SmartAds:
Bringing Contextual Ads
to Mobile Apps
Mike Lin
Outline
Authors
Introduction
Background
Characterization
SmartAds Architecture
Results
Discussion
Comment
Authors
Suman Nath
A Ph.D student at the School of Computer
Science, Carnegie Mellon University .
A masters student in University of Illinois at
Urbana-Champaign
work at Microsoft Research.
Suman Nath’s Work
Authors
Felix Xiaozhu Lin
PhD student in Computer
Science at Rice University
Software Developer at Collabera
Research Intern at Microsoft
Research Visiting Student at IBM
Research Intern at Nokia
Intern at Baidu, Inc.
Education
Rice University
Tsinghua University
Tsinghua University
Authors
Lenin Ravindranath Sivalingam
A PhD student at CSAIL.
I received Microsoft Research Graduate
Fellowship in 2011.
I joined MIT in the fall of 2008.
I was on leave (from May 2011 to Jan
2013) building AppInsight and related
systems at Microsoft Research.
Authors
Jitendra Padhye
A Principal Researcher at Microsoft Research.
He is interested in all aspects of computer
networking and networked systems. His
recent work has focused on data center
networks and mobile computing. He has
published numerous research papers in top
conferences, and holds over 25 US patents.
He is the recipient of the ACM SIGCOMM
Test of Time award. He received his PhD in
Computer Science from University of
Massachusetts Amherst in 2000.
Outline
Authors
Introduction
Background
Characterization
SmartAds Architecture
Results
Discussion
Comment
Introduction
US consumers spent 30% more time on mobile apps
than on traditional web.
Today's mobile ads are not contextual and mostly
irrelevant.
To scrape the content at runtime, extract keywords and
fetch contextually relevant ads.
Introduction
Introduction
i. PHYSICAL CONTEXT-AWARE ADVERTISING:
based on users' locations, activities, and other
physical contexts.
ii. BEHAVIORAL TARGETING:
based on each individual user's past web usage
history.
Outline
Authors
Introduction
Background
Characterization
SmartAds Architecture
Results
Discussion
Comment
Background
Contextual Advertising: matching ads to web pages
(1)Offline ad labeling:
Ads in ad network's inventory are labeled with
keywords, called bidding keywords.
(2)Offline keyword extraction:
Ad network employs a bot that crawls web pages
and uses machine learning algorithms (such as
KEX) to extract keywords.
(3)Online web page to ad matching.
Background
Mobile Ads:
(1)main signals today:
the app metadata
(2)focus in this paper:
contextual advertising-challenging
-the apps often transform the content in a variety of ways, sometimes combining
content from multiple sources, such as a news aggregator app.
Two challenging points:
a. limited resources
b. user privacy
Outline
Authors
Introduction
Background
Characterization
SmartAds Architecture
Results
Discussion
Comment
Characterization
Page Data:
A rich source of ad keywords
More keywords than the metadata
Change significantly over a period of time
Example:
If a user uses the app for nding local restaurants, he
should be shown restaurant related ads
Characterization
Methodology
Apps
PhoneMonkey
emulate various user interactions (touch, swipe etc.)
App session(each run):
run each of top 1200 app 30 times with the PhoneMonkey.
Methodology
Keyword Extraction
A modified version of the well-known KEX keyword
extractor
Example
the words “is” and “zebra”:both get a low score
the word “pipe”:
get higher weight because of plumbing business
Methodology
Page data is a good source of ad keywords.
Methodology
Page data yields more keywords than metadata.
Methodology
Page data yields more keywords than metadata.
Methodology
Page data is dynamic, and requires online keyword extraction.
Outline
Authors
Introduction
Background
Characterization
SmartAds Architecture
Results
Discussion
Comment
SmartAds Architecture
Utility. Efficiency. Privacy.
SmartAds Architecture
Achieving Good Utility
Local features
AnywhereCount: total times
NearBeginningCount: times the word appears in the beginning of the page
SentenceBeginningCount: times the word starts a sentence.
PhraseLengthInWord:words in the phrase.
PhraseLengthInChar:characters in the phrase.
MessageLength: The length of the line, in characters,containing the word.
Capitalization: times the word is capitalized in the page.
Font size: Font size of the word.
SmartAds Architecture
Global knowledge
the knowledge about how often advertisers bid on a keyword
The frequency is how many times the word appears in the bidding keyword trace.
SmartAds Architecture
Achieving Efficiency
Addressing memory overhead
we partition x into a vector of local features xl
(e.g.,anywhereCount) and a vector of global features xg
(e.g, global knowledge of a keyword), with weight vectors
wl and wg respectively.
SmartAds Architecture
Achieving Efficiency
Addressing communication overhead
Bloom filter: a space-efficient probabilistic data structure
Bloom filter size at the client.
SmartAds Architecture
Achieving Efficiency
Addressing communication overhead
Dynamics of bidding keywords.
SmartAds Architecture
Achieving Privacy
The ad server knows only the ad keywords in the page and nothing else.
End-to-end workflow in SmartAds
Outline
Authors
Introduction
Background
Characterization
SmartAds Architecture
Results
Discussion
Comment
Results
Relevance
Results
Relevance
Results
End-to-end performance
Results
Overheads
CPU Overhead:
The combined runtime. As seen from Table 1,this overhead is minimal.
Memory Overhead:
The size of our bloom filter is 1MB. Based on our
measurements, SmartAds control consumes around 2.8MB of memory on an average.
Network Overhead:
From our analysis of 1200 apps, we find that the average number of keywords extracted
per page is 7.6. Hence, the average extra bytes sent is 91.
But on average various ad controls upload 1.5KB and download 5KB of data.
Battery Overhead:
The increase in power consumed was less than 1% and well within experimental noise.
Outline
Authors
Introduction
Background
Characterization
SmartAds Architecture
Results
Discussion
Comment
Discussion
Optimizations
Addressing lack of text:
Level 1: keywords from the current page
Level 2 keywords from all the pages
Level 3 keywords for each app, learned offline from that app's metadata.
Handling related keywords
{HDTV}
{HDTV; LED TV; LCD TV}.
Sources:
Analyzing Bing web queries and click logs
http://veryrelated.com
Discussion
Dealing with Tail Bidding Keywords
(1) by serving them when their keywords are semantically related to any
keyword in the Bloom filter
(2) by prioritizing them when the current page does not contain any important
ad keywords
(3) by occasionally serving them even if the contextual signals do not match
them.
Outline
Authors
Introduction
Background
Characterization
SmartAds Architecture
Results
Discussion
Comment
Comment
Comment