ºÝºÝߣ

ºÝºÝߣShare a Scribd company logo

Demo David Sun AdFlow

?
?
P
Peng Sun

AdFlow is a bidding system that uses machine learning algorithms to deliver personalized ads to target users in real time. It quantifies user preferences and measures affinity between users and products using text analysis. Advertisers provide bids that are matched to detected user events and interests through real-time bidding. The system was tested on 90k tweets generating 1-10 word tweets at 100 tweets/sec with 500 advertisers updating bids. It generated $500k in revenue per hour on average.

1 of 10
AdFlow A bidding system for ads suppliers to deliver personalized ads Dr. David P. Sun Insight Data Engineering Fellow Fall 2016, New York
My motivation ? Find the target users for ads suppliers? ? Push personalized products in real time? ? Receive bids from product providers? ? Real time dashboard update and ad-exchange? ?
Design the product to answer: ? What is the right product? ? Who are the target users? ? Where are the target users? ? When is the right time to Deliver?
Quantify user preference and user-product affinity Corr=U dot V [¡®I¡¯, ¡®love¡¯, ¡®mac¡¯, ¡®cheese¡¯] Tweets [¡®Nike is an American multinational corporation that is engaged in the design¡­¡¯] Ad U=[.64, .52, .22, -.21, ¡­] V=[.64, .52, .22, -.21, ¡­] Measure the distance between text messages 1.0 like -1.0 dislike
Detect ads pushing events Use groups of messages to describe a person OR Bob 19:00:05, I want to have some pizza 19:00:12, mozzarella sounds good 19:33:24, I will meet you at 9pm
Realtime bidding with a complete set of bids from bidders Trade some costs from communication for completeness Bids Stream Events Stream Updating Bids pid bid price pid00000 1.54 pid00001 16,99 pid00002 7.34 pid00003 2.31 uid time dictionary of candidates pid:score 10057 timestamp {pid00000:0.91, pid00305: 0.94, ¡­ 58738 timestamp {pid00006:0.95, pid00134: 0.99, ¡­
Preliminary pipeline bookkeeping biddings detect events match bids extract features
Results 90k twitters generate a tweet flow of 1-10 words at 100 tweets/sec 500 advertisers updating the bid flow 18k-word vocabulary Each word vector is of dimension 10 Bids follows a Pareto distribution Generating a revenue of 500k per hour on average
About me Scientist PhD Engineer Entrepreneur TransportCalc 10k+ lines in c++ Insight data engineer fellow business sense anisotropic turbulence bid data tools build products solve problems modeling experiment visualize machine learning algo & libs know customers & create values
Demo http://www.adsflow.cc

More Related Content

Demo David Sun AdFlow

  • 1. AdFlow A bidding system for ads suppliers to deliver personalized ads Dr. David P. Sun Insight Data Engineering Fellow Fall 2016, New York
  • 2. My motivation ? Find the target users for ads suppliers? ? Push personalized products in real time? ? Receive bids from product providers? ? Real time dashboard update and ad-exchange? ?
  • 3. Design the product to answer: ? What is the right product? ? Who are the target users? ? Where are the target users? ? When is the right time to Deliver?
  • 4. Quantify user preference and user-product affinity Corr=U dot V [¡®I¡¯, ¡®love¡¯, ¡®mac¡¯, ¡®cheese¡¯] Tweets [¡®Nike is an American multinational corporation that is engaged in the design¡­¡¯] Ad U=[.64, .52, .22, -.21, ¡­] V=[.64, .52, .22, -.21, ¡­] Measure the distance between text messages 1.0 like -1.0 dislike
  • 5. Detect ads pushing events Use groups of messages to describe a person OR Bob 19:00:05, I want to have some pizza 19:00:12, mozzarella sounds good 19:33:24, I will meet you at 9pm
  • 6. Realtime bidding with a complete set of bids from bidders Trade some costs from communication for completeness Bids Stream Events Stream Updating Bids pid bid price pid00000 1.54 pid00001 16,99 pid00002 7.34 pid00003 2.31 uid time dictionary of candidates pid:score 10057 timestamp {pid00000:0.91, pid00305: 0.94, ¡­ 58738 timestamp {pid00006:0.95, pid00134: 0.99, ¡­
  • 8. Results 90k twitters generate a tweet flow of 1-10 words at 100 tweets/sec 500 advertisers updating the bid flow 18k-word vocabulary Each word vector is of dimension 10 Bids follows a Pareto distribution Generating a revenue of 500k per hour on average
  • 9. About me Scientist PhD Engineer Entrepreneur TransportCalc 10k+ lines in c++ Insight data engineer fellow business sense anisotropic turbulence bid data tools build products solve problems modeling experiment visualize machine learning algo & libs know customers & create values