Big data panel slides
Download as PPT, PDF1 like429 views
The document describes a splash page experiment testing different button text and media variations on a splash page. It tested 4 button options ("Sign Up", "Learn More", "Join Us Now", "Sign Up Now") paired with 6 different media options. The results showed a 40.6% increase in email subscriptions, volunteers, and amount raised when compared to the original splash page. The new splash page performed better, obtaining 10 million email subscriptions, 1 million volunteers, and raising $200 million.
Big data panel slides
- 2. 2
- 5. Media Button
- 6. Splash page experiment Variations: Button: Media: • Sign Up 1. Get Involved Image • Learn More 2. Family Image • Join Us Now 3. Change Image • Sign Up Now 4. Barack’s Video 5. Springfield Video 6. Sam’s Video
- 9. Button: “Join Us Now”
- 10. Button: “Sign Up Now”
- 11. Splash page experiment Variations: Button: Media: • Sign Up 1. Get Involved Image • Learn More 2. Family Image • Join Us Now 3. Change Image • Sign Up Now 4. Barack’s Video 5. Springfield Video 6. Sam’s Video
- 13. Media: “Family”
- 14. Media: “Change”
- 18. Splash page experiment Variations: Button: Media: 1. Sign Up 1. Get Involved Image 2. Learn More 2. Family Image 3. Join Us Now 3. Change Image 4. Sign Up Now 4. Barack’s Video 5. Springfield Video 6. Sam’s Video
- 19. Splash page experiment results
- 20. Splash page experiment results
- 21. Splash page experiment results
- 22. Splash page experiment results Email Volunteers Amount Raised Subscriptions Original: 7,120,000 712,000 $143,000,000 +40.6% +2,880,000 +288,000 +$57,000,000 New: 10,000,000 1,000,000 $200,000,000
- 23. Splash page experiment results Email Volunteers Amount Raised Subscriptions Original: 7,120,000 712,000 $143,000,000 +40.6% +2,880,000 +288,000 +$57,000,000 New: 10,000,000 1,000,000 $200,000,000
- 24. Splash page experiment results Email Volunteers Amount Raised Subscriptions Original: 7,120,000 712,000 $143,000,000 +40.6% +2,880,000 +288,000 +$57,000,000 New: 10,000,000 1,000,000 $200,000,000
- 25. Email me: Follow me: dan@optimizely.com @dsiroker
- 26. 26
- 27. Modeling is using the data you have about a voter to make an informed judgment about: - whether they will vote - who they will vote for - what issues affect their vote - any other question you think has predictable (reproducible) behaviors
- 28. humans make models all the time, as we collect data & make informed judgments: 80% ? likelihood of voting for Obama
- 29. SURVEY GOP=94% (94/100) STATEWIDE GOP~94% (94,000/100,000) 50 years old white 56% 59% 63% 68% 85% 89% 94% % GOP
- 30. - knowledge management - master data management - data harmonization - voter relationship management OBAMA CAMPAIGN’S PROJECT NARWHAL
- 31. Slowly lowering the wall between online & offline data 1. Data providers William Alexander Lundry, 2. Targeted Display Ads Registered Republican 3. Facebook Apps 4. Volunteered Association
- 32. The Conservative Data Ecosystem RNC Data Trust Themis United In Purpose
- 33. 33
- 35. Big Political Data for the masses
- 37. Periodic to real time
- 38. Responds to DM Likelihood to vote Partisan Affiliation Top Issue Likelihood to unsubscribe Likelihood to volunteer Best Channel for Giving Receptiveness to treatment
- 39. 39
- 40. The world of survey research is changing rapidly.
- 41. The average phone survey response rate is around 20% and declining.
- 42. 27% of US households are cell only. (CDC)
- 43. 46% of American adults own a smartphone.
- 44. Telephone Consumer Protection Act of 1991
- 45. Online research tools are emerging.
- 46. Whether online works as a tool depends on what we want to measure.
- 47. As it gets harder to ask... What if we get better about listening?
- 50. Sentiment? I can go to twittersentiment.appspot.com and get an analysis.
- 51. Does “Who said it - Newt or Buzz Lightyear?” really count as a positive tweet?
- 52. “Gingrich had 6 percent more activity than the other candidates and the positive sentiment on him related to Super Tuesday is at 84 percent. Sentiment in general online conversation about him is only at 45 percent. So it seems his folks are Photo: Marc Grob for Time working the online world hard.” -From POLITICO “Playbook,” Super Tuesday (March 6 2012), quoting an email from a Washington-based public affairs consultant
- 54. Survey Sentiment Analysis • Landline bias • Online/activist bias • Contained universe • Variable universe • Concrete results • Subject to interpretation • “Snapshot” in time • Real-time, evolving • Message testing • Identify new trends
Editor's Notes
- #35: Talk about how historically, data has been tightly guarded by political parties and big organizations. And it’s also been limited by available tech. Technology to handle terabytes of data was expensive, and even out of reach of a lot of organizations.
- #36: But we’re seeing change. One great example of this is NationBuilder. Anyone who signs up can get access to the voter file. This is a slow but steady movement to a greater number of people having access to the data. Some examples: (1) Democratic Data Coop, (2) Catalist, (3) New Koch Brothers Data Organization on the right
- #37: But we’re also seeing a democratization of the technology. It used to be that if you wanted to handle terabytes of data, you had to go to either a Teradata, Netezza, or Oracle. More and more, organizations are using Hadoop + Hive and other technologies in order to process and query large sets of data. This also corresponds to the commoditization of the hardware required for these big clusters (and the cloud is becoming a better and better option).
- #38: Talk about starting to put more real time measures into models -- VF data updates once a month, maybe, but online data sources are dynamic and change on a day to day basis. Your voter contact system is producing new day thousands of times throughout a day. This shouldn’t be a process that moves once a month.
- #39: Shift of what we are modeling FOR
- #45: Calling more cell phones isn’t the answer because of TCPA - regulates when/how pollsters can call, makes cell interviews expensive and a major burden
- #47: Some online panels are great for understanding specific audiences but there’s still debate about their use and ability to be representative of large populations for something like a ballot test.
- #48: This is creating a massive new industry at the intersection of tech and opinion research. I’ve even had friends and colleagues tell me they’ve built models that can predict ballot share by looking just at online conversation and activity data.
- #49: Here’s one race that threw a wrench into the idea of monitoring online indicators as a way of predicting ballot share.
- #50: You can see in that 2010 election, sure Christine O’Donnell had a LOT of search traffic, and up until election day always had a higher search volume than her opponent. But did it matter? Nope.
- #51: It’s not hard to generate a sentiment analysi.. Take Twitter. Here, I can plug in and get an analysis fast. However, note in my example the first “positive” tweet...
- #52: Really? And look - nobody gets it perfect. That’s the point. It is very, very, very hard to get this right.
- #53: Sentiment analysis is working its way into the political conversation as a way, in addition to polls, to understand public opinion - but does it matter?
- #54: If Gingrich’s online buzz was so great ahead of Super Tuesday, why did he only win one of these ten states that day? (TELL ANECDOTE ABOUT NG CAMPAIGN)
- #55: The point isn’t to rain on the parade of sentiment analysis. I think traditional research’s days are numbered if it doesn’t evolve. But I think sentiment analysis has a lot to learn from traditional survey research for campaigns. (Explain pros and cons)