Solutions and Challenges to Computation Advertising MediaV Peng 北冥乘海生.

Slides:



Advertisements
Similar presentations
Google News Personalization: Scalable Online Collaborative Filtering
Advertisements

How To Generate Targeted Traffic With Real-Time Bidding
Microsoft Media Network. 2 =+ Combining two great networks To build 1 industry-leading platform Microsoft Media Network The right impression. Every time.
UNDERSTANDING THE DIGITAL ECOSYSTEM. Copyright ©2013 The Nielsen Company. Confidential and proprietary. 2 CONSUMERS SPEND 12 HOURS PER MONTH WITH DIGITAL.
Modelling Relevance and User Behaviour in Sponsored Search using Click-Data Adarsh Prasad, IIT Delhi Advisors: Dinesh Govindaraj SVN Vishwanathan* Group:
■ Google’s Ad Distribution Network ■ Primary Benefits of AdWords ■ Online Advertising Stats and Trends ■ Appendix: Basic AdWords Features ■ Introduction.
1 Unified Approach: Display March PROPOSAL Creative Spec’s Delivery Traffic Billing The Method.
Digital Marketing Overview Tpugliese Adapted from Anton Koekemoer | April 2012.
The process of increasing the amount of visitors to a website by ranking high in the search results of a search engine.
Chapter Eight Traffic Building “A wealth of information creates a poverty of attention.” ~ Herbert Simon.
chapter 9 Communication McGraw-Hill/Irwin © 2004 The McGraw-Hill Companies, Inc., All Rights Reserved.
A Search-based Method for Forecasting Ad Impression in Contextual Advertising Defense.
Digital Marketing Paid Media New York Mayor Rudy Giuliani spent $60 million on his presidential campaign and won only one delegate.
Get new customers through the Internet By: Eng. Ahmed Sabry CEO, IT Vision Mobile: E-Marketing Campaign.
Search Engine Optimization (SEO)
Overview of Search Engines
Jennifer Ford.  Blog – A type of website or online journal that allows you to publish articles and updates that visitors.
The Google Display Network. Why Display Matters.
_______________________________________________________________________________________________________________ E-Commerce: Fundamentals and Applications1.
© 2006 Pearson Education Canada Inc Canadian Advertising in Action Chapter 12 Internet Communications.
Artur Strzelecki.  10 teams  10 non-profit organizations  6 students per team  2 weeks of developing campaigns  ~50€
Information Architecture and Web Advertising Xiaojing Feng.
PRIMISTA ONLINE MARKETING MADE EASY. Slide 2 Agenda Presentation Topics: 1.Introduction to Targeted Marketing 2.Ad Distribution Network 3.Primary Benefits.
3 rd Party Data & Audience Targeting © All rights reserved. 3 rd Party Data – Collected both online and offline by 3 rd party data companies such.
Buyer Advertising & UMass Boston Navigating the Changing Landscape of Recruitment Communications Presented to: November 18, 2014.
SOCIAL MEDIA OPTIMIZATION – GOOGLE ADSENSE, ANALYTICS, ADWORDS & MUCH MORE Ritesh Ambastha, iWillStudy.com.
CONCEPTUAL PRESENTATION ON ODINO.
Confidential 2008 Roadmap. confidential 2008 Solution Roadmap Main Themes The ChallengeOur Approach Actionable Analytics Non effective data analysis with.
About me: Michael Braems Freelancer Online Marketing AdWords Specialist.
Chapter 15 Using Digital Interactive Media William F. Arens Michael F. Weigold Christian Arens McGraw-Hill/IrwinCopyright © 2013 by The McGraw-Hill Companies,
Interactive Media The Basics. 2 Today’s Topic – Interactive Media Who we are and what we do –Strategy –Banner Advertising –Sponsorships –Search Engine.
InTopic Media, provides turnkey and white-label technology for in-text display advertising. We partner with Advertising Networks and Agencies to deliver.
