Machine learning for profit: Computational advertising landscape
4 likes617 views
The document outlines the landscape of computational advertising, highlighting the integration of machine learning, economics, and data-driven strategies for targeted ad delivery. It discusses the challenges in advertising, such as audience targeting, bidding, and pricing strategies, emphasizing the role of algorithms in optimizing ad performance. Additionally, it details the economic growth of the advertising sector and the importance of diverse technologies and methodologies in this domain.
![About me
• UB Alum, 2005 (CUBS)
• MIT Alum, 2010 (CBCL)
• Google, Senior Software Engineer
• AdWords pCTR modeling (McMahan et al. [2013])
• UI and position effect modeling
• Microsoft, Senior Software Engineer (AzureML [2016])
• Azure machine learning
• Nanigans, Principal data scientist
• Value modeling for Facebook/Twitter/Instagram
• Modeling stack for Real time display (RTB)
2](https://image.slidesharecdn.com/advertising-landscape-170414014046/85/Machine-learning-for-profit-Computational-advertising-landscape-2-320.jpg)
![Ranking: ad selection
• Ad selection can be framed as ranking problem given (ad,
display context, user) features
• Common metric: E[revenue] = p(Click) ∗ CPC
Position dependence
• Click through rates depend on display position.
• Factored click through rate model (Richardson et al. [2007])
p(Click|ad, context) = p(view|ad, context)∗p(click|view, ad, context)
• Factorization machine (Rendle [2012])
f (ad, context) = b+
i
(wa
i ai )+
j
(wc
j cj )+
i j
vi , vj ai cj
23](https://image.slidesharecdn.com/advertising-landscape-170414014046/85/Machine-learning-for-profit-Computational-advertising-landscape-30-320.jpg)
![Bidding: Factored value formulation
• Bidding value is the optimal strategy (Vickrey [1961])
• Value is factored based on a conversion funnel
• View → Click → Convert → Value
Factored formulation for downstream (revenue) optimization
VPI = E[V |A] ∗ P(A|Click) ∗ P(Click|Impression)
E[V |A] Value per action Linear/Log-Linear/Poisson regres
P(A|Click)orE[A|Click] Actions per click Logistic/Poisson Regression
P(Click|Impression) CTR Logistic regression
Fortunately, all of these regression can be implemented using a
general setup.
24](https://image.slidesharecdn.com/advertising-landscape-170414014046/85/Machine-learning-for-profit-Computational-advertising-landscape-32-320.jpg)
![Optimization: summary
Type Bidding
CPM value
CPC p(CTR) * value
CPA p(CTR) * p(Conversion) * value
OCPM E[value]
OCPC p(CTR) * E[value]
OCPA p(CTR) * p(Conversion) * E[value]
28](https://image.slidesharecdn.com/advertising-landscape-170414014046/85/Machine-learning-for-profit-Computational-advertising-landscape-37-320.jpg)
![Pacing types (Lee et al. [2013])
33](https://image.slidesharecdn.com/advertising-landscape-170414014046/85/Machine-learning-for-profit-Computational-advertising-landscape-44-320.jpg)
![Feedback control (Zhang et al. [2016])
34](https://image.slidesharecdn.com/advertising-landscape-170414014046/85/Machine-learning-for-profit-Computational-advertising-landscape-45-320.jpg)
Machine learning for profit: Computational advertising landscape
- 1. Machine learning for profit Computational advertising landscape Sharat Chikkerur sharat@alum.mit.edu Principal Data Scientist Nanigans Inc. 1
- 2. About me • UB Alum, 2005 (CUBS) • MIT Alum, 2010 (CBCL) • Google, Senior Software Engineer • AdWords pCTR modeling (McMahan et al. [2013]) • UI and position effect modeling • Microsoft, Senior Software Engineer (AzureML [2016]) • Azure machine learning • Nanigans, Principal data scientist • Value modeling for Facebook/Twitter/Instagram • Modeling stack for Real time display (RTB) 2
