unintrusive customization techniques for web advertising
play

Unintrusive Customization Techniques for Web Advertising Marc - PowerPoint PPT Presentation

Unintrusive Customization Techniques for Web Advertising Marc Langheinrich Atsuyoshi Nakamura Naoki Abe Tomonari Kamba Yoshiyuki Koseki NEC Corporation, C&C Media Research Laboratories, Japan Overview Introduction Ad targeting


  1. Unintrusive Customization Techniques for Web Advertising Marc Langheinrich Atsuyoshi Nakamura Naoki Abe Tomonari Kamba Yoshiyuki Koseki NEC Corporation, C&C Media Research Laboratories, Japan

  2. Overview � Introduction � Ad targeting and current methods � Targeting with ADWIZ � The ADWIZ System � Architecture and basic interaction � The learning process � Experimental results Overview � Conclusions NEC Corporation 2

  3. Ad Targeting � Goal � Show advertisement only to desired target audience � Means 1.1 Ad Targeting � Dynamically select different ad for each Web site visitor � Targeting Parameters (Examples) � Browser, OS, time of day, country NEC Corporation 3

  4. Manual Ad Targeting � Method � Manually define targeting parameters for each ad � Advantages � Reaches only desired target audience 1.1 Ad Targeting � Predictable (How many ads will be shown?) � Disadvantages � Laborious to setup and maintain NEC Corporation 4

  5. Automated Ad Targeting � Method � Neural network learns user interests � Advantages � Fully automated 1.1 Ad Targeting � Disadvantages � User tracking violates privacy � Unable to predict number of times an ad is shown (contract constraints) NEC Corporation 5

  6. Targeting with ADWIZ � Automated Targeting 1.2 Targeting with ADWIZ � based on search keywords or page URI � Respects User Privacy � No user tracking necessary � Handles Contract Constraints � Supports minimum number of displays and other constraints NEC Corporation 6

  7. Control & Data Flow Return Return HTML HTML 2.1 Control & Data Flow Content Site Request Request Display Request Request Display Parse Parse page ad image page page ad image page User Extract Select Return Extract Select Return parameters ad GIF/JPG parameters ad GIF/JPG Ad Server NEC Corporation 7

  8. Basic Interaction 2.2 Basic Interaction User searches User searches for "car" for "car" Keyword-Based Ad Customization Keyword-Based Ad Customization NEC Corporation 8

  9. Basic Interaction 2.2 Basic Interaction System selects a car System selects a car related advertisement related advertisement Keyword-Based Ad Customization Keyword-Based Ad Customization NEC Corporation 9

  10. Basic Interaction 2.2 Basic Interaction User browses sports User browses sports section in directory section in directory System selects a sports System selects a sports related advertisement related advertisement Page-Based Ad Customization Page-Based Ad Customization NEC Corporation 10

  11. System Components Content Provider User Content Site 2.3 ADWIZ Architecture Ad System Ad Server Database Server Learning System Advertiser Administration Server Advertiser NEC Corporation 11

  12. Scheduling Ad Displays 2.4 Administrative Interface 1. Select advertisement 1. Select advertisement graphic to display graphic to display 2. Set minimum number 2. Set minimum number of necessary displays of necessary displays 3. What is the timeframe 3. What is the timeframe for showing the ad? for showing the ad? 4. Any special keyword 4. Any special keyword you want to reserve? you want to reserve? NEC Corporation 12

  13. Updating Display Weights 2.4 Administrative Interface Automatically updates Automatically updates every 3, 10 or 30 Minutes every 3, 10 or 30 Minutes NEC Corporation 13

  14. Inspecting the Weights 2.4 Administrative Interface List of ads and their probabilities of List of ads and their probabilities of being displayed for a certain keyword being displayed for a certain keyword NEC Corporation 14

  15. Inspecting the Weights 2.4 Administrative Interface List of keyword weights List of keyword weights per advertisement per advertisement NEC Corporation 15

  16. Inspecting the Weights 2.4 Administrative Interface List of page weights List of page weights per advertisement per advertisement NEC Corporation 16

  17. Inspecting the Weights 2.4 Administrative Interface List of advertisement List of advertisement weights per page weights per page NEC Corporation 17

  18. Keyword based Learning Inputs Maximize expected total click-through rate Advertisements A j Required displays h j 2.5 The Learning Process m n ∑ ∑ Toyota Camry 110 000 Toyota Camry 110 000 c ij d k i ij = = i 1 j 1 Cyberwing Golf 50 000 Cyberwing Golf 50 000 Keywords W i Usage rate k i 1. Show all required displays car 17 462 car 17 462 n ∑ = = k d h i ij j golf 34 921 i 1 golf 34 921 2. Weights sum up to 100% Click-through rate c ij car golf m ∑ = = d 1 ij j 1 Toyota Camry 7% 8% 7% 8% 3. No negative weights allowed Cyberwing Golf 1% 11% 1% 11% ≥ d 0 ij NEC Corporation 18

