4. What are we really looking for?
9/3/20154
Not the home location — First Location of
Persistence (FLOP)
Does not necessarily correlate to reported home
or billing address
Challenges:
Perhaps less RAN usage at FLOP
Median user may have predictable FLOP time
range, but many user profiles do not conform
SLOP may have equal time duration and be
indistinguishable
5. Geomarketing and Using FLOP
9/3/20155
Device location enables:
1. Sending personalized offers, directions to targeted stores
and venues etc.
2. Discovering mobility patterns
Finding the persistent location for a user allows:
1. Precise definition of how a venue or retailer fits into the daily
life of consumers
2. Determining competition closest to customer base, not to
retail location
3. Improved means of marketing via stronger, more relevant
demographics rather than push/individually-targeted
advertisements
4. Detection of significant changes in targeted consumers
8. Approaches
9/3/20158
To algorithmically determine the location which
corresponds to decreased mobility, three methods
were investigated:
Curve fitting
Local minima analysis
Weighted K-means clustering
9. Curve Fitting
9/3/20159
Tried to fit a binomial curve, Gaussian
distribution, beta distribution and Rayleigh
distribution
Utilized chi-square test for goodness of fit
Poor performance
where
Oi : Observed value
Ei: Expected value
10. Curve Fitting Pitfalls
9/3/201510
Does not work for sparse data (very misleading)
Implies some level of symmetry in movements
Only works on single 24 hour window
Need mixture of curves and more computational
power for higher time range
Can only get FLOP
11. 9/3/201511
• Local minima
analysis results for
the most recent two
days showing time
and location spread
for the persistent
locations
• Matched for 735609
users out of total
1187116
Local Minima
Analysis Results
12. 9/3/201512
• Revisiting previous
plot with smaller bin
size to better
appreciate results
• Median of the FLOPs
returned were within
200m. from each
other
Local Minima
Analysis Results
13. 9/3/201513
Local Minima
Analysis Results
• Local minima
analysis results for all
the 5 days showing
time and location
spread for the
persistent locations
• Matched for 833453
users out of 1187116
users successfully
(73%)
14. Weighted K-
Means Clustering
9/3/201514
• Two approaches were
taken
• First one shown
here
• Two clusters
formed for data
over two days
• A cluster for each
day and variation
studied
• Matched for
886228 users out
of 1187116 (75%)
15. 9/3/201515
Weighted K-
Means Clustering
• Second approach: For
data over recent most
2 days
• Different weighing and
filtering method
• 3 cluster problem
reduced to 2 clusters
by appropriate
selection of 24-hour
window
• Matched for 847317
users out of 1187116
(71%)
• Measure of typical
time obtained
• Typical duration for
location and its
16. 9/3/201516
Weighted K-
Means Clustering
• Intuition: Better
clustering with larger
data set.
• Attempted to see if
analysis of data over
entire 5 days
produced any better
results.
• Matched for 887682
users out of 1187116
users (75%)
• Better than 2 day
analysis yet poorer
than local minima
analysis
17. Performance Comparison
9/3/201517
Weighted K-Means Clustering Local Minima Analysis
Can be used to measure typical
time and duration at persistent
location for the user
Doesn’t say anything about typical
time or duration at location for user
SLOP and TLOP can be obtained,
given sufficiently large data
No SLOP or TLOP
Lots of post processing of data
needed
(filtering and weighing of data).
SLOW
Less post processing required.
FAST
High memory consumption due to
weighing method
Low memory requirement
Requires larger data set to work on Works on sparse data set as well
Calibration of returned values of
typical time and duration needed
Confidence measure is hard to
establish here
18. Conclusion
9/3/201518
Tested 3 approaches to determine FLOP of users
Local minima analysis: Very efficient and
reasonably accurate
2 days of DLS data sufficient to work on
Use search space 12 hours post/prior maxima
On average, user generates 100 locations per day
A location record is typically 85 bytes
For one user, about 17 kB of data
19. Future Prospects
9/3/201519
We may consider “slope-analysis” based
algorithm that can work on-the-fly, without post-
processing
Using GMMs (Gaussian Mixture Model) or other
statistical modellings
Other machine learning algorithms
Picking location for minimum mobility time in local
minima analysis