- 6. What is computational advertising Computational advertising is a new discipline that spans areas of computing, economics and machine learning Feature Discipline Web-scale audience Distributed computing Value estimation Machine learning Targeted delivery Information retrieval Personalized content Machine learning Economics (Explore Exploit) Recommendation system Dynamic pricing Game theory 5
- 8. Economic scale: worldwide • 11% annual growth rate • 260B by 2020 6
- 9. Economic scale: US • 17% annual growth rate, 100% annual growth for mobile 7
- 11. Market landscape
- 12. Entities • Advertiser • Publisher • Demand side platform • Supply side platform • Ad networks • Exchanges • Data aggregators 9
- 13. 10
- 14. © LUMA Partners LLC 2 Performance Video / Rich Media Targeted Networks / AMPs Horizontal Vertical / Custom Mobile ExchangesDSPs Publisher Tools Data Suppliers Ad Servers DMPs and Data Aggregators Measurement and Analytics Creative Optimization Agency Trading Desks Ad Networks Media Planning and Attribution Verification / Privacy Ad Servers Retargeting Media Mgmt Systems and Operations Sharing Data / Social Tools SSPs DISPLAY LUMAscape M A R K E T E R Tag Mgmt Agencies Denotes acquired company Denotes shuttered company 11
- 16. Search ads • Intent is well known • Search engine acts as a publisher, exchange and data aggregator • Targeting includes keywords, demographics, geo and user history 12
- 17. Display ads : Retargeting • Targets user who have almost converted elsewhere. • Users are tracked through cookie syncing between advertiser and exchange • Higher intentionality, time sensitive and very profitable 13
- 18. Display ads: Prospecting and brand • Typically used for brand advertising and optimized for reach • Primarily targeted using demographics • Low intentionality 14
- 19. Native ads • Sponsored content is mixed with native content • Very popular within social networks and portal sites • Requires good recommendation scheme 15
- 20. Multiformat • Sponsored content is mixed with organic content • Ads from other networks also displayed 16
- 21. Market landscape
- 22. Marketplace • Grauranteed delivery (Futures contract) • Direct advertiser publisher contract • Publisher guarantees delivery of impression to a given audience at a fixed price. • Publisher on the hook for shortfall • forecasting is the key to make a profit • Header bidding/First look (Options contract) • Advertiser pays a premium to get the first look at buy at a lower price than auction • Publishers get a better margin than the open exchange 17
- 23. Marketplace(cont.) • Non guaranteed delivery (Auctions) • Each impression goes through an auction in one or multiple exchanges • Publisher chooses the best bid • Less expensive 18
- 25. Challenges
- 26. The challenge of advertising • Find the best match between a given user in a given context and a suitable advertisement (when the advertisement is relevant in the context) 1 Context Channel Context Search Keywords, Query, Ads, Geolocation, user behavior Display Page content, user behavior, geolocation, demographics Native Page content, user behavior, social connections 1 Introduction to computational advertising https://web.stanford.edu/class/msande239/ 20
- 27. Problem domains Advertising can be thought off as a sequential decision process • Targeting: • Who do we show the advertising to and when • Ranking/Matching/Recommendation • Which set of ads do we show • Bidding • How much is this opportunity worth to the advertiser ? • Optimization • What’s the objective of this ad campaign? • Pricing • How much is this opportunity worth to the publisher ? • Budgeting & Pacing • When should we spend and how much to maximize yield 21
- 28. Challenge: Targeting and ranking
- 29. Targeting: audience selection • Targeting is usually manually specified based on campaign goals • Demographic (age, gender, income bracket) • User supplied • Credit card data • Shipping address etc. • Inferred demographics • Audience targeting can be framed as search/ranking task given (context, ad) pair • contrast with Ad selection = search(context, user) 22
