Electronic Advertisements & Trading Agents. Enrique Areyan Viqueira

Transcription

1 Electronic Advertisements & Trading Agents Enrique Areyan Viqueira CS1951k/2951z, Spring 218

2 Electronic Advertisements (Ads) 3

3 Massive Market 32

4 Massive Market Example Google AdWords: 72 billion dollars worth of payper-click advertisements were purchased in 215. This was 97% of Google s overall income in

5 Ad Exchange & Ad Networks An ad exchange is a technology platform that facilitates the buying and selling of media advertising from multiplead networks An on-line advertising network or ad network is a company that connects advertisers to web sites that want to host advertisements 35

6 Ad Exchange Schematic Cartoon Ad Networks Publishers Impression Opportunity Ad Exchange Bid Bid Bid Advertisement Campaign Advertisement Campaign... Advertisement Campaign 39

7 Ad Network Challenge Fulfill advertising contracts at minimum cost while sustaining and attracting advertisers by achieving high quality targeting. Efficient Marketing Effectiveness 41

8 Trading Agent Competition - AdX Oneofmany Trading Agent Competition Games In a nutshell: the game reflects the challenges faced by an Ad Network as it bids for display ads opportunities. 42

9 Agents TAC AdX Bird s Eye View Implement an strategy for the Ad Network challenge, i.e., perform a bidding strategy. Game Server Simulates the behavior of users, web sites, advertisers, and an Ad Exchange for a number of days. 46

10 Next Lab: One Day Game We will play a simplified version of AdX game. In this simplified version we care only about: Campaigns Impressions Market 5

11 One Day Game - Campaigns Advertising campaigns are created by advertisers to promote their brands, and the AdNet s role is to carry them out. A campaign isdefined by: A market segment: a demographic(s) to be targeted A reach: the number of ads to be shown. A budget: the amount of money the advertiser will pay if this contract is fulfilled. 55

12 One Day Game - Impressions Market Each run of the game, a number of users from different Market Segments is simulated. Each user generates an impression opportunity. Each impression opportunity is allocated via a second price auction. 58

13 One Day Game - Mechanics Agent 1 Agent 2. Server Agent n 59

14 One Day Game - Mechanics Start of the Simulated Day Agent 1 Agent 2. Server Agent n 6

15 One Day Game - Mechanics Start of the Simulated Day Agent 1 Campaign 1 Agent 2. Agent n Campaign 2 Campaign n Server 61

16 One Day Game - Mechanics Agent 1 1 seconds Agent 2. Server Agent n 63

17 One Day Game - Mechanics 1 seconds Agent 1 Bid Bundle 1 Agent 2. Agent n Bid Bundle 2 Bid Bundle n Server 64

18 One Day Game - Mechanics Agent 1 Agent 2. Agent n Server Simulate impressions one at a time. 67

19 One Day Game - Mechanics Agent 1 Agent 2. Agent n Server Simulate impressions one at a time. For each impression: Run second price auction. Allocate and price impression. 68

20 One Day Game - Mechanics Agent 1 Agent 2. Agent n Server Simulate impressions one at a time. For each impression: Run second price auction. Allocate and price impression. Compute statistics for each agent. 69

21 One Day Game - Mechanics Agent 1 Agent 2. End Of Simulated Day Server Agent n Simulate impressions one at a time. For each impression: Run second price auction. Allocate and price impression. Compute statistics for each agent. 7

22 One Day Game Agent Evaluation Agents are evaluated by the profit obtained from their campaign as: profit( x)! = budget * min valid(! x) reach,1 price( x)! 73

23 A Minor Complication What are some of the main differences with a sealed-bid, one-shot auction studied in class? Agents can only bid in these auctions once! before their simulation begins. Repeated auction in unknown order! Complicated complimentary structure 77

24 One Day Game - Summary Your Agent s job is to determine: 1. Whereto bid? (which market segment(s)) 2. What to bid? (bid and limits) Knowing: 1. Distribution of impressions 2. Number and distribution of campaigns (but not exactly other agents campaigns!) 84