Why Display matters. The Google Display Network The Google Display Network. Why Display Matters.
The Internet Industry Week Four. RISE OF THE INTERNET THE INTERNET – a global system of interconnected private, public, academic, business, and government.
Display & Remarketing What You Need to Know. PROPRIETARY AND CONFIDENTIAL / COPYRIGHT © 2013 BE FOUND ONLINE, LLC 2 WHAT IS DISPLAY?
Monetize Your Website Audience and Manage Digital Ad Campaigns with Admixer.Publisher, Built on the Powerful Microsoft Azure Platform MICROSOFT AZURE ISV.
ValueAd Inc. AdXpress ® Enterprise Ad Serving platform.
Canadian Advertising in Action, 6th ed. Keith J. Tuckwell ©2003 Pearson Education Canada Inc Elements of the Internet World Wide Web World.
Online Advertising Greg Lackey. Advertising Life Cycle The Past Mass media Current Media fragmentation The Future Target market Audio/visual enhancements.
Copyright © 2010 Pearson Education, Inc.Copyright © 2007 Pearson Education, Inc. Slide 1-1 ELC 200 Day 15.
Online Advertising Core Concepts are Identical to traditional advertising: –Building Brand Awareness –Creating Consumer Demand –Informing Consumers of.
Section 4 & 5 Review Google Adwords.  Contextual Targeting.
MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation 1 Dynamic Sensor Resource Management for ATE MURI.
Stephen Panjaitan PRESIDENT UNIVERSITY ORGANIZATION BEHAVIORAL.
Chapter Twelve Digital Interactive Media Arens|Schaefer|Weigold Copyright © 2015 McGraw-Hill Education. All rights reserved. No reproduction or distribution.
Triton Digital Presented to: Illinois Lottery. Engagement Network ADVERTISERS LEADING THE SPACE 2 OVERVIEW The Triton Digital® engagement network enables.
DIGITAL ADVERTISING Standard 4. THE ROLE OF DIGITAL ADVERTISING IS TO INCREASE SALES OR IMPROVE BRAND AWARENESS.
Predicting Winning Price In Real Time Bidding With Censored Data Tejaswini Veena Sambamurthy Weicong Chen.
Programmatic Buying Simplified
Chapter 1: Internet Marketing Foundations. Chapter Objectives Describe how computers and servers communicate to enable people to interact with webpages.
NA Sales Training 2007 The Digital Marketing Space.
Google Display Network. Targeting options.
1 DATA-DRIVEN SOLUTIONS. 2 KEYWORD-LEVEL SEARCH RETARGETING TARGET USERS BASED ON THEIR RECENT SEARCH HISTORY AND SEARCH QUERIES. A user performs a search.
Google Confidential and Proprietary Real Time Bidding Powers Data Driven Display Boris Kurschinski, Head of AdX Buyer Development DACH & Nordics.
We help businesses achieve online success! © All rights reserved. 8-digital.com - Proprietary and Confidential.
The Google Display Network. Why Display Matters..
The following Pitch Deck aims to help you sell Facebook Paid Ads to your clients. REMINDER: Get statistics and talking points on Facebook Paid Ads to.
Lecture 9 Communication.
Channel Connect for Mobile
User Modeling for Personal Assistant
Scalable Web Apps Target this solution to brand leaders responsible for customer engagement and roll-out of global marketing campaigns. Implement scenarios.
The Internet Industry Week Two.
Audience Ads Greece.
Scalable Web Apps Target this solution to brand leaders responsible for customer engagement and roll-out of global marketing campaigns. Implement scenarios.
9 Communication chapter McGraw-Hill/Irwin
Web Advertising and Cookies
E-commerce 2017 business. technology. society.
SEO Tutorial Search Engine Optimization
TARGET DISPLAY AUDIENCE
Presentation transcript:

Solutions and Challenges to Computation Advertising MediaV Peng 北冥乘海生

Brand Advertising Create a distinct favorable image

Direct Marketing Advertising that involves a "direct response" : buy, subscribe, vote, donate, etc, now or soon.

