Feedback Control of Real-Time Bidding Advertising Fatemeh Gheshlaghpour Advisors: Maryam Babazadeh Amin Nobakhti Sharif University of Technology, EE Department Apr. 2018
Overview Introduction • Business Model of the RTB Markets • – Key Roles in RTB Market – The Business Process of RTB Ad Delivery • Second-Price Sealed-Bid Auction Key Research Issues in the RTB • IPinYou Dataset • – Basic Statistics – User Feedback RTB Feedback Control System • – Bidding Strategy – Logistic Estimator – Actuator Control Issues of the Problem • – Reference and Feedback Signals – Model Results • 2/31
Introduction • In the previous generations of the display advertising the advertiser paid in two different platforms: – CPM (cost per mille) -> do not getting satisfactory numbers of clicks. – CPC (cost per click) -> being subjected to click frauds. • Internet users, on the other hand, faced with lots of irrelevant advertisements on their screens -> Ad-Blockers ! 3/31
Introduction • The Solution was the RTB in which the advertisers should bid for each impression based on some behavioral and contextual data. • The business process, including audience identification, auction and ad display, will be finished in exactly 10 to 100 milliseconds, and hence it is named "real-time bidding". • A basic problem for RTB bidding agents is to figure out how much to bid for an incoming bid request. 4/31
Business Model of the RTB Markets The key roles in RTB markets • ⁻ Advertiser ⁻ Demand side platform (DSP) is a platform that helps advertisers optimize their strategies. ⁻ Ad exchange (AdX) is an ad exchange market that matches the buyers and sellers . ⁻ Supply side platform (SSP) is a platform that helps publishers optimize the strategies. ⁻ Data management platform (DMP) is a platform that analyzes the cookie data of Internet users. ⁻ Publisher 5/31
Business Model of the RTB Markets The Business Process of RTB Ad Delivery • Myerson has proved that its optimal mechanism is second-price sealed-bid auction. The business process of RTB ad delivery.* * Yuan, Y ., Wang, F ., Li, J., & Qin, R. (2014). A survey on real time bidding advertising. Proceedings of the IEEE International Conference on Service Operations and Logistics, and Informatics (SOLI) , Qingdao, China (pp. 418-423). 6/31
Key Research Issues in the RTB Inventory Pricing and Channel Allocation • • In RTB markets, publishers and SSPs constitute the supply side of ad resources. Their key decisions, such as inventory pricing and multi-channel allocation of ad impressions, are major re search topics in literatures. • Business Model and Mechanism Design • Similarly working like the stock markets, AdX can bridge the gap in RTB markets by matching advertisers to publishers via real-time auctions. The existing works are focused on the design of the business model and auction mechanisms of AdXs and DSPs. 7/31
Key Research Issues in the RTB • Market Segmentation and Ad Performance Analysis Via designing the audience classification category and attribute • labels, DSPs can divide the Internet users into large amounts of niche markets with different kinds of demographic characteristics or shopping interests, and display best-matched ads accordingly. Bidding Behavior Analysis and Strategy • Optimization In RTB markets, advertisers and DSPs constitute the demand side of • ad resources, seeking to buy best-matched ad impressions via real- time auctioning and bidding. In literatures, bidding behavior analysis and strategy optimization for advertisers and DSPs attract intensive interests. 8/31
Second-Price Sealed-Bid Auction 9/31
Second-Price Sealed-Bid Auction • Nash Equilibrium: The only NE which does not need to know other players’ • private values is to bid truthfully, i.e. to bid same as the private value. 10/31
IPinYou Dataset • Fortunately, a leading Chinese advertising technology company iPinYou decided to release the dataset used in its global RTB algorithm competition in 2013. • The dataset includes logs of ad auctions, bids, impressions, clicks, and final conversions. These logs reflect the market environment as well as form a complete path of users’ responses from advertisers’ perspective. 11/31