25 One Day Game Users Distribution Segment Avg. # of Users Male, Young, Low Income 1,836 Male, Young, High Income 517 Male, Old, Low Income 1,795 Male, Old, High Income 88 Female, Young, Low Income 1,98 Female, Young, High Income 256 Female, Old, Low Income 2,41 Female, Old, High Income 47 Distribution derived from data found in 85

26 Strategies for the One Day Game Computing the optimal strategy is a daunting task. What other options we have? Heuristics! Strategies with no optimality guarantee that (we hope) work well in practice. Can you think of one? 88

27 Heuristics Big picture: Predict, Optimize, Bid 89

28 Walrasian (Competitive) Equilibrium Roughly speaking, an equilibrium is given by allocation and prices such that: ü All agents are happy with their impressions at the given prices ü Supply meets demand 93

29 Walrasian Equilibrium (cont.) Idea: compute a Walrasian Equilibrium as a prediction of the outcome of the game Bid based on the Walrasian Equilibrium Warning: Well, actually, an exact Walrasian Equilibrium might not exist, so we just approximate it. Also, an approximation might be NP-hard to compute, so we heuristically compute it 97

30 Waterfall Heuristic The game is complicated by at least two factors: random order of impressions random campaigns of other agents. Suppose other agent s campaigns are known other agents bid truthfully (budget/reach) 98

31 Waterfall Heuristic (cont.) Waterfall heuristic solve for an allocation and a matrix of prices. An agent playing Waterfall strategy builds an internal representation of the market and runs the Waterfall algorithm. Prices for each market segment resemble a waterfall, where campaigns with higher budget/reach are served first and pay the price of the campaign with immediate lower budget/reach. 99

32 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 $5/5 = 1 7 M2 C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 1

33 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 $5/5 = 1 7 M2 C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 11

34 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 $5/5 = 1 7 M2 C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 12

35 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 $5/5 = 1 7 M2 55 C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 13

36 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 $5/5 = 1 Allocation M2 7 M2 55 C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 Prices (per impression) M2 14

37 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 $5/5 = 1 Allocation M M2 55 C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 Prices (per impression) M2 2 15

38 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 $5/5 = 1 Allocation M M2 55 C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 Prices (per impression) M2 2 16

39 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 $5/5 = 1 Allocation M M2 55 C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 Prices (per impression) M2 2 17

40 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 $5/5 = 1 Allocation M M2 55 C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 Prices (per impression) M2 2 18

41 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 $5/5 = 1 Allocation M M2 55 C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 Prices (per impression) M2 2 19

42 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 $5/5 = 1 Allocation M M C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 Prices (per impression) M2 2 11

43 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 $5/5 = 1 Allocation M M C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 Prices (per impression) M

44 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 $5/5 = 1 Allocation M M C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 Prices (per impression) M

45 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 $5/5 = 1 Allocation M M C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 Prices (per impression) M

46 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 $5/5 = 1 Allocation M M C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 Prices (per impression) M

47 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 $5/5 = 1 Allocation M M C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 Prices (per impression) M

48 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 $5/5 = 1 Allocation M M C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 Prices (per impression) M

49 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 35 $5/5 = 1 Allocation M M C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 Prices (per impression) M

50 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 35 $5/5 = 1 Allocation M M C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 Prices (per impression) M

51 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 35 $5/5 = 1 Allocation M M C2 4, $2 $4/2= 2 C3 15, $15 $15/15 = 1 Prices (per impression) M

52 Waterfall Algorithm - An example Impressions Campaigns 5 C1 5, $5 35 $5/5 = 1 Allocation M Price 7 M Market Segment M2 C2 4, $2 $4/2= 2 C3 15, $ $15/15 = 1 # Impressions Prices (per impression) M

53 Waterfall Heuristic (cont.) Waterfall takes you only so far! Waterfall works IF you know the exact configuration of the market. How would you use this algorithm if you don t know other agent s campaigns? and supply? 121

54 Image Credits Electronic Advertisement (Ads) webcapitaltop.com/servicios/campanas-display Google Ad Words Ad Exchanges Campaigns Heuristics Walrasian Equilibrium blog.mindsetwarrior.co/confidence-happiness/ 122