCRIPTION:Title: Unbiased Boosting Estimation for Censored Survival Data\nby Grac
e Yi (University of Western Ontario) as part of BIRS workshop: Statistical
Methods for Computational Advertising\n\n\nAbstract\nBoosting methods hav
e been broadly discussed for various settings\, especially for cases with
complete data. This talk concerns survival data which typically involve ce
nsored responses. Three adjusted loss functions are proposed to address th
e effects due to right-censored responses where no specific model is impos
ed\, and an unbiased boosting estimation method is developed. Theoretical
results\, including consistency and convergence\, are established. Numeric
al studies demonstrate the promising finite sample performance of the prop
osed method.\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/10/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Xuan Bi (University of Minnesota)
DTSTART:20211005T180000Z
DTEND:20211005T185500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/11
DESCRIPTION:Title: Improving Sales Forecasting Accuracy: A tensor factorization approac
h with demand awareness\nby Xuan Bi (University of Minnesota) as part
of BIRS workshop: Statistical Methods for Computational Advertising\n\n\nA
bstract\nDue to accessible big data collections from consumers\, products\
, and stores\, advanced sales forecasting capabilities have drawn great at
tention from many companies especially in the retail business because of i
ts importance in decision making. Improvement of the forecasting accuracy\
, even by a small percentage\, may have a substantial impact on companies'
production and financial planning\, marketing strategies\, inventory cont
rols\, supply chain management\, and eventually stock prices. Specifically
\, our research goal is to forecast the sales of each product in each stor
e in the near future. Motivated by tensor factorization methodologies for
personalized context-aware recommender systems\, we propose a novel approa
ch called the Advanced Temporal Latent-factor Approach to Sales forecastin
g (ATLAS)\, which achieves accurate and individualized prediction for sale
s by building a single tensor-factorization model across multiple stores a
nd products. Our contribution is a combination of: tensor framework (to le
verage information across stores and products)\, a new regularization func
tion (to incorporate demand dynamics)\, and extrapolation of tensor into f
uture time periods using state-of-the-art statistical (seasonal auto-regre
ssive integrated moving-average models) and machine-learning (recurrent ne
ural networks) models. The advantages of ATLAS are demonstrated on \\iv{ei
ght datasets} collected by the Information Resource\, Inc.\, where a total
of 165 million weekly sales transactions from more than 1\,500 grocery st
ores over 15\,560 products are analyzed.\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/11/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Nianqiao (Phyllis) Ju (Purdue University)
DTSTART:20211005T200000Z
DTEND:20211005T205500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/12
DESCRIPTION:Title: Towards Cost-Efficient A/B Testing\nby Nianqiao (Phyllis) Ju (Pu
rdue University) as part of BIRS workshop: Statistical Methods for Computa
tional Advertising\n\n\nAbstract\nOnline A/B tests play an instrumental ro
le for Internet companies to improve products and technologies in a data-d
riven manner. An online A/B test\, in its most straightforward form\, can
be treated as a static hypothesis test where traditional statistical tools
such as p-values and power analysis might be applied to help decision mak
ers determine which variant performs better. However\, a static A/B test p
resents both time cost and the opportunity cost for rapid product iteratio
ns. While some works try to tackle these challenges\, no prior method focu
ses on a holistic solution to both issues. We propose a unified framework
utilizing sequential analysis and multi-armed bandit to address time cost
and the opportunity cost of static online tests simultaneously. In particu
lar\, we present an imputed sequential Girshick test that accommodates bot
h streaming data and dynamic treatment allocation. The unobserved potentia
l outcomes are treated as missing data and are imputed using empirical ave
rages. Focusing on the binomial model\, we demonstrate that the proposed i
mputed Girshick test achieves Type-I error and power control with both a f
ixed allocation ratio and an adaptive allocation such as Thompson Sampling
through extensive experiments.\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/12/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Nathaniel Stevens (University of Waterloo)
DTSTART:20211005T210000Z
DTEND:20211005T215500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/13
DESCRIPTION:Title: Modern Design of Experiments for Computational Advertising\nby N
athaniel Stevens (University of Waterloo) as part of BIRS workshop: Statis
tical Methods for Computational Advertising\n\n\nAbstract\nDesigned experi
ments have long been regarded as the backbone of the scientific method use
d as the gold standard for causal inference. Although DOE has traditionall
y been applied in the realms of agriculture\, manufacturing\, pharmaceutic
al development\, and the physical and social sciences\, in recent years\,
designed experiments have become commonplace within internet and technolog
y companies for product development/ improvement\, customer acquisition/ r
