Computational Marketing at Groupon - JCSSE 2017
Download as PPTX, PDF2 likes415 views
The document discusses computational marketing at Groupon, outlining its growth and operational strategies since its inception in 2008. It highlights the use of data-driven marketing channels, campaign management, and attribution techniques to optimize user engagement and sales across its platform. Key challenges include keyword selection, performance tracking, and the complex interplay of multiple marketing channels to enhance customer conversion rates.
Computational Marketing at Groupon - JCSSE 2017
- 1. Computational Marketing at Groupon Clovis Chapman, PhD email: cchapman@groupon.com Research opportunities and challenges
- 2. About Me University College London, London, UK: Grid Computing Dell Cloud Research & Development Center, Dublin, IE: Cloud Computing Groupon, Seattle, USA: Big Data and Computational Marketing
- 3. About Groupon Founded November 2008. One deal per day in Chicago. Every day, a special deal for a business near you is sent by email as a coupon. Collective buying: If enough coupons were bought, the deal would become active, hence: “GROUPon”
- 4. Explosive Growth 2009 2013 40 K Active customers 41.7 M Active customers Spread to over 40 countries, over 200 million customers, and 40 million active customers.
- 5. Today “eCommerce Marketplace”: 200,000+ businesses, connecting tens of millions of customers on multiple sites across the world. Top 5 eCommerce brand in 2016 and top 25 mobile app by unique users in the US. Businesses include restaurants, health and beauty services, activity parks, etc. Groupon also provides goods (electronics, groceries, etc.) and travel deals.
- 6. How it works (local) Deal Centric: Groupon will promote “deals” that are typically near you. E.g. “20 minute Seattle Plane Ride” at a discount Bring voucher to merchant (using mobile phone) Groupon specific considerations: • Location Targeting: our deals are mostly location specific • Time Bound: New deals are up every day, only for a limited time • Variety: wide range of different type of experiences, and merchants (beauty, things to do, etc.)
- 7. PUSHPULL Over 200 Million customers internationally. Emails sent once or multiple time daily. Personalised and geolocated emails. Email Take deals to the customer Search Engine Optimization Search Engine Marketing ($) Display Advertising ($) Affiliate Marketing ($) Invite the customer to you. Present deals where she is looking. Marketing Channels
- 8. Talk Overview Computational Marketing at Groupon ● What is Computational Marketing? ● Campaign Management ● Marketing Channel Attribution ● Bidding ● In Summary
- 9. Find the “best match” between a given user in a given context and a suitable advertisement Broder, A. Z.
- 10. AdvertiserAdvertising platform Advertising Creative Landing Page Advertising context Advertiser domain Redirected Campaign setup: Creative, match criteria, bids Clicks 1 2 3 User
- 14. Display Advertising: context is everything! /
- 15. Marketing Optimization loop Publish / Update Campaigns (on Google, Bing, Facebook). • Several advertising platforms (Google, Facebook, Bing, etc.) • Hundreds of thousands of deals worldwide changing frequently • Fully automated optimization loop Analysis Monitor Performance (attribution) Select Deals / Content / Assets Define Match criteria (Keywords / Audience / location / time etc.) Compute Bids (Max cost per click)
- 16. Talk Overview Computational Marketing at Groupon ● What is Computational Marketing? ● Campaign Management ● Marketing Channel Attribution ● Bidding ● In Summary
- 17. Campaign Management 4 Core elements of an SEM ad campaign: ● Keywords to target ● Creative ● Landing Page ● Bid Other matching criteria important for Groupon: ● Location targeting Campaigns are created / refreshed multiple times daily for every deal. Automatic Keyword & Creative Generation. Use of Adwords API to push updates. Keywords Advertising Creative Search Engine Marketing Landing Page Click
- 18. Campaign Management Keyword Examples "massage groupon" "massage & wellness spa" "massage acupuncture" "massage addiction" "massage advertisement" "massage albuquerque" "massage and bodywork" "massage and facial" "massage and pedicure" "massage and manicure" "massage and spa" … “message spa" "message spas” “masge” “massahe” We advertise ~14 000 + massage related keywords alone in the U.S. 200 Million+ Keywords worldwide to manage Keywords advertised across multiple locations / Campaigns Billions of bid & performance records to manage: • How much did we bid • How many impressions per day • How many clicks per day • How many purchases per day • Etc. Keywords can be obtained from a variety of sources including deal and page meta-data: • Category • Merchant details • Template based generation
- 19. Campaign Management Keyword generation challenges Must be selective in our use of keywords Example: • “Pizza” • Unclear intent (Delivery? Recipe? …) • Minimum bid: $4.90 • 1 million daily searches Can quickly drain our budget! Lower price for lower search terms “market inefficiency” (Bartz K et al.) Cost per Click increases with Search volume (Bartz K et al.)
