2. Malcolm Gladwell Title:
What Chipotle, Glenn Beck and
Alien Abductions Teach Us About
the Future of the Web
3. Graphs 101
Node
A B Node
Edge
Social networks (Facebook): nodes are people, edges “friendship”
Communication graph (Skype): nodes are people, edges communications
Search ranking graph (Google): nodes are pages, edges links
Taste graph (Hunch): nodes are people, edges taste similarity
Interest graph (Twitter, Instagram): nodes are people, edges interest
4. First Graph Theory:
Euler’s 7 bridges of Koeningsberg
•Is it possible to traverse the town & cross •Convert land to nodes & bridges to edges
each bridge exactly once? •Any node that is passed through must have
even number of edges
•Thus only solvable if you have 0 or 2 nodes
with odd number of edges
6. Directed Graph: Relationship Non-symmetric
(Like, follow, subscribe)
One could argue that Twitter’s main innovation was making edges non-
symmetric (directed), turned social network into publishing platform
Facebook began as undirected friend graph but has since bolted directed
“like” graph on top of it.
8. Averages
Twitter:
Number of followers: 62.97 per user
Number of followees: 43.52 per user
Facebook:
Number of facebook likes: 217.2 per item (liked)
Number of facebook likes: 29.30 per user
But distributions are interestingly different...
9. Twitter distributions are power curves
Distribution of # of followers you have Distribution of # of people you follow
Spike of “# following” curve around 20 due to old onboarding process (?)
10. Facebook friends is more like a bell curve
y = number of people; x = number of friends for those people
13. Marketing
C
similar demographics to A
A
B
purchased product communicates with A
B more likely to buy than C
Telecom company tested using phone call graph to use for direct mail*
Targeting network neighbors of purchasers dominated other targeting techniques.
Today, Facebook and many ad networks use similar targeting for online ads.
* “Network-Based Marketing: IdentifyingLikely Adopters via Consumer Networks - Shawndra Hill, Foster
Provost and Chris Volinsky
14. Defense
You can infer organizational hierarchies from communication patterns.
Governments use this to map rogue organizations.
calls
A B
responds immediately
calls
B A
responds slowly
THEREFORE
A B
Boss Henchman
15. Google founders’ $200B idea
Words and documents are nodes, connected by occurrence
PageRank: Links are directed graph
Node Node
18. Start with smaller graph:
Bowling Pin Strategy
Everyone
Everyone
More colleges
More colleges
area colleges
area colleges
Boston
Boston
Harvard
• Utility is proportional to square of network coverage, but how to start?
• Shrink size of the initial network and grow from there
• Also try to choose a sub-network with natural ‘spillover’ effects
•In this example, students at one college tend to have friends at others
19. Find clusters within existing graphs
A lot of people in the 90s thought dating would be “winner
take all” - but didn’t account for clustered graph structure
20. Introducing Overlap of Buyers/Sellers can add
Differentiation even in Entrenched Graphs
Heterogeneous Homogenous
buyers/sellers Hybrid buyers/sellers
For heterogenous buyers/sellers consider “Ladies night strategy”
23. When to Interoperate?
Metcalfe’s Law Corollary:
Network value ~ (nodes)2 Little guy benefits more than big guy
Little guy
Big guy
Little guy joins network and:
•Big guy gains small incremental increase in connections
•Little guy gains value of the many existing connections
•That’s why AIM (as incumbent big player) resisted when
Yahoo! & Google wanted to interoperate for IM
24. On the other hand…
• Each little guy benefits more than the big guy from interoperating
• But thousands of little guys relying on the big guy solidifies big guy position
• Facebook realized this and introduced Facebook Apps, Connect and other
“interoperating” features to prevent the “social network decay” that destroyed
previous social networks.
Facebook dev platform
26. Tastemates as Basis of a Graph
Someone out there must enjoy the same tile/strategy games I do…
And chances are they are not (yet, anyway) my friend
?
Enigmo Modern Conflict Carcasonne
27. The “Cold Start” Challenge
for Taste-Based Predictions
How to provide initial recommendations for a new user?
Force train, then predict
Assume tastes are driven by social graph
Leverage cross-vertical knowledge and
adjacent known nodes in Taste Graph
28. One Cold Start Solution:
Propagate Known Data to Unknown Nodes
• Iteratively propogate with adjacent data
• Dynamically adjust with ‘hard’ data
• Lather, rinse, repeat
= Known data
= Unknown data
31. Communications Graphs:
How Related are they to Social or Taste Graphs?
My iPhone contacts include some of my friends…
…but also my plumber, doctor, network administrator, United
Airlines and the Chinese restaurant around the corner
A lot of people were surprised that their email contacts were
assumed to be active social contacts
32. Could We Use Ad Preferences to
Cold Start Restaurant Recs?
hotpot
+
32
33. We know this person likes Classical Music, Yoga, Poetry, and Hiking
33