Most location sharing applications display people's locations on a map. However, people use a rich variety of terms to refer to their locations, such as "home," "Starbucks," or "the bus stop near my house." Our long-term goal is to create a system that can automatically generate appropriate place names based on real-time context and user preferences. As a first step, we analyze data from a two-week study involving 26 participants in two different cities, focusing on how people refer to places in location sharing. We derive a taxonomy of different place naming methods, and show that factors such as a person's perceived familiarity with a place and the entropy of that place (i.e. the variety of people who visit it) strongly influence the way people refer to it when interacting with others. We also present a machine learning model for predicting how people name places. Using our data, this model is able to predict the place naming method people choose with an average accuracy higher than 85%.
Authors are Jialiu Lin, Guang Xiang, Jason Hong, and Norman Sadeh
Artificial intelligence in the post-deep learning era
Modeling People’s Place Naming Preferences in Location Sharing, at Ubicomp2010
1. Jialiu Lin, Guang Xiang, Jason Hong, Norman Sadeh
School of Computer Science
Carnegie Mellon University
2.
3. We seldom say …
• Hey mom, I am still at 55.66 north,
12.59 east. I will be home soon.
•Let’s have some coffee at 417 S
Craig st.
• I’m at 2039 Main st and already in
bed.
Instead, we say …
• Hey mom, I am still at school. I will be
home soon.
• Let’s have some coffee at Starbucks.
• I’m at home and already in bed.
4. Provide more flexible and pertinent
information
Can infer not only position, but also activities,
availability, safety and etc.
Multiple dimensions to tailor the information
Preserve privacy
Obfuscate physical location
e.g. sharing ‘Home’ instead of physical location of
home
Better Integration (vs. map)
More information in smaller space
5.
6. What if I want to share different location
names to different people?
I want to know I’m at work
I want to know I’m in the conference room.
I want to know I’m in RM 4221
……….
7. Ideally:
First Step: ← Focus of this paper
e.g:
f(40.44,-79.94, ‘sister’, work day night,…) ‘grocery store’
f(36.49,-79.22, ‘close friend’, weekend,…) ‘@ Starbucks’
e.g:
f(40.44,-79.94,‘sister’, work day night,…) place’s functionality
f(36.49,-79.22, ‘close friend’, weekend,…) Business name
8. Empirical study of location naming
2 weeks study with 26 participants in 2 cities
Result analysis
Place naming diversity
Taxonomy on location naming methods
Modulated location information
Influencing factors in place naming
Predictive model of place naming methods
▪ Top level accuracy 93%
▪ Sub level accuracy 68%
▪ Granularity accuracy 89%
Discussion and conclusion
9. Time: August 2009
Duration: Two weeks
Number of participants: 26
12 female, 14 male
Age 20-44, mean=25.6
Locations:
CMU Pittsburgh campus (18)
CMU Silicon Valley campus (8)
10. Entrance survey:
List names under different social groups
In study:
Mobile application recorded participants location
information (GPS+ wifi)
Participants uploaded this information through our
web application every day
Participants answered a set of question regarding
to the places they had been.
Exit survey:
General attitudes toward location sharing
Procedure
11. 12
You were observed at this location
from 15:35 Aug 12 (Wed) to 16:24 Aug 12 (Wed) • Maps reminds participants of the locations
they visited.
• Questions were asked for four social groups
•family member,
•close friend,
•acquaintance,
•stranger.
Imagine that Mary (your family member) wanted
to know where you were at the given time period.
1.How comfortable would you be to let her know
where you were at this time?
1:not comfortable at all, 7: extremely comfortable
2.How familiar is Mary with this location?
1: not familiar at all, 7: extremely familiar
3. What terms or phrases (place name) would you
use to refer to this location if you want to tell her
where you were?
