Research Fro rontier of f Real-Time Bid idding based Dis ispla - PowerPoint PPT Presentation

Research Fro rontier of f Real-Time Bid idding based Dis ispla lay Advertising Weinan Zhang University College London w.zhang@cs.ucl.ac.uk http://www0.cs.ucl.ac.uk/staff/w.zhang August 2015

Bas asic RTB Pro rocess Data User Information Management User Demography: Platform Male, 26, Student User Segmentations: Ad science, London traveling Page 1. Bid Request (user, page, context) 0. Ad Request Demand-Side RTB Platform Ad 2. Bid Response 5. Ad Exchange (ad, bid price) (with tracking) User 3. Ad Auction 4. Win Notice Advertiser (charged price) 6. User Feedback (click, conversion) 2

Model Bidding Str trategy Bid Request Bidding (user, ad, page, context) Strategy Bid Price • A function mapping from bid request feature space to a bid price • Design this function to optimise the advertising key performance indicators (KPIs) 3

Bidding Str trategy in in Pra ractice Bidding Strategy Feature Eng. Whitelist / Bid Request Blacklist (user, ad, Frequency Capping page, context) CTR / CVR Estimation Retargeting Campaign Budget Pricing Pacing Scheme Bid Price Bid Bid Landscape Calculation 4

Bidding Str trategy in in Pra ractice: : New Per erspective Bidding Strategy Bid Request Preprocessing (user, ad, page, context) CTR, Utility Cost Bid landscape Estimation Estimation CVR, revenue Bid Price Bidding Function 5

Dis iscussed Topics of f This Tal alk Fundamentals ls • CTR/CVR Estimation • Bid Landscape Forecasting • Bidding Strategies Advances • Arbitrage • Unbiased Training and Optimisation • Conversion Attribution

CTR/CVR Est stimatio ion • A seriously unbalanced-label binary regression problem – Negative down sampling, calibration • Logistic Regression [Lee et al. Estimating Conversion Rate in Display Advertising from Past Performance Data. KDD 12]

CTR/CVR Est stimatio ion • Follow-The-Regularised-Leader (FTRL) regression [McMahan et al. Ad Click Prediction : a View from the Trenches. KDD 13] Closed-form solution

CTR/CVR Est stimatio ion • Factorisation Machines [Oentaryo et al. Predicting response in mobile advertising with hierarchical importance-aware factorization machine. WSDM 14] – Explicitly model feature interactions – Empirically better than logistic regression – A new way for use ser pro rofilin iling • GBDT+FM [http://www.csie.ntu.edu.tw/~r01922136/kaggle-2014-criteo.pdf]

Deep Learning Models [our working project]

Bid Lan andscape Forecasting Auction Count Winning Probability Win bid Win probability: Expected cost:

Bid Lan andscape Forecasting Auction Winning Probability • Log-Normal Distribution [Cui et al. Bid Landscape Forecasting in Online Ad Exchange Marketplace. KDD 11]

Bid Lan andscape Forecasting • Price Prediction via Linear Regression [Wu et al. Predicting Winning Price in Real Time Bidding with Censored Data. KDD 15] – Modelling censored data in lost bid requests

Bidding Str trategies • How much to bid for each bid request? Bid Request Bidding (user, ad, page, context) Strategy Bid Price • Bid to optimise the KPI with budget constraint

Bidding Str trategies • Truthful bidding in second-price auction [Chen et al. Real-time bidding algorithms for performance-based display ad allocation. KDD 11] – Bid the true value of the impression • Non-truthful linear bidding [Perlich et al. Bid Optimizing and Inventory Scoring in Targeted Online Advertising. KDD 12] – With budget and volume consideration

Bidding Str trategies • Direct functional optimisation [Zhang et al. Optimal real-time bidding for display advertising. KDD 14] winning function CTR bidding function budget Est. volume • Solution: Calculus of variations 16

Optimal Bidding Str trategy Solu lution [Zhang et al. Optimal real-time bidding for display advertising. KDD 14] 17

Overall Performance – Optimising Cli licks or r Conversions iPinYou dataset [Zhang et al. Optimal real-time bidding for display advertising. KDD 14] 18

Dis iscussed Topics of f This Tal alk Fundamentals ls • CTR/CVR Estimation • Bid Landscape Forecasting • Bidding Strategies Advances • Arbitrage • Unbiased Training and Optimisation • Conversion Attribution

Dis iscussed Topics of f This Tal alk Fundamentals ls • CTR/CVR Estimation • Bid Landscape Forecasting • Bidding Strategies Advances • Arb rbit itrage • Unbiased Training and Optimisation • Conversion Attribution