- 20. Campaign Management Research opportunities, examples Semantic Keyword Extraction algorithms & natural language processing: • Can use deal, merchant page, and related pages as corpus • Shallow NLP, N-Gram, TFIDF Cyclic keyword graph (Joshi P. et al. 2014) Graph theory (Joshi P. et al.) Use directed graph to compute relevance between terms
- 21. Talk Overview Computational Marketing at Groupon ● What is Computational Marketing? ● Campaign Management ● Marketing Channel Attribution ● Bidding ● In Summary
- 22. 2 days later $ t 12:00 Groupon Email 12:10 Google Search (Google Ad) 17:10 Mobile App 13:00 Facebook My devices
- 23. Rule based attribution models Multi-channel Attribution: Algorithms and Modeling Given a series of landings on the site from a variety of external sources, which should be associated with the purchase? These models differ greatly in terms of results and don’t take into account: • Complex interplay between channels (eg. Display > SEM > SEO may be a frequent path to conversion) • Differences between channels and product types (e.g. Travel deals convert better on desktops and will involve longer time periods allocated to research) • Static models that do not evolve over time Last click First click Linear Time Decay U-Shape
- 24. Multi-channel Attribution Multi-channel Attribution: Algorithms and Modeling Display Adv. Direct Organic Search (Google / Bing) Search Engine Marketing Berman, R . Beyond the last touch Traffic sources of conversion
- 25. Game theory / Shapley Value (Source: Shapley, L.): Markov Chains (Source: Bryl’, S.): Journey 1: C1 – C2 – C3 - $ Journey 2: C1 – END Journey 3: C2 – C3 - END Other Algorithms Multi-channel Attribution: research opportunities, examples Facebook Clicks Adwords Clicks Sales 5 0 $10 0 5 $5 5 5 $20 Facebook value: Adwords value: Cooperative effect: 2 1 20 – (2x5 + 1x5) = 5
- 26. Analytics Pipeline N-Tier architecture Homepage Deal page Shopping cart Frontend services Click Stream Tracking parameters / Cookies / Javascript / etc. Significant volumes of data must be analyzed : • Hundreds of thousands of user events per minute (views, purchases, cart, etc.). (Clickstream) • Service oriented architecture likely involves numerous loosely coupled independent services, powered by hundreds of servers • Service interfaces and event likely differ significantly Data Storage Clicks, views, purchases, etc. Service instance / server
- 27. Clickstream Analytics Deal Click groupon.com/deal? utm=UK_RTC_X Deal Purchase groupon.com/checkout?... Click Stream Logged in? No Cookie X IP address Y Browser thumbprint Z Logged in? Yes Cookie X IP address Y Browser thumbprint Z • Compute individual user journeys in order determine channel attribution based on selected model • Individual campaigns can be identified via tracking parameters set on the advertising platform • Cross device tracking: 67% of customers start shopping on one device and continue on another (Google) … User A User A
- 28. Supporting Architecture Lambda Architecture (Source: Marz, N. et al.) Data Stream Archive Job2Job1All events Recent events Batch Layer Speed Layer Worker nodes Results
- 29. Open Source Implementation Lambda Architecture (Source: Marz, N. et al.) HDFS All events Recent events Batch Layer Speed Layer Results
- 30. Talk Overview Computational Marketing at Groupon ● What is Computational Marketing? ● Campaign Management ● Marketing Channel Attribution ● Bidding ● In Summary
- 31. Bidding basics (SEM) Second price auctions (sealed) Keyword: “Food in Seattle” Other match criteria: User location Time of day Known user Etc. Max Cost per Click $5 $4 $3 $2 Pays second highest bid: $4.01
- 32. Bidding basics (SEM) At it’s simplest – bid can be calculated as: Avg(Revenue Per Click) / (Target Return on spend) How much should be willing to pay per click? For example – for “Food in Seattle”: - 100 people clicked the campaign - $500 was generated as per attribution method - Our target is: %200 – make $2 for every $1 spent on marketing - Avg(Revenue Per Click) / (Target Return on spend) = $5 / 2 = $2.5 Source: Ortec Mktg
- 33. Bidding Context (SEM) Many contextual modifiers can be applied – resulting in millions of combinations possible. Some examples: • Location targeting • Device • Time of day • Remarketing: Should we pay more for known users? Other challenges: do we have enough history for a given campaign? Can we exploit? • Historical performance of merchant / deal / category / etc ? • Performance across channels Challenges and Opportunities
- 34. Bidding Opportunities Research opportunities & examples Decision tree style approach: - Non linear supervised learning method - Can be used for conversion estimation (Wang J. et al) - Simple but suffers from high variance - Other approaches: collaborative filtering, etc. Source: Wang J. et al Keyword Clustering: (Source: Regelson et al.) - Compensate for sparse data and lack of history - Group campaigns with similar expected performance
- 35. In summary Computational marketing is a complex science involving a wide range of disciplines, provides numerous opportunities for research. • Information retrieval, Big data, Machine learning, Natural Language Processing, etc. • Financial scale is huge Many other topics not touched upon but equally challenging: • Landing page and customer journey optimization. Is the whole journey consistent? • A/B testing of creatives and landing pages • Re-targeting of known users
- 36. Thank you!