ABSTRACT
Restricted by the budget, the goal is to buy a set of ad impressions to reach as many targeted users as possible.
An online approach to the smooth budget delivery while optimizing for the conversion performance.
1 INTRODUCTION
RTB exchanges provide a technology for advertisers to algorithmically place a bid on any individual impression through a public auction. This functionality enables advertisers to buy inventory in a cost effective manner, and serve ads to the right person in the right context at the right time.
Demand-side platforms (DSPs) offer such a solution called real time bid optimization to help advertisers find the optimal bid value for each ad request.
The process of real time bid optimization tries to maximize the campaign performance goal under the delivery constraint within the budget schedule.
Performance goal: minimizing cost-per-click(CPC) or cost-per-action(CPA); maximizing click-through-rate(CTR) or conversion-rate (CVR). Under the constraint: smooth budget.
2 BACKGROUND AND RELATED WORK
Two situations won’t occur:
Premature Campaign Stop: Advertisers do not want their campaigns to run out of the budget prematurely in the day so as not to miss opportunities for the rest of day.
Fluctuation in Spend: Advertisers would like to be able to analyze their campaigns regularly and high fluctuations in the budget make the consistency of the results questionable.
Two main issues:
Traffic Issue: Depending on the target audience, the volume of the online traffic varies a lot throughout the day.
Performance Issue: The quality of the online traffic changes over the course of the day for different groups of audience.
3 ONLINE BID OPTIMIZATION
Two different classes of campaigns:
flat CPM campaign: CTR or CVR
dynamic CPM campaign: CPC or CPA
3.1 Smooth Delivery of Budget
S(t) is the dollar amount of money spent
reqs(t) is the number of incoming ad requests that satisfy the audience targeting constraints of the campaign
bids(t) is the number of ad requests that the campaign has bid on
imps(t) is the number of impressions of the campaign
pacing rate(t) = bids(t)/reqs(t)
win_rate(t) = imps(t)/bids(t)
3.2 Selection of High Quality Ad Requests - Flat CPM Campaigns
flat CPM campaign: CTR or CVR
win_rate(t) = imps(t)/bids(t)
goal is to simply select a set of ad requests to bid on considering the current time slot pacing rate
imps*(t) = s(t)/c*
For flat CPM campaigns that always submit a fixed bid price c* to RTB exchanges, the goal is to simply selct a set of ad requests to bid on considering the current time slot pacing rate.
In all, its CTR or CVR value is first estimated by the statistical model. If the predicted value is larger than the upper bound of the threshold, this ad request will be kept and the fixed bid price v* will be submitted to the RTB exchange. If the predicated value is smaller than the lower bound of the threshold, this ad request will be simply dropped without further processing.
3.3 Selection of High Quality Ad Requests - Dynamic CPM Campaigns
dynamic CPM campaign: CPC or CPA
For dynamic CPM campaigns are free to change the bid price dynamically for each incoming ad request, the goal is to win enough number of high quality impressions for less cost.
Base bid price ui = CVR*G
Thresholds beta1 and beta2:
Safe region: no under-delivery
Critical region: delivery is normal
Danger region: campaign has a hard time to find and win enough impressions.
For the case where our campaign is in the critical region, submitted ci_hat = u_i
For the case where our campaign is in the safe region, submitted ci_hat=theta*u_i, and paid ci, compute histogram of theta=ci/ci_hat.
For the case where our campaign is in the danger region,
If the audience targeting constraints are too tight and hence, there are not enough incoming ad requests selected for a bid, nothing could do.
If the bid price is not high enough to win the public auction even if we bid very frequently, boost the bidding price.
3.4 Estimation of CTR and CVR
Firstly, the estimated CVR provides a quality assessment for each ad request helping to decide on taking action on that particular ad request.
Secondly, the base bid price is set to be the estimated CVR multiplied by the CPA goal, which directly affects the cost of advertising.
The hierarchy starts with the root and continues layer after layer by advertiser category, advertiser, insertion order, package, lineitem, ad and creative.
4 PRACTICAL ISSUES
4.1 Cold Start Problem
4.2 Prevention of Overspending
4.3 Distributed Architecture
5 EXPERIMENTAL RESULTS
5.1 Comparison of Pacing Strategies
No comments:
Post a Comment