Dis isplay Advertising In Intermediaries This work: Intermediary arbitrage algorithms in RTB display advertising. 21 [Zhang et al. Statistical Arbitrage Mining for Display Advertising. KDD 15]

Intermediary’s Statistical Arbitrage via RTB • Statistical arbitrage opportunity occurs, e.g., when (CPM) cost per conversion < (CPA) payoff per conversion 1000 impressions * 5 cent < 8000 cent for 1 conversion 22

Sta tatistical Arbitrage Min ining • Expected utility (net profit) and cost on multiple campaigns Bid request vol. Est. payoff winning function CVR estimation bidding function Cost upper bound Prob. of selecting Campaign i 23

Sta tatistical Arbitrage Min ining • Optimising net profit by tuning bidding function and campaign volume allocation Total arbitrage net profit Total cost constraint Risk control M-Step E-Step • Solve it in an EM fashion 24

M-Step: Bidding fu function optimisatio ion • Fix v and tune b () 25

E-Step: Campaign volume allo llocation • Multi-campaign portfolio optimisation Portfolio margin Portfolio margin mean variance where Net profit margin on each campaign 26

Campaign Portfolio Opti timisation Results 27

Dynamic Portfolio Optimisation 28

Onli line A/B Test on Big igTree ™ DSP • 23 hours, 13-14 Feb. 2015, with $60 budget each 29

Dis iscussed Topics of f This Tal alk Fundamentals ls • CTR/CVR Estimation • Bid Landscape Forecasting • Bidding Strategies Advances • Arbitrage • Unbia iased Tr Train inin ing an and Opti timis isatio ion • Conversion Attribution

Pro roblem of f Tra raining Dat ata Bia ias • Data observation process If win A bid Data Bid request observation • We want to train the model • But we train on the biased data [Zhang et al. Learning and Optimisation with Censored Auction Data in Display Advertising. AAAI 2016 Submission]

Unbiased Tra raining • Training target • Eliminate the data bias via importance sampling • Modelling winning probability via bid landscape

Unbiased Tra raining • Modelling winning probability via bid landscape • Only use observed impression data [UOMP] • Also use lost bid request data (censored data) [KMMP] n j : # {winning prices > b j } dj: # {winning prices = b j }

Exp xperimental Results • Winning probability estimation

Exp xperimental Results • CTR estimation: immediate performance improvement

Dis iscussed Topics of f This Tal alk Fundamentals ls • CTR/CVR Estimation • Bid Landscape Forecasting • Bidding Strategies Advances • Arbitrage • Unbiased Training and Optimisation • Co Conversio ion Att ttrib ributio ion

Conversion Attribution • Assign credit% to each channel according to contribution • Current solution: last-touch attribution [Shao et al. Data-driven multi-touch attribution models. KDD 11]

Multi-Touch Attribution • How to estimate the contribution of each channel? [Shao et al. Data-driven multi-touch attribution models. KDD 11] • A more general formula [Dalessandro et al. Casually Motivated Attribution for Online Advertising. ADKDD 11]

[Shao et al. Data-driven multi-touch attribution models. KDD 11]

Bidding in in Multi-Touch Attribution Mechanism • Current bidding strategy – Driven by last-touch attribution • A new bidding strategy – Driven by multi-touch attribution [Xu et al. Lift-Based Bidding in Ad Selection. ArXiv 1507.04811. 2015]

Val alue-based bidding v.s .s. . Lif ift-based bid idding

Val alue-based bidding v.s .s. . Lif ift-based bid idding • Comparison – Lift-based bidding help brings more conversions to advertisers – but its eCPA is higher than value-based bidding because of last-touch attribution • Lift-based bidding with multi-touch attribution could bring a better eco-system

Tak aking-home Messages • St Statis istic ical l Arb rbit itrage Mini ining: The internal auction selects the ad with highest arbitrage margin instead of the highest bid price. • Unbia iased Tr Train inin ing: Add the weight to each instance to eliminate the auction-selection bias. • Attrib ibutio ion and Bid iddin ing: Bidding proportional to the CVR lift instead of CVR value.

Computatio ional l Advertising Research in in Academia Disa isadvantages • Lack of data and online test platform • Lack of specific domain knowledge Advantages • Good at mathematic modelling • Focus on knowledge collection and communication • More research human resource

OpenBidder Pro roje ject: : www.openbidder.com • Online open-source benchmarking project – Bid optimisation, CTR estimation, Bid landscape etc. • Bridge academia and industry research on computational advertising 45