Post on 04-Aug-2020
ARCHAN MISRA SINGAPORE MANAGEMENT UNIVERSITY
Mobile Analytics: The Ongoing
Revolution in Consumer Sense-making
THE BIG PICTURE
Unique Insights into Human
Behavior in Urban Settings
=
UNDERSTANDING CONSUMERS
Offline
•Scanner Data
•Consumer Panel Data
Online
•Clickstream Analysis
•Online Search Data
In-Store
•Movement & Navigation
•Activity Patterns & Insights
Long Term Purchase
Behavior
Medium-term Browsing
and Search Behavior
Real-time Adaptation
and Intervention
CONSUMERS & SOCIAL INTERACTION
ONLINE BROWSING &
PURCHASE
PHYSICAL WORLD IN-
MALL/STORE BEHAVIOR
Collaborative
Filtering
Influence
Networks and
Propagation
Opinion &
Sentiment
Mining
Who are you
with?
What are you
doing together?
What parts of your
in-store browsing
are similar?
What are you (not)
interested in?
The Analogue?
MOBILE: JUST A BETTER SHOPPING TOOL?
•Only user-initiated
actions & preferences
•No system-inferred
adaptation to implicitly
infer consumer
activities/preferences.
What’s Missing?
Reproduced from: M. Sneathen, MobileU 2013 Summit, Heartland Mobile Council
Mobile Analytics & LiveLabs
Real-time Mobile
Sensing (Activity, Indoor location,
Browsing, SMS ...)
Real-time Individual and
Server Group Analytics (Dynamic Group Detection, Queuing,
Preferences)
Context-Driven
Intentions (Incentives, Promotions,
Recommendations ...)
AN EXAMPLE OF LIVELABS’ CAPABILITIES
7
LiveLabs Cloud Service
5 minutes later
Lifestyle
Company
If a group of 4 or more people exit
from Café after sitting down for 10
minutes, send SMS with a “Movie
Discount”
10 minutes later
4 in a group
sitting down at
a Café
4 in a group left
after 10 mins
LiveLabs software
continuously monitors
(location, activity, …)
Show this
notification and get
20% on all Movies
RESEARCH THREADS
• Analytics: Queuing Detection
• Analytics: Group Detection
• In-Store Shopper Classification
QUEUING 101
Japan
• Americans spend 4 years of their lives queuing.
• Average wait time at lunch in Singapore ~= 10+
mins
Western: Nothing is certain in life but death and taxes
Asian: Nothing is certain in life but death and QUEUING
Singapore
QUEUING: 3-TIER DETECTION ARCHITECTURE
QUEUING: CHALLENGES & SOLUTIONS
Variability in
Service Times
Concurrent HA
Detection
Start & End of Queue
Detection
Queue class Service Time (seconds)
Min. Ave. Max. Stdev.
Airport check-in (*1) 10.7 102.1 421.7 136.9
Airport boarding 4.3 7.3 10.9 1.7
Café and food court 10.3 32.9 77.0 17.9
Movie ticketing (*1) 5.0 7.9 11.5 2.3
*0: Numbers were based on serving (“dequeue”) intervals at the head of the queue.
*1: Multiple serving counters serving for one queue.
Queue detected at any profile
Location info.-based trigger
Leave detected
Queue
detection
Leave
detection
Disabled
“Food court” FrameLen=48sec
“Air. Checkin” FrameLen=120sec
“Air Boarding” FrameLen=10sec
Detection of signature activity of “Queue Leaving” e.g., “5 sec. continuous walking”
RESEARCH THREADS
• Analytics: Queuing Detection
• Analytics: Group Detection
• In-Store Shopper Classification
13
• Use spatiotemporal analytics to
identify individual vs. group
relationships among visitors. recommendations
• Identify group-based social
relationships.
• Predict future
locations/activities likely to be
visited/performed by
individual/group.
GROUP DETECTION ON INDOOR
MOVEMENT DATA
•Indoor Location Data
• With +-4/5 meter accuracy
•Dense, Indoor Urban Spaces
14
Unified Metric Learning Framework
● Combines spatial coordinates, semantic location labels and activity profiles
Visualization of group and individual movement patterns
● Test on LiveLabs testbeds (SMU, Malls)
● Combine with observations of social media data
Real-time measurement of
● User indoor location
● User context (e.g., sitting, walking)
ALGORITHM & RESULT
Outperforms individual trajectory or
social media based group detection
RESEARCH THREADS
• Analytics: Queuing Detection
• Analytics: Group Detection
• In-Store Shopper Classification
THE PROBLEM: CLASSIFYING IN-STORE
SHOPPER BEHAVIOR
Using Mobile Sensing to
Infer
Consumer Preferences,
Interests, Behavior
OVERALL CLASSIFICATION APPROACH
THINGS NOT OBVIOUS
What are the MA labels?
(Different for different
stores?)
What are the HA
classifiers?
How many classifiers?
AN INITIAL CONTROLLED STUDY
16 “shoppers” in a shopping mall in Singapore.
Provided with smartphone to collect sensor data.
Shadowed by observer for “ground truth”.
Given specific tasks corresponding to the HA labels.
A mix of male and female shoppers with varying physical
characteristics.
Pre and Post surveys to verify their shopping behavior profile.
2 Types of Stores: Clothing and Shoe/Accessories
Shopping episodes lasting 3 -10 mins.
RESULTS ON HA CLASSIFICATION..INITIAL
Classifying Focused versus Confused customers:
Classification run on (i) Entire data set (ii) gender-specific data
0.00%
20.00%
40.00%
60.00%
80.00%
100.00%
Accu
racy
Grouping the dataset
Locomotive + Trajectory 71.43% 50% 100%
Only Locomotive 50% 66.67% 75%
M+F M F
Gender matters: It seems
that “focused” female
shoppers have a very
different characteristic
than “confused” female
shoppers.
Need a way to discover
and account for such
“attributes that matter”
OPEN CHALLENGES
21
Energy Overheads of
Continuous Sensing
•
Power Consumption
Observed on a Test
Samsung Galaxy S3
Improve Classification
Accuracy
•
Boxplot of Service
Time Estimates
(Queuing at F&B)
Privacy for Data
Need Anonymization and
Provenance Solutions!
ACKNOWLEDGMENTS
22
• LiveLabs Research Center: performed jointly with Assoc. Professor Rajesh BALAN (SMU)
• Queuing Research: Tadashi OKOSHI, Rahul MAJETHIA and Rajesh BALAN
• Group Detection Research: Siyuan Liu (CMU), Kasthuri Jeyarajah and Ramayya KRISHNAN (CMU),
• In-Store Shopping Research (jointly with IBM Research): Sougata SEN, Vigneshwaran SUBBARAJU, Dipanjan CHAKRABORY (IBM Research), Nilanjan Banerjee (IBM Research)