The document proposes a pay-per-attention (PPA) auction scheme for online advertisements that measures consumer attention to ads using mouse cursor movement analysis. It establishes that mouse cursor data can predict visual attention and presents a PPA auction model that generates higher expected revenue than pay-per-impression or pay-per-click schemes. An experiment uses mouse cursor logs and neural networks to accurately predict users' self-reported attention to ads with differences found across gender and age groups. The PPA scheme and attention prediction model provide scalable ways to quantify ad attention.
3. §The construct of attention has become the common
currency on the Web
§Payments in online ad auctions are typically derived
from click-through or pageview rates
§Various elements compete for user attention
4. §Existing online ad auctions adopt either pay-per-
impression (PPI) or pay-per-click (PPC) schemes
§Advertisers ultimately may care about knowing if
consumers notice the ad
§We propose a a pay-per-attention (PPA) scheme that
measures consumer attention to ads using mouse
cursor movements analysis
5. §Mouse cursor data have been used to predict user’s emotional
state and demographic attributes like gender and age
§Established hand-eye relationship and utility
of mouse cursor analysis as a low-cost and
scalable proxy of visual attention
§Tracking mouse cursor movements can be
used to predict attention on heterogeneous
web page layouts
7. Baseline Model
§Advertisers 𝑖 ∈ {𝟣, …, 𝑛}; single advertisement slot on sale
§𝑝: prob. a user notices the ad
§𝑥: ad’s CTR, conditional on ad being noticed
§𝑝𝑥: unconditional CTR
§For each 𝑖 ∈ 𝐼
§𝛾𝑖: prob. 𝑖 realises a sale conditional on its ad being clicked
§𝑞𝑖: prob. 𝑖 realises a sale conditional on ad being noticed but not clicked
§𝑣𝑖: value of a sale for bidder 𝐼
§(𝛾𝑖, 𝑞𝑖, 𝑣𝑖) are 𝑖’s private info; 𝑝 and 𝑥 are known to all advertisers
9. PPA Auction
§With a single item on sale, the standard benchmark in the Second-Price Auction
(SPA):
• Bidders simultaneously submit bid (𝑏1, 𝑏2, …, 𝑏 𝑛)
• The highest bidder 𝑖* gets the item and pays the second-highest bid (𝑏2nd)
§Existing formats differ in when such price is paid:
• PPI schemes: you pay as soon as you get the slot/win the auction (Expay = 𝑏2nd)
• PPC schemes: you pay if and only if your ad gets clicked (Expay = CTR ⋅ 𝑏2nd)
§The idea of the PPA scheme is to have advertisers pay only if the ad is noticed
(whether or not it’s clicked); If one could measure the probability 𝑝 that the ad is
noticed, then you pay 𝑝⋅𝑏2nd
11. PPA Auction
§Relaxing (ii): replace 𝑝 with an advertiser specific attention probabilities (𝑝𝑖) 𝑖∈𝐼
§The revenue ranking depends on the join distribution of the 𝑝𝑖‘s and 𝑉𝑃𝐴𝑖‘s (e.g.,
if 𝑝𝑖 and 𝑉𝑃𝐴𝑖 are positively correlated, the ExRevPPA = ExRevPPC > ExRevPPI
§Relaxing (i): Suppose bidders are naive, in that they follow the “standard bidding
strategies”, which only reflect the values made salient by the auction rules: 𝑏PPI=
𝑉𝑃𝐼𝑖 , 𝑏PPC= 𝑉𝑃𝐶𝑖 , 𝑏PPA= 𝑉𝑃𝐴𝑖
§As long as some advertisers are naive, the PPA does better than both its PPI and
PPC counterparts under weak conditions which are expected to hold in relevant
economic settings
12. Simulation 1: Uniformly Distributed Attention Prob.
(𝑎 = 𝟶)
§(𝑎 ∈ [𝟶, 𝟣] denotes the fraction of naive buyers)
§(𝜌 ∈ [𝟶, 𝟣] denotes the correlation between valuations 𝑣𝑖 and attention probabilities 𝑝𝑖)
(𝑎 = 𝟶.𝟧) (𝑎 = 𝟣)
§ 𝛾𝑖: and 𝑞𝑖 are i.i.d. draws from a uniform distribution over [𝟶, 𝟣]
§the 𝑝𝑖 are independently drawn from a uniform distribution over [𝟶, 𝟣]
§with probability (𝟣 - 𝜌), 𝑣𝑖 is an i.i.d. draw from a uniform over [𝟶, 𝟣𝟶𝟶]; with prob. 𝜌 instead of 𝑣𝑖 = 𝑝𝑖 ⋅𝟣𝟶𝟶
14. Experimental Design
§Brief transactional search task where participants were presented
with a predefined search query and the corresponding SERP, and were
asked to click on any element of the page that answered it best
§Between-subjects design with two independent variables:
• ad format (organic1 and direct display ads)
• ad position (top-left and top-right position)
§Each participant was exposed to a unique combination of query, ad
format, and ad position
1 Organic ads are only shown in the left part of Google
15. Experimental Design
§We used EvTrack1, an open-source JavaScript library that allows event
tracking via event listeners or via event polling
§Collected data from 3,206 participants, of age 18 − 66
§We collected ground-truth labels at post-task and asked the users to
what extent they paid attention to the ad using a 5-point Likert-type
scale: “Not at all” (1), “Not much” (2), “I can’t decide” (3), “Somewhat”
(4), and “Very much” (5)
1 https://github.com/luileito/evtrack
17. Experimental Design
§After excluding logs with incomplete mouse cursor data, we concluded
to 2,289 search sessions (45,082 mouse cursor positions):
• 763 correspond to the organic ad condition
• 793 correspond to the left-aligned direct display ad
• 733 correspond to the right-aligned direct display ad
§Ground-truth labels were converted to a binary scale (66% of positive
cases)
§Used 60-10-30 (%) disjoint stratified splits for training, validation, test
22. Main Findings
§We showed the PPA second-price auction inherits the same virtues (strategy-proofness
and efficiency) of its PPI and PPC counterparts
§In relevant economic environments (i.e. if at least some bidders are not fully sophisticated,
and if 𝑝𝑖 are positively correlated with the 𝑉𝑃𝐴𝑖), then it generates higher expected
revenues
§Our findings, suggest perceptual differences across the examined user groups i.e.
differences in the reliability of the attention predictions across gender and age groups
§We have introduced a scalable diagnostic technology that estimates user attention to ads
using raw mouse cursor data and a recurrent neural network which does not rely on
handcrafted features nor page-level information
23. Thank you for your attention!
Dataset Available Here
iarapakis
ioannis.arapakis@telefonica.com
http://iarapakis.github.io
luileito
luis.leiva@aalto.fi
https://luis.leiva.name/web/
antonio.penta@upf.edu
https://www.icrea.cat/Web/ScientificStaff/anton
io-penta-296186