Embed Size (px)
Citation preview
Intelligence Through
Location Analytics
Sense Networks Overview
• Founded in 2006 by world-class MIT and Columbia Computer Scientists interested in understanding human behavior through location information
• Proprietary technology and deep expertise in geospatial and temporal analysis to deliver unique insights, trends and intelligence based on behavior patterns
• Team of 16, based in New York and San Francisco, funded by Intel Capital and Javelin Ventures
2
2009 Excellence Award 2009 AlwaysOn Award
2009 Cool Vendor 2009 Company to Watch Award
World’s Most Intriguing Startups
Location-Based Services and Mobile Advertising: For Years,
Focused on Simple Proximity-based Coupons
• Maybe some context like the weather or day of week
• No personalization – ads based either on opt-in lists from
retailers or mass marketing
3 Proprietary & Confidential
=>
A Better Way: Use The Best Context Available – Current and
Historical Location Information
• Existing targeting: context from mobile content consumption
• What about location history? Best predictor of behavior and
how we interact with the real world
4 Proprietary & Confidential
Location History Can Drive Much Better Recommendations and
Mobile Advertisements
=>Segment:
Health & Fitness
Young Adult
Outdoorsy
+ Location History: Parks
. . . A Different Ad
5 Proprietary & Confidential
+ =
Location History: Parks
and Recreational Areas
Current Context:
Location, Weather,
Time-of-Day, User
“Mode” (e.g. shopping
or commuting)
Sense Networks Has Built a High-Capacity Platform for
Extracting Intelligence From Location Data and Summarizing It
Ex
tra
ct I
nfo
Ex
tra
ct I
nfo
Cle
an
an
d
No
rma
lize
Cle
an
an
d
No
rma
lize
Da
ta R
ed
uct
ion
Da
ta R
ed
uct
ion
Se
gm
en
tati
on
Se
gm
en
tati
on
Pre
dic
tio
nP
red
icti
on
MacroSense Software Platform
Input Locations
Info
rmation O
utp
ut
6 Proprietary & Confidential
Ex
tra
ct I
nfo
Ex
tra
ct I
nfo
Cle
an
an
d
No
rma
lize
Cle
an
an
d
No
rma
lize
Da
ta R
ed
uct
ion
Da
ta R
ed
uct
ion
Se
gm
en
tati
on
Se
gm
en
tati
on
Pre
dic
tio
nP
red
icti
on
Input Locations
Info
rmation O
utp
ut
We can extract thousands of location
“features” from tens of millions of users and
tens of millions of points-of-interest
Proprietary algorithms to summarize all this information more efficiently (“jpeg for data”)
Example: A Mobile User’s Location Data and Call Activity Can
Be Abstracted to Commercial, Advertising Exposure . . .
Flow Call Activity Demographics, Commercial,
Ad Exposure
7 Proprietary & Confidential
Week
Hour
FLO
1
FLO
2
… FLO2
0
SIC
1
SIC
2
… SIC
97
DEM
1
DEM
2
… DEM
78
1 .03 .31 .14 .03 .05 .41 .11 .04 .01
2 .14 .34 .02 .04 .05 .52 .01 .01 .00
…
168 .07 .34 .51 .02 .06 .48 .02 .01 .00
. . . And Compacted Into A “Location DNA” for Each User . . .
DNA User 1 DNA User 2
Tim
e o
f D
ay,
Da
y o
f W
ee
k
8 Proprietary & Confidential
Category of Commercial Exposure (i.e. restaurants, schools, golf courses)
Category of Commercial Exposure (i.e. restaurants, schools, golf courses)
Tim
e o
f D
ay,
Da
y o
f W
ee
k
. . . To Create Segments – All From Anonymous Location Data
Nightlife Profiles, Primary Clusters
How often do they go out each
day of the week?
Where do they hang out?
Are the places they spend
What is the avg age of most
people in the neighborhoods
they spend time in?
How racially diverse are the
neighborhoods they spend
time in?
9 Proprietary & Confidential9
Are the places they spend
time in rich neighborhoods
or poor neighborhoods?
“Young & Edgy”•Out every night in young, racially diverse, low income neighborhoods
“Young & Edgy”•Out every night in young, racially diverse, low income neighborhoods
“Weekend Mole”•Out occasionally on weeknights, typically middle-aged, Latino, middle-income neighborhoods
“Weekend Mole”•Out occasionally on weeknights, typically middle-aged, Latino, middle-income neighborhoods
“Mature Homebody”•Rarely goes out, typically spends nights in mature, white, higher income neighborhoods
“Mature Homebody”•Rarely goes out, typically spends nights in mature, white, higher income neighborhoods
Location-Based Segments Proven to Drive User Behavior
• Example: Using data from a mobile location app, we predicted
new places that users would go based on their “tribe”
• If we gave users 500 recommendations, 20% were acted on
15%
20%User Response To Sense’s Top Recommendations
10 Proprietary & Confidential
5%8%
20%
0% 1% 4%
0%
5%
10%
15%
50 100 500
Sense Networks
Baseline
We examined 30k users and 5k points of interest. If users were presented with 50, 100, or 500 place recommendations they had not previously visited, what % of those
would they visit and check in? For 100 Sense recommendations, 8% were acted upon by users. For 500 Sense recommendations, 20% were acted upon.
Contact Info
1123 Broadway (between 25th and 26th Streets)
Suite 817
New York, NY 10010
+1 646 758 6227
Mikki Nasch, EVP Business Development: [email protected], +1 646-845-9859
Christine Lemke, COO: [email protected], +1 917-284-8384
David Petersen, CEO: [email protected], +1 415-336-3948
11 Proprietary & Confidential
![Universal approximations of invariant maps by neural networks · in the sense of the universal approximation theorem for neural networks Pinkus [1999]. Designing invariant and equivariant](https://img.pdfslide.net/doc/110x75/5f18775525025331252c439e/universal-approximations-of-invariant-maps-by-neural-networks-in-the-sense-of-the.jpg)