  19. Keyword based Learning Inputs Maximize expected total click-through rate Advertisements A j Required displays h j 2.5 The Learning Process m n ∑ ∑ Toyota Camry 110 000 Toyota Camry 110 000 c ij d k i ij = = i 1 j 1 Cyberwing Golf 50 000 Cyberwing Golf 50 000 Keywords W i Usage rate k i Output 1. Show all required displays car 17 462 car 17 462 n ∑ = = k d h Display rate d ij i ij j car golf golf 34 921 i 1 golf 34 921 Toyota Camry 91% 74% 2. Weights sum up to 100% 91% 74% Click-through rate c ij car golf m Cyberwing Golf ∑ = 9% 26% = d 1 9% 26% ij j 1 Toyota Camry 7% 8% 7% 8% 3. No negative weights allowed Total: 100% 100% Cyberwing Golf 1% 11% 1% 11% ≥ d 0 ij NEC Corporation 19

  20. Ad Selection Process Ad Server 2.5 The Learning Process HTTP "car" Extract Lookup Keyword Weights Administration Administration Ad Server Weights Ad Server P(A i |"car") Server Server HTTP A i Return Select GIF/JPG Ad Required Displays Database Click-Through Database Learning Learning Server & Usage rate Server System System NEC Corporation 20

  21. Experimental Setup � Keyword based � Methods compared � 32 Ads � Random Selection � 128 Keywords � Constraint-based Learning � Setup � Max-Click Method � Simulated keyword search 2.6 Experiments � Artificial User Interest Models Always select the advertisement Always select the advertisement � Repeated 1 million which had the highest click-through which had the highest click-through rate for given keyword in the past times rate for given keyword in the past � Averaged over 5 runs NEC Corporation 21

  22. Experimental Results Random Selection � Number of times 2.6 Experiments Advertisement ID NEC Corporation 22

  23. Experimental Results Random Selection Max-Click Method � � Number of times 2.6 Experiments Advertisement ID More total clicks � Fails to show more than � half of the ads NEC Corporation 23

  24. Experimental Results Random Selection Max-Click Method Constraint-based Learning � � � Number of times Number of times 2.6 Experiments Advertisement ID Advertisement ID More total clicks Increases click-through for � � all ads Fails to show more than � half of the ads Shows minimum number of � required displays NEC Corporation 24

  25. Experimental Results Max-Click Method Random Selection Constraint-based Learning � � � Max-Click and Random Number of times Number of times Method identical Max-Click Method better than Random Method 2.6 Experiments Random Method better than Max-Click Method Advertisement ID Advertisement ID More total clicks Increases click-through for � � all ads Fails to show more than � half of the ads Shows minimum number of � required displays NEC Corporation 25

  26. Experimental Results II Random Method Number of times 2.6 Experiments Advertisement ID NEC Corporation 26

  27. Experimental Results II Random Method Max-Click Method Number of times 2.6 Experiments Advertisement ID NEC Corporation 27

  28. Experimental Results II Random Method Max-Click Method Number of times Constraint-Based Learning 2.6 Experiments Advertisement ID NEC Corporation 28

  29. Experimental Results II Random Method Max-Click Method Number of times Constraint-Based Learning 2.6 Experiments More than 15% � improvement over Advertisement ID Max-Click Method NEC Corporation 29

  30. Conclusions � Current Ad Targeting Solutions � Manual: � Laborious � Automated: � Threatens privacy � Difficult to incorporate contract constraints 3.1 Conclusions � ADWIZ � Offers Automated Targeting � Respects User Privacy � Handles Contract Constraints NEC Corporation 30

  31. Future Work � Scaling Up � Thousands of keywords, pages, ads � Clustering techniques � Faster Learning for New Ads 3.2 Future Work � How to reuse previously learned parameters for new advertisements � Real-World Deployment � "Real" experiments NEC Corporation 31

  32. Related Work � Web Advertisement � effectiveness [Risden98] � alternative forms [Kohda96, Briggs97] � customization [Baudisch97] 3.3 Related Work � Privacy � user surveys [Rogers98, Cranor99] � cookies & profiling � FTC reports, EU Directive NEC Corporation 32

  33. ADWIZ Homepage For More Information http://www.ccrl.com/adwiz/ NEC Corporation 33

Recommend


More recommend