- 30. Ranking: ad selection • Ad selection can be framed as ranking problem given (ad, display context, user) features • Common metric: E[revenue] = p(Click) ∗ CPC Position dependence • Click through rates depend on display position. • Factored click through rate model (Richardson et al. [2007]) p(Click|ad, context) = p(view|ad, context)∗p(click|view, ad, context) • Factorization machine (Rendle [2012]) f (ad, context) = b+ i (wa i ai )+ j (wc j cj )+ i j vi , vj ai cj 23
- 31. Challenge: Bidding and value estimation
- 32. Bidding: Factored value formulation • Bidding value is the optimal strategy (Vickrey [1961]) • Value is factored based on a conversion funnel • View → Click → Convert → Value Factored formulation for downstream (revenue) optimization VPI = E[V |A] ∗ P(A|Click) ∗ P(Click|Impression) E[V |A] Value per action Linear/Log-Linear/Poisson regres P(A|Click)orE[A|Click] Actions per click Logistic/Poisson Regression P(Click|Impression) CTR Logistic regression Fortunately, all of these regression can be implemented using a general setup. 24
- 33. Value estimation: Generalized linear models A generalized linear predictor specifies • A linear predictor of the form η(x) = wT x • A mean estimate µ • A link function g(µ) such that g(µ) = η(x) that relates the mean estimate to the linear predictor. This framework supports a variety of regression problems Linear regression µ = wT x Log-linear regression log(µ) = wT x Logistic regression log( µ 1−µ) = wT x Poisson regression log(µ) = wT x 25
- 34. Challenges: data issues • Optimization objectives can be different • (CPM, CPC, CPA, VPA etc.) • Downstream objectives may conflict with upstream objectives • high CTR impression/user may be low CVR impression/user • Data arrives at different rates, quantity (attribution windows) and have different maturity scales • CTR models have high volume and granular features • CVR models will have lower data volume • Need to handle delayed rewards 26
- 36. Optimization: types • Maximize volume • Bid high enough to guarantee delivery • Targeting should be narrow to avoid inefficiency • Spend needs to be paced during the day • Target cost (Average cost bidding) • Advertiser specifies CPA, CPC • Bid = p(CTR) ∗ p(Conversion|Clck) ∗ CPA ∗ pacing(t) • Pacing accounts for second price discount • Target yield (Revenue optimization) • Advertiser specifies yield on CPA, CPC • Bid = p(CTR) ∗ p(Conversion|Click) ∗ CPA yield ∗ pacing(t) 27
- 37. Optimization: summary Type Bidding CPM value CPC p(CTR) * value CPA p(CTR) * p(Conversion) * value OCPM E[value] OCPC p(CTR) * E[value] OCPA p(CTR) * p(Conversion) * E[value] 28
- 38. Challenge: Pricing (Exchange side)
- 39. Challenges: Pricing (Auctions) • First price sealed bid • Highest bidder wins • Unstable • Second price sealed bid • Highest bidder wins but pays second price • Stable • Incentive to bid true value • English auction (ascending price) • Seller increases price until a single bidder remains • Dutch auction (descending price) • Seller decreases price until a single bidder accepts it 29
- 40. Second price auction • Assume value is 100 • If you bid 100, net gain is 100 - cost • If you bid > 100, increase chances that cost > 100, does not affect chance of price < 100 • If you bid < 100, reduce chances of winning • Optimal strategy: bid 100, payout price cannot be affected • Model exposure fatigue Vickrey-Clark-Groves (VCG) auction • Maximizes total social value • Everybody pays their externality (TSV w/o bidder participation) - (TSV w/ participation) • Optimal strategy: all bidders bid their true value 30