etention\, and just about anything that impacts a business’s bottom line
. These online controlled experiments\, known colloquially as A/B tests\,
provide an especially lucrative opportunity for modern advertisers to unde
rstand market sentiment and consumer preferences. In this talk we provide
an overview of A/B testing and online controlled experiments and we descri
be ways in which these experiments and this context differ from that of cl
assical experiments. Although this modern “backyard” (as Tukey might c
all it) is somewhat under-appreciated in the field of industrial statistic
s\, we discuss several important and impactful research opportunities that
traditional industrial statisticians could and should get involved with.\
n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/13/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Yiyun Luo (University of North Carolina Chapel Hill)
DTSTART:20211006T150000Z
DTEND:20211006T155500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/14
DESCRIPTION:Title: Distribution-Free Contextual Dynamic Pricing\nby Yiyun Luo (Univ
ersity of North Carolina Chapel Hill) as part of BIRS workshop: Statistica
l Methods for Computational Advertising\n\n\nAbstract\n.: Contextual dyna
mic pricing aims to set personalized prices based on sequential interactio
ns with customers. At each time period\, a customer who is interested in p
urchasing a product comes to the platform. The customer's valuation for th
e product is a linear function of contexts\, including product and custome
r features\, plus some random market noise. The seller does not observe th
e customer's true valuation\, but instead needs to learn the valuation by
leveraging contextual information and historical binary purchase feedbacks
. Existing models typically assume full or partial knowledge of the random
noise distribution. In this paper\, we consider contextual dynamic pricin
g with unknown random noise in the linear valuation model. Our distributio
n-free pricing policy learns both the contextual function and the market n
oise simultaneously. A key ingredient of our method is a novel perturbed l
inear bandit framework\, where a modified linear upper confidence bound al
gorithm is proposed to balance the exploration of market noise and the exp
loitation of the current knowledge for better pricing. We establish the re
gret upper bound and a matching lower bound of our policy in the perturbed
linear bandit framework and prove a sub-linear regret bound in the consid
ered pricing problem. Finally\, we show the superior performance of our po
licy on simulations and a real-life auto-loan dataset\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/14/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Aiyou Chen
DTSTART:20211006T160000Z
DTEND:20211006T165500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/15
DESCRIPTION:Title: Robust Causal Inference for Incremental Return on Ad Spend with Rand
omized Paired Geo Experiment\nby Aiyou Chen as part of BIRS workshop:
Statistical Methods for Computational Advertising\n\n\nAbstract\nEvaluatin
g the incremental return on ad spend (iROAS) of a prospective online marke
ting strategy has become progressively more important as advertisers incre
asingly seek to better understand the impact of their marketing decisions.
Although randomized “geo experiments” are frequently employed for thi
s evaluation\, obtaining reliable estimates of the iROAS can be challengin
g as oftentimes only a small number of highly heterogeneous units are used
. In this talk\, we formulate a novel statistical framework for inferring
the iROAS of online advertising in a randomized paired geo experiment desi
gn\, and we propose and develop a robust and distribution-free estimator
“Trimmed Match” which adaptively trims poorly matched pairs. Using num
erical simulations and real case studies\, we show that Trimmed Match can
be more efficient than some alternatives\, and we investigate the sensitiv
ity of the estimator to some violations of its assumptions. This is joint
work with my colleague Tim Au at Google.\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/15/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Jason Poulos (Harvard)
DTSTART:20211006T170000Z
DTEND:20211006T173500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/16
DESCRIPTION:Title: Retrospective and Forward-Looking Counterfactual Imputation via Matr
ix Completion\nby Jason Poulos (Harvard) as part of BIRS workshop: Sta
tistical Methods for Computational Advertising\n\n\nAbstract\nI will discu
ss the matrix completion method for counterfactual imputation in standard
and retrospective panel data settings\, with applications to the social sc
iences. This talk is partly based on joint work with Andrea Albanese (LISE
R)\, Andrea Mercatanti (Bank of Italy)\, and Fan Li (Duke).\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/16/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Yi Guo (Duke University)
DTSTART:20211006T174000Z
DTEND:20211006T181500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/17
DESCRIPTION:Title: Multiparty Auctions without Common Knowledge\nby Yi Guo (Duke Un
iversity) as part of BIRS workshop: Statistical Methods for Computational
Advertising\n\nAbstract: TBA\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/17/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Maggie Mao
DTSTART:20211006T182000Z
DTEND:20211006T185500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/18