Audience targeting Which is more effective?

Advertising effectiveness exposure attention comprehension message acceptance retention purchase stageprinciples 1.1. Nature of the ad position 2.1 Don’t interrupt user’s task 2.2 Disclose the recommendation reason 2.3 Match user’s interests or needs 3.1 Convey message in user’s interest scope 3.2 Attention-aligned understanding barrier 4.1 Brand and creative recognition 4.2 Publisher/channel recognition 5.1 artistic quality 6.1 Set appropriate barrier for price- sensitive users selection interpre -tation attitude

Targeting technologies exposure attention comprehension message acceptance retention purchase stage targeting contextual (2.1, 3.1) re-targeting (2.2, 2.3, 3.1) behavioral (2.3, 3.1) geo (2.3, 4.1) demographical (2.3, 3.1, 6.1) channel (2.3, 3.1, 4.2) effectiveness applicable stages hyper-local (2.3, 4.1) look-alike (2.3, 3.1, 4.1, 6.1) group purchase (2.3, 4.1, 6.1)

ROI & computational advertising ROI of an advertising market –Investment = #X ⅹ CPX; –Return = #impression ⅹ CTR ⅹ click value = #impression ⅹ e(xpected)CPM CPM market: static eCPM CPC market: dynamic CTR, static click value CPA/CPS/ROI market: dynamic CTR and click value Key problem: eCPM estimation –Who: Various market place designs –How: Computational advertising

Ad network Ad network: –Connects advertisers to web sites that want to host advertisements –Estimate CTR and Matching ad with (context, user) by itself –Charge advertisers with CPC, CPM or other contracts –Hard to support various audience segments

Ad exchange Mission –Platforms that facilitate the bided buying and selling of online media ad inventory from multiple ad networks Key features: –Bridging ad with (context, user) by real time bidding (RTB) –Charge advertisers with real time bids on impressions.

Demand side platform Mission: –Allows digital advertisers to manage multiple ad exchange and data exchange accounts through one interface Key features: –Demand defined audience segments –Cross-media traffic acquisition –Evaluate advertiser ROI to support RTB

An ad flow illustration DSP 1 Media 1 Media 2 Media 3 Ad net 1 Ad net 2 Adx DSP 2 Agency 1 ATD Advertiser SSP Agency 1

Display ad market place

Computational advertising Main challenge –Find the best match between a given user in a given context and a suitable advertisement. Examples: –Context = Web search results -> Sponsored search –Context = Publisher page -> Content match, banners –Other contexts: mobile, video, newspapers, etc

Essentials A scientific sub-discipline at the intersection of –Web-scale search and text analysis –Information retrieval –Statistical modeling and machine learning –Numerical optimization –Microeconomics –Recommender systems The goal is to satisfy both quality and quantity requirements in ad serving.

Key messages A principled way to find the "best match" between a user in a context and a suitable ad. The financial scale for comp. advertising is huge –Small constants matter –Expect plenty of further research Advertising is a form of information –Adding ads to a context is similar to the integration problem of other types of information –Finding the “best ad” is an information retrieval problem with multiple, possibly contradictory utility functions

A skeleton of advertising system Ad retrieval User Context Ad candidates Freshness Popularity Quality … result Explore & Exploit User response optimization Relevance ranking Offline click feedback Online click feedback

Conceptual architect Ad server Storm Zookeeper CTR Model Audience Targeting Anti -spam Realtime CTR feedback User Profiling Realtime Billing Cache Thrift + Scribe

Comparison between several web scale applications SearchSponsored search contextual ads Display adsRecommen -dation Top CriteriarelevanceROIuser interest Other aspects application dependent qualitydiversity, freshness Index scale~billions~dozens of millions~millions~hundreds of millions User scaleshallow personalization~billions Contextualinsensitivesensitive Retrieval signals concentratedrichdemand side defined rich Downstream optimization not applicableapplicable

