Auction Mechanisms for Efficient Advertisement Selection on Public Displays

1
Auction Mechanisms for Efficient Advertisement
Selection on Public Displays

• Terry Payne
• Ester David
• Alex Rogers
• Nicholas R. Jennings
• Matthew Sharifi
• Maria Karam

2
Outline

• Adaptive Advertising - who needs it?
• Situating and Justifying the approach
• The Use of Agents and Auctions
• Architecture
• Bidding Strategies
• Evaluation
• Simulate users to investigate mechanism
• Evaluate feasibility though live deployment
• Summary Future Work

3
Adaptive Advertising
Source BBC News Site 25th Aug 2006
4
Adaptive Advertising
Source BBC News Site 25th Aug 2006
5
Scheduling Dynamic, Personalised Advertisements

• Need to Identify Audience Members
• Not just one observer, but audience of observers
• Audience composition is dynamic and uncertain
• Requires ubiquitous technology with market
  penetration
• To identify observers within audience
• For pragmatic deployment in open environments
• Can detect users though their wireless devices
• Not ideal, but pragmatic

6
Scheduling Dynamic, Personalised Advertisements

• Game-based
• Current advertisements can affect audience
  behavior
• based on each observers viewing history
• Incomplete Knowledge
• No knowledge of observers preferences at startup
• No a priori knowledge of
• The number of adverts
• The number of observers
• Size or dynamic of audience (location specific?)
• Knowledge only of current/past advertisement
  schedule, and audience composition

7
Characterisation of the Problem Addressed

• To advertise
• a fixed number of adverts to (trackable)
  observers
• where adverts have fixed, homogonous duration
• and the number of observers is fixed
• Aim is to
• Minimise number of advertising cycles necessary
  to expose all advertisements to all observers
• Evaluation
• Simulation to evaluate model behavior
• Real-world deployment to assess practicality

8
Bluetooth Devices

• Tracking devices through Bluetooth
• Only partial market saturation
• But simple deployment with little initial cost!
• Short-range (2-5m) wireless network

• Discovery Mode
• Each Device broadcasts its own unique MAC
  identifier

9
Distributed Agents

• Agent approach adopted
• To encapsulate specific functionality
• Sensor Agents to detect observers Bluetooth
  Device
• Advertising Agents to represent each Advert
• Marketplace Agents which manages Public Display
• Coordination
• Utilise contract net protocol to announce
  upcoming advertising cycle and solicit bids
• Auction approach to allow agents to bid for
  (time-limited) access to the Public Display

10
Architecture
Sensor Agent
11
Auctions for resource allocation

• Each Advertising Agent is
• Self Interested - aims to
• Maximise its exposure to novel observers
• Minimise repeated exposure to existing observers
• Each agent bids value that reflects expected
  exposure to observers in the next advertising
  cycle
• Each Advertising Agent knows
• Advert and audience composition at end of current
  advertising cycle
• What observers have been present during its past
  advertising cycles, and for how long

12
Choice of Auction

• Repetitive, second-price sealed-bid auction
• Dominant strategy of truth revealing
• Obviates the need for expensive bidding
  strategies
• One-shot auctions simplify auction game
• Rapid conclusion of each auction
• Simple protocol assumed (contract net)
• Winner of auction
• Agent with highest bid pays second highest bid
  price

13
Valuation Strategy

• For each device d
• find the longest segment the device was present
  during past cycles won by aj
• Calculate minimum duration of unseen segment
• Sum unseen segments for each of the devices
  observed at end of cycle Ci

14
Example
Prior to Advertising cycle C1
No Advert Being Displayed
a ? end(C0)
v(a?, C1) 100 1 History0, 0, 0
?
v(a?, C1) 100 1 History0, 0, 0
?
v(a?, C1) 100 1 History0, 0, 0
?
15
Example
Prior to Advertising cycle C2
Observations a,4, b,2
?
a, c ? end(C1)
v(a?, C2) 001 1 History4, 2, 0
?
v(a?, C2) 101 2 History0, 0, 0
?
v(a?, C2) 101 2 History0, 0, 0
?
16
Example
Prior to Advertising cycle C3
Observations b,1, c,4
?
b, c ? end(C2)
v(a?, C3) 00.250 0.25 History4, 3, 4
?
v(a?, C3) 011 2 History0, 0, 0
?
v(a?, C3) 00.750 0.75 History0, 1, 4
?
17
Example
Prior to Advertising cycle C4
Observations a,2, b,4, c,4
?
a, b, c ? end(C3)
v(a?, C4) 00.250 0.25 History4, 3, 4
?
v(a?, C4) 0.500 0.5 History2, 4, 4
?
v(a?, C4) 10.750 1.75 History0, 1, 4
?
18
Evaluation

• Simulate users to investigate mechanism
• Investigate the auction behavior given simulated
  observer activity
• Vary behavior of audience
• Evaluate feasibility though deployment
• Can we track users using Bluetooth devices?
• What are the implementation issues?
• Preliminary User Trial
• How effective was the deployment for users?