- 41. Generalized second price auction • Rank by bid : cost(i) = bid(i + 1) • Rank by revenue: cost(i) = bid(i + 1) ∗ ctr(i + 1)/ctr(i) • Bidder always pays less than bid • Better CTRs generate discount • Not identical to VCG • Revenue may be higher than VCG 31
- 42. Challenge: Budgeting and pacing
- 43. Pacing • Purpose of pacing • Budget pacing - distributes spend over time period • KPI pacing - achieves target KPI (control vs. optimization) • Method of pacing • Throttling (allocation) - useful when bid adjustments are not allowed • Bid modification • Feedback control 32
- 44. Pacing types (Lee et al. [2013]) 33
- 45. Feedback control (Zhang et al. [2016]) 34
- 46. Explicit feedback control The bid is computed using ba(t) = b(t)exp{φ(t)} Here φ(t) is the control signal. Other approaches such as ba(t) = b(t)(1 + φ(t)) were found to be ineffective. 35
- 47. Controller • PID controller: e(tk) = xr − x(tk) φ(tk + 1) = λPe(tk) + λI k j=1 e(tj )∆tj + λD ∆e(tk) ∆tk Here, xr is the reference/desired KPI, x(tk), KPI at time tk. λP, λI , λd are conotroller gains. All control factors remain the same between updates. • Water-level based controller φ(tk + 1) = φ(tk) + γ(xr − x(tk)) 36
- 48. Performance 37
- 50. Subjective quality Ads blindness • Repeated exposures leads to learned blindness/adblock • We need to predict when not showing ads are better • Approaches • Thresholding based on scores • Utility function (lifetime value) • Learning from data Ads diversity • Similar search results lead to redundant information • Maximize overall utility (hard) • Maximize incremental utility (easier) • Tradeoff marginal utility with a diversity metric (easy) 38
- 51. Summary • Computational advertising invovles multiple subfields of computer science, machine learning, operations research, economics and game theory • Entire industrity is driven exclusively by data • Diverse eco-system with lots of opportunities for data scientists © LUMA Partners LLC 2016 Performance Video / Rich Media Targeted Networks / AMPs Horizontal Vertical / Custom Mobile ExchangesDSPs Publisher Tools Data Suppliers Ad Servers DMPs and Data Aggregators Measurement and Analytics Creative Optimization Agency Trading Desks Ad Networks Media Planning and Attribution Verification / Privacy Ad Servers Retargeting Media Mgmt Systems and Operations Sharing Data / Social Tools SSPs DISPLAY LUMAscape M A R K E T E R P U B L I S H E R P E O P L E Tag Mgmt Agencies Denotes acquired company Denotes shuttered company 39
- 52. References
- 53. AzureML. Azureml: Anatomy of a machine learning service. Journal of Machine Learning Reserach, 50, 2016. C. Karande, A. Mehta, and R. Srikant. Optimizing budget constrained spend in search advertising. WSDM ’13, page 697, 2013. Kc Lee, Ali Jalali, and Ali Dasdan. Real Time Bid Optimization with Smooth Budget Delivery in Online Advertising. arXiv preprint arXiv:1305.3011, pages 1–13, 2013. H Brendan McMahan, Gary Holt, D Sculley, Michael Young, Dietmar Ebner, Julian Grady, Lan Nie, Todd Phillips, Eugene Davydov, Daniel Golovin, Sharat Chikkerur, Dan Liu, Martin Wattenberg, Arnar Mar Hrafnkelsson, Tom Boulos, and Jeremy Kubica. Ad click prediction: a view from the trenches. Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 1222–1230, 39
- 54. 2013. doi: 10.1145/2487575.2488200. Steffen Rendle. Factorization machines with libFM. ACM Trans. Intell. Syst. Technol., 3(3):57:1–57:22, May 2012. ISSN 2157-6904. Matthew Richardson, Ewa Dominowska, and Robert Ragno. Predicting clicks: estimating the click-through rate for new ads. In Proceedings of the 16th international conference on World Wide Web, pages 521–530. ACM, 2007. William Vickrey. Counterspeculation, auctions, and competitive sealed tenders. The Journal of finance, 16(1):8–37, 1961. Weinan Zhang, Yifei Rong, Jun Wang, Tianchi Zhu, and Xiaofan Wang. Feedback control of real-time display advertising. Technical report, 2016. 39