The iPinYou data format.* * Zhang, W., Yuan, S., Wang, J., & Shen X. (2014). Real-time bidding benchmarking with ipinyou dataset. arXiv:1407.7073. 12/31
Basic Statistics 13/31
User Feedback • Some statistics of user feedback on campaigns 1458 and 3358 are shown bellow. 14/31
User Feedback 15/31
User Feedback 16/31
RTB Feedback Control System The traditional bidding strategy is represented as the bid calculator module in the DSP bidding agent. The controller plays as a role which adjusts the bid price from the bid calculator. Feedback controller integrated in the RTB system.* * Zhang, W., Yuan, S., & Wang, J. (2014). Optimal real-time bidding for display advertising. Proceedings of the 20 th ACM SIGKDD international conference on Knowledge discovery and data mining (KDD) , New York City, NY , USA (pp. 1077-1086). 17/31
Bidding Strategy • b θ b t ( ) t = 0 θ 0 18/31
Logistic Estimator •Logistic estimator predicts the CTR (a real value between 0 and 1) of an ad given a set of features. •Where f i (ad) is the value of the i th feature for the ad, and w i is the learned weight for that feature. 19/31
Logistic Estimator • In our experiment, all the features for LR are binary. • The weekday and hour feature are extracted from timestamps. • The floor price is processed by buckets of 0, [1,10], [11,50], [51,100] and [101,+ ∞ ). • We do not include the features of Bid ID, Log Type, iPinYou ID, URL, Anonymous URL ID, Bidding Price, Paying Price, Key Page URL. • In sum, we have 937,748 binary features for LR training and prediction. 20/31
Logistic Estimator Validation of LR for impressions with click=1 600 450 300 150 0 1458 2259 2261 2821 2997 3358 3386 3427 3479 #ectr > 0.5 #ectr < 0.5 Validation of LR for impressions with click=0 700000 525000 350000 175000 0 1458 2259 2261 2821 2997 3358 3386 3427 3479 #ectr > 0.5 #ectr < 0.5 21/31
Actuator • For example it can be chose to use: For instance in a PID controller we have: 22/31
Control Issues of the Problem • Same as any other control problems we should take care of these: • Reference Signal -> What is our controlling goal? • Feedback Signal -> Do we have access to this signal? • Model -> Can we have a static/dynamic model of the process? 23/31
Reference and Feedback Signals • The advertisers are provided with the data of the impressions. So we can define a feedback signal. • Reference signal is the output of an optimization problem. • In order to have smooth budget delivery we should set another constraint, too. 24/31
Reference and Feedback Signals Illustration of different budget pacing schemes with respect to the portion of the budget spent every time interval.* *Lee, Kuang-Chih, Ali Jalali, and Ali Dasdan. "Real time bid optimization with smooth budget delivery in online advertising." Proceedings of the Seventh International Workshop on Data Mining for Online Advertising . ACM, 2013. 25/31
Reference and Feedback Signals • Finally, we should optimize the following goal: 26/31
Model • As the time-constant of the dynamic model is not short enough we can not have it using the data. 27/31
Results • Here is the AWR signal for the campaign 2998 using impression-based PI controllers. 28/31
Results • Bellow you can see the overall control performance on AWR for the campaign 2998 using some different methods. Performance of different AWR controllers Method ref #Ad #clk total_cost AWR CPC CPM CTR rise-time settling-time rmse-ss awr_rand_pi 0.6 6055 9 200341 0.5896 20034.1 33086.8 0.0015 4 6 0.00196 control_awr_pi 0.6 6167 13 352707 0.5875 25193.3 57192.6 0.0021 252 2016 0.02587 awr_pi_constant 0.6 6053 7 209847 0.5943 26230.8 34668.2 0.0011 4 6 0.00203 awr_pi 0.6 6052 10 226075 0.5999 20552.2 37355.4 0.0017 4 6 0.00199 static_bid_awr_K 0.6 6191 5 142775 0.6136 23795.8 23061.7 0.0008 15 930 0.03102 29/31
Results • Bellow you can see the overall control performance on CPC for the campaign 2998 using some different methods. Performance of different CPC controllers Method ref #Ad #clk total_cost AWR CPC CPM CTR rise-time settling-time rmse-ss control_ecpc_pi 10000 1940 4 54510 0.1939 10902 28097.9 0.00206 5250 5250 0.06144 cpc_constant_pi 10000 153 0 10057 0.0152 10057 65732.1 0 134 134 0.00556 cpc_rand_pi 10000 153 0 10057 0.0153 10057 65732.1 0 134 134 0.00556 30/31
Thank you for your time… 31/31
Recommend
More recommend