DESCRIPTION:Title: The Fight for Best Practices in Experimentation\nby Maggie Mao a
s part of BIRS workshop: Statistical Methods for Computational Advertising
\n\n\nAbstract\nOnline controlled experiments provide a scientific approac
h to understand how product changes affect user behavior and site performa
nce. It is also called the A/B test\, and it is a golden standard to testi
fy ideas\, quantify improvements\, and build causal relationships. At eBay
\, we have built a self-service experimentation platform to facilitate run
ning experiments at scale. However\, challenges raise when democratizing e
xperimentation and ensuring best practice (e.g.\, power analysis\, multipl
e testing\, metrics with highly skewed distribution\, etc.). In the talk\,
I will introduce the challenges we are facing and our current coping stra
tegies.\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/18/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Guy Aridor (Columbia University)
DTSTART:20211007T150000Z
DTEND:20211007T155500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/19
DESCRIPTION:Title: The Effect of Privacy Regulation on the Data Industry: Empirical Evi
dence from GDP\nby Guy Aridor (Columbia University) as part of BIRS wo
rkshop: Statistical Methods for Computational Advertising\n\n\nAbstract\nU
tilizing a novel dataset from an online travel intermediary\, we study the
effects of EU’s General Data Protection Regulation (GDPR). The opt-in r
equirement of GDPR resulted in 12.5% drop in the intermediary-observed con
sumers\, but the remaining consumers are trackable for a longer period of
time. These findings are consistent with privacy-conscious consumers subst
ituting away from less efficient privacy protection (e.g\, cookie deletion
) to explicit opt out—a process that would make opt-in consumers more pr
edictable. Consistent with this hypothesis\, the average value of the rema
ining consumers to advertisers has increased\, offsetting some of the loss
es from consumer opt-outs.\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/19/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Fiammetta Menchetti (University of Florence)
DTSTART:20211007T160000Z
DTEND:20211007T165500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/20
DESCRIPTION:Title: ARIMA Models and Multivariate Bayesian Structural Models for Causal
Inference from Sales Data\nby Fiammetta Menchetti (University of Flore
nce) as part of BIRS workshop: Statistical Methods for Computational Adver
tising\n\nAbstract: TBA\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/20/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Ernest Fokoue (University of Rochester)
DTSTART:20211007T170000Z
DTEND:20211007T175500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/21
DESCRIPTION:by Ernest Fokoue (University of Rochester) as part of BIRS wor
kshop: Statistical Methods for Computational Advertising\n\nAbstract: TBA\
n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/21/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Ron Berman (The Wharton School)
DTSTART:20211007T180000Z
DTEND:20211007T185500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/22
DESCRIPTION:Title: Latent Stratification for Advertising Experiments\nby Ron Berman
(The Wharton School) as part of BIRS workshop: Statistical Methods for Co
mputational Advertising\n\n\nAbstract\n.: We develop a new estimator of t
he ATE for advertising incrementality experiments that improves precision
by estimating separate treatment effects for three latent strata -- custom
ers who buy regardless of ad exposure\, those who buy only if exposed to a
ds and those who do not buy regardless. The overall ATE computed by averag
ing the strata estimates has lower sampling variance than the widely-used
difference-in-means ATE estimator. The variance is most reduced when the t
hree strata have substantially different ATEs and are relatively equal in
size. Estimating the latent stratified ATE for 5 catalog mailing experimen
ts shows a reduction of 36-57% in the posterior variance of the estimate.\
n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/22/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Michael Braun (Southern Methodist University)
DTSTART:20211007T200000Z
DTEND:20211007T205500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/23
DESCRIPTION:Title: The A/B Test Deception: Divergent Delivery\, Ad Response Heterogenei
ty\, and Erroneous Inferences in Online Advertising Field Experiments\
nby Michael Braun (Southern Methodist University) as part of BIRS workshop
: Statistical Methods for Computational Advertising\n\n\nAbstract\n.: Onl
ine advertising platforms provide tools to make it easy for advertisers to
conduct randomized experiments: so-called “A/B Tests”. In a targete
d advertising environment\, true A-B tests are comparing two different mix
tures of experimental subjects. We characterize how bias in the aggregate
estimate of the difference between two ads’ lifts is driven by the inte
rplay between heterogeneous responses to different ads and how platforms d
eliver ads to divergent subsets of users. We also identify conditions for
an undetectable “Simpson’s reversal\,” in which all unobserved types
of users may prefer ad A over ad B\, but experimental results lead the ad
vertiser mistakenly infer that users prefer ad B over ad A.\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/23/
END:VEVENT
BEGIN:VEVENT