19
Real-World Deployment

• Each installment uses
• Mac Mini
• With Bluetooth Wifi
• 23 LCD Screen
• Two displays deployed to date
• Three more to be installed in January, 2007
• Used to evaluate the use of such devices as
  proxies for users

20
Modeling Observer Activity

• Observer presence measured in discrete
  sample-intervals
• Duration of adverts are in whole sample-intervals
• An advert is fully-seen only when an observer has
  been present for the full advert duration
• Observers can arrive at any time with probability
  Parrive
• Observers can depart at any time with probability
  Pdepart
• An observer will leave if they have fully-seen an
  advert

21
Benchmark Selection Mechanisms

• Round Robin
• Random Selection

22
Benchmark Selection Mechanisms

• Round Robin
• Random Selection

23
Varying Device Number

• Evaluate performance in exposing all adverts for
  different numbers of observers/devices
• Fixed Parameters
• Parrival 50
• Pdepart 5
• 10 adverts of duration 6 time segments
• 10,000 experimental runs
• Variable Parameters
• Number of Devices, Nd ? 1, , 100

24
Varying Device Number
25
Varying Arrival

• Investigate performance with different observer
  arrival probability
• Fixed Parameters
• Number of Devices, Nd 50
• Pdepart 5
• 10 adverts of duration 6 time segments
• 10,000 experimental runs
• Variable Parameters
• Parrival ? 5, , 100

26
Varying Arrival
27
Varying Departure

• Investigate performance with different observer
  departure probability
• Fixed Parameters
• Number of Devices, Nd 50
• Parrival 50
• 10 adverts of duration 6 time segments
• 10,000 experimental runs
• Variable Parameters
• Pdepart ? 5, , 100

28
Varying Departure
29
Evaluating the Behaviour Space
Round Robin Selection
Auction-based Selection
30
Preliminary User Study

• To evaluate effectiveness in exposing content to
  users through two deployed screens
• 8 users with bluetooth devices over a 1 week
  period
• Each user was interviewed at the end of the trial
  period
• A touchscreen client was used to identify users
  and to recall adverts displayed to that user
• Users questioned about
• Adverts seen by user
• Advert relevance to user

31
Preliminary User Trial Results

• Metrics
• Recall - adverts remembered
• Relevant - adverts deemed relevant
• Non-Relevant - adverts noticed but not deemed
  relevant
• Extra - adverts also noticed but not recorded by
  BluScreen

32
Ongoing Work (1)

• Managing Deadlines
• Some adverts have limited lifetime
• Upcoming Seminars or Talks
• Concerts, Exhibitions etc.
• Need to schedule these adverts to
• Maximise exposure for a given budget
• Utilise full budget by deadline

33
Ongoing Work (2)

• Modify rationality assumption to consider Fair
  Allocation
• Current approach minimises exposure
• With unknown future audience, reduce expenditure
  of budget for each observer
• New approach considers differential exposure
• Agents bid based on comparison of exposure of
  its content with respect to exposure of that for
  other agents

34
Summary

• Weve introduced the idea of adaptive advertising
  using an agent environment
• Agents compete for advertising space based on
  their expected exposure, based on current and
  past exposure
• Observers are tracked using Bluetooth devices
• The approach has been evaluated empirically,
  through simulation and proof-of-concept
  deployment
• Results suggest that 36 fewer advertising cycles
  are needed to expose all observers to all adverts
• Deployment confirms that Frequent observers can
  be identified by their personal Bluetooth devices

35
Future Work

• Extend the evaluation to consider many screens,
  where the same advert may be observable in
  different locations
• Utilise observed behaviour to model return on
  bids, to improve economy
• Encourage formation of closed economy across
  multiple screens
• Investigate use of semantic tags to label adverts
  and sensors
• Acquire contextual evidence that could be used to
  build personal profiles for observers

36
Thank YouQuestions?
http//www.ecs.soton.ac.uk/research/projects/BluSc
reen