CTR prediction Challenges: –Dynamics: high churn rate of items, rapid evolution of user interests –Scale: Billions of training samples, and online features Offline CTR modeling –Click prediction model trained with offline data: –It is a regression problem rather than a ranking problem –Click feedback features for dynamic signal capturing –Logistic regression in widely adopted. Online CTR modeling

Logistic regression Objective –p (click | ad, user, context) = 1 / (1 + e -z ) (z = w 1 f 1 + w 2 f 2 + … + w N f N ) Optimization: L-BFGS, Trust-region Map-reduce –Mappers: derivative computation –Reducers: Parameter update Weakness: –Can hardly model non-linear behavior patterns

Click feedback Effective impression –EC: Expected Click, the sum of reference CTR for each impression. Normalized impression number Reference CTR –A CTR model for different presentation biases. Normalized CTR –COEC: Click Over EC = Click / EC. equally viewed?

Major bias features Position of ad space Size of ad space File size of creative Type of ad space (Homepage, channel homepage, content page, client, etc.) Type of creative (Picture, flash, rich-media) Mobile or desktop Browser & OS Time & Date

Multiple level click feedback Impression triple: (user, ctxt, ad) Multi-level descriptors: –User: cookie, gender, interest, city, … –Ctxt: url, topic, domain, channel, … –Ad: banner, solution, campaign, advertiser, topic Segment traffic by 1-3 descriptor combos, and use the statistics as features. –Current: ck, (ck, banner), (city, banner), … –Future: (interest, ad), (domain, ad topic), …

CTR smoothing Goal: Reliably estimate CTR with sparse data. A Empirical Bayesian treatment: Click generation: Treat as a random, and apply a conjugate prior to regularize it: How to make use of item hierarchy for better regularization?

Evaluation PR curve ROC curve AUC

Apply linear model in retrieval Baseline ad selection methodology –1. Retrieve all feasible candidates. –2. Apply ranking model to each feasible one and select top K. How can we handle a large feasible set? –If the model is linear, the ad selection problem can be view as a two step process: Calculate S = F ⅹ W (S :scores, F : features, W : weights) Select top K socres from S. –With elegant index and algorithm design, We can merge the two steps and selected top K results in a parsimonious way.

Weighted AND algorithm [A. Broder, etc. 2003] Index –Enhanced reverted index, with keeping all the documents sorted by their contribution to the score on the index key. Retrieval –Estimate the upper bound of a (query, document) given the visited features on the fly, and prune candidates as early as possible. –The algorithm can make sure infeasible candidates not touched in retrieval, and only part of the weighted sum operator need to be calculated.

Click value estimation Applications: –Real time bidding for DSP –Bidding tool for ad network –Smart price Challenges: –Extremely sparse training data –Advertiser-dependent behavior modes Principle in click value prediction –Aim at a smooth modeling approach with small variance but relatively large bias

Stream computing for online click feedback: S4 as an example

Explore & Exploit Goal: –Assign tail items adequate changes to obtain sufficient statistics –A multi-arm bandit (MAB) problem. Approaches: –ε–greedy: A small random traffic for exploration –Upper Confidence Bound: Play the arm with largest upper bound of expected return [J. Audibert, 2006] Challenge: –Large number of items to be explored.

ROI with quantity constraint Goal: –Optimize ROI w.r.t. a given ad spend in one or more ad networks. Examples: SEM, ATD Two key problems: Set solutions on appropriate traffic segments (keywords in SEM, labels in ad network). Bid on each solution to optimize the entire ROI. State-of-the-art: Portfolio selection theory. [H. Markowitz, 1959]

Portfolio and Efficient frontier Treat the group of keywords/solutions as a portfolio of stocks, and find the efficient frontier to optimize ROI given an ad spend.