SUMMARY:George Michailidis (University of Florida)
DTSTART:20211007T210000Z
DTEND:20211007T215500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/24
DESCRIPTION:Title: Regularized and Smooth Double Core Tensor Factorization for Heteroge
neous Data\nby George Michailidis (University of Florida) as part of B
IRS workshop: Statistical Methods for Computational Advertising\n\n\nAbstr
act\nTensor factorization based models have been extensively used in devel
oping recommender systems. In this talk\, we introduce a general tensor m
odel suitable for data analytic tasks for heterogeneous datasets\, wherein
there are joint low-rank structures within groups of observations\, but a
lso discriminative structures across different groups. To capture such com
plex structures\, a double core tensor (DCOT) factorization model is intro
duced together with a family of smoothing loss functions. By leveraging th
e proposed smoothing function\, the model accurately estimates the model f
actors\, even in the presence of missing entries. A linearized ADMM method
is employed to solve regularized versions of DCOT factorizations\, that a
void large tensor operations and large memory storage requirements. The ef
fectiveness of the DCOT model is illustrated on selected real-world exampl
es including image completion and recommender systems.\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/24/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Simon Mak (Duke University)
DTSTART:20211008T150000Z
DTEND:20211008T155500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/25
DESCRIPTION:Title: TSEC: a framework for online experimentation under experimental cons
traints\nby Simon Mak (Duke University) as part of BIRS workshop: Stat
istical Methods for Computational Advertising\n\n\nAbstract\nThompson samp
ling is a popular algorithm for solving multi-armed bandit problems. In ma
ny applications\, however\, the number of choices (or arms) can be large\,
and the data needed to make adaptive decisions require expensive experime
ntation. One is then faced with the constraint of experimenting on only a
small subset of arms within each time period\, which poses a problem for t
raditional Thompson sampling. To address this\, we propose a new Thompson
Sampling under Experimental Constraints (TSEC) method\, which makes use of
a Bayesian interaction model to model reward correlations between differe
nt arms. This fitted model is then integrated within Thompson sampling\, t
o jointly identify a good subset of arms for experimentation and to alloca
te resources over these arms. We demonstrate the effectiveness of TSEC in
two applications with arm budget constraints: the first on website optimiz
ation\, and the second for portfolio optimization.\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/25/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Sammy Natour
DTSTART:20211008T160000Z
DTEND:20211008T165500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/26
DESCRIPTION:by Sammy Natour as part of BIRS workshop: Statistical Methods
for Computational Advertising\n\nAbstract: TBA\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/26/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Edoardo Airoldi
DTSTART:20211008T170000Z
DTEND:20211008T175500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/27
DESCRIPTION:Title: Estimating Peer-Influence Effects Under Homophily: Randomized Treatm
ents and Insights\nby Edoardo Airoldi as part of BIRS workshop: Statis
tical Methods for Computational Advertising\n\n\nAbstract\n: Classical app
roaches to causal inference largely rely on the assumption of lack of inte
rference\, according to which the outcome of an individual does not depend
on the treatment assigned to others\, as well as on many other simplifyin
g assumptions\, including the absence of strategic behavior. In many appli
cations\, however\, such as evaluating the effectiveness of health-related
interventions that leverage social structure\, assessing the impact of pr
oduct innovations and ad campaigns on social media platforms\, or experime
ntation at scale in large IT organizations\, several common simplifying as
sumptions are simply untenable. Moreover\, being able to quantify aspects
of complications\, such as the causal effect of interference itself\, are
often inferential targets of interest\, rather than nuisances. In this tal
k\, we will formalize issues that arise in estimating causal effects when
interference can be attributed to a network among the units of analysis\,
within the potential outcomes framework. We will introduce and discuss sev
eral strategies for experimental design in this context centered around a
useful role for statistical models. In particular\, we wish for certain fi
nite-sample properties of the estimates to hold even if the model catastro
phically fails\, while we would like to gain efficiency if certain aspects
of the model are correct. We will then contrast design-based\, model-base
d and model-assisted approaches to experimental design from a decision the
oretic perspective.\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/27/
END:VEVENT
BEGIN:VEVENT
SUMMARY:David Banks
DTSTART:20211008T180000Z
DTEND:20211008T185500Z
DTSTAMP:20260422T185904Z
UID:BIRS-21w5508/28
DESCRIPTION:Title: Closing Remarks\nby David Banks as part of BIRS workshop: Statis
tical Methods for Computational Advertising\n\nAbstract: TBA\n
LOCATION:https://researchseminars.org/talk/BIRS-21w5508/28/
END:VEVENT
END:VCALENDAR