12. Place naming diversity
Taxonomy on location naming methods
Modulated location information
Influencing factors in place naming
Predictive model of place naming methods
14. Taxonomy on Place Naming Method
Place Names
Hybrid
e.g. :
home, work,
friend’s house…
e.g. :
gym, restaurant,
grocery store…
e.g. :
McDonalds,
Hilton…
Semantic
Personal Functional
Business
name
Geographic
Address Landmark
e.g. :
5000 Forbes ave,
Rued Langgaards Vej, 2300
København…
e.g. :
near the Liberty Bridge,
outside city library…
17. Pattern 2: Modulate Location Information
Personal
47.1%
Functional
12.8%Business
name
9.3%
Landmark
1.3%
Address
23.6%
Hybrid
6.0%
Distilling: Viewer of this information
extract physical position by using
shared knowledge
27. Predictive Model of Place Naming Methods
•14 different attributes direct captured or derived
attributes, 3 labels
•Attributes: e.g. social relation (1-
4), frequency, comfort level(1-7), familiarity (1-
7), place entropy, duration (in sec), arriving
time, physical distance, and etc…
•Labels: top level category label, sub category
label, granularity label
• Use Weka 3 as the major tool
• Training and testing data separated by participant ID
• Randomly select 5 participants as testing
• Remaining as training
• Results averaged over 50 rounds
28. Accuracy%
J48 Decision Tree
Support Vector
Regression Naive Bayes
Top level
category
85.50
(3.14)
76.21
(4.27)
80.33
(3.51)
Sub-Class 60.74
(1.50)
54.26
(3.34)
56.19
(1.93)
Granularity 71.25
(3.44)
68.55
(4.58)
67.48
(2.67)
Predictive Model of Place Naming Methods
30. Accuracy can be boosted when learn from similar people
•Calculate similarity (Kappa value) among participants based on their exit survey
Accuracy% Max
Top level
category
93.2
Sub-Class 67.8
Granularity 88.7
31. Participants from one university
community
No real sharing happened during study
32. Conducted empirical study on how people
name places in different context
Proposed taxonomy of place naming methods
Identified several typical patterns
Place naming diversity
Location information Modulation
Significant influencing factors: social
relation, privacy, familiarity, place entropy.
33. Used machine learning to predict place
naming methods
Top categories accuracy 93%
Sub categories accuracy 68%
Granularity accuracy 89%
34.
35. Attribute
s
Explanations
(lat, lon) Geo-coordinates of the place
FromTime P’s arrival time to the place
ToTime P’s departure time from the place
Group The social group of R (Family member, close
friend, acquaintance, or stranger)
PhyDist The physical distance between P and R, in a
scale of 1 to 4 (1=same city, 2=same state diff
cities, 3=same country diff states, and 4=diff
countries).
CmftShare How comfortable of P letting R know where
he/she was at that moment, in a scale of 1 to 7
(1= not comfortable at all, 7= fully
comfortable)
Familiarity How familiar R with this place, in a scale of 1
to 7 (1=don’t know this place, 7=extremely
familiar. P can input “not sure” if they don’t
know the answer)
PlaceName The place name which P would like to use in
the specific scenario.
Attributes Explanations
DistHome Distance from this place to P’s home
DistWork Distance from this place to P’s work place
Duration The amount of time P spent at this place
Freq Number of times P visited this place
UserCount Number of participants who visited this
place
Entropy * The diversity of users visiting a particular
place.
* J. Cranshaw, E.Toch, J. Hong, A.
Kittur, and N. Sadeh, "Bridging the Gap
Between Physical Locaation and
Online Social Networks," in Proc.
UbiComp, 2010
Derived Attributes
Editor's Notes
Self-introduction
Thanks to the fast advancing mobile technology, knowing where we are becoming easier and easier.There are many LSA allow ppl sharing location with each other.Google latitude, loopt, ipokiTheseLSAs typically display ppl’s location on a map. This presentation is perfect for navigation, however, is not the way we usually use in daily life.
We seldomly say…..Instead we say…Like school, starbucks, home, ppl use a rich set of terms to describe their locations, we refer to these terms as place names. there is an obvious semantic gap lying between what we used to and what technology can offer,
Someone may say: wait a minute, There are already some LSA which features place names, like …….These check-in based application let users proactively ‘check-in’ their location by sharing one place names with others.
Emphasis that the diversity of place naming is created by ourselveMum -> at work. It’s not the right time to call meBoss -> in the conference room, such that he know I’m in a meetingClients-> they haven’t been to conference room b4.We can name lots of these similar examples. General speaking, people usually associate multiple place names with a single location, and used them in different situations respectively. We refer to these phenomenon as the diversity of place naming.
Create a system that can automatically generate appropriate place names based on real-time context and user preferences.Study how people refer to places in location sharingBuild a model to predict the method people might use to name the place
Pittsburgh,PensylvaniaMoffett Field, CA
Comparing to ESM, Day Reconstruction Method (Kahneman, Krueger, Schkade, Schwarz,& Stone, 2004),
Imagine.. In fig
Then, I will present you what we learned from the study. focusing on the observed behavior pattern such as: place naming diversity, modulated location informationThe identified influencing factors, like social relation, privacy concern, viewer’s familiarity and place entropy.Also, the predictive model we build based on the data we collected.
It does confirm our tatement b4 about
In order to have a clear idea of what kind of information is embedde in place names, we go through all the place names shared and organized them in to the following hairarchy we proposed. Hybrid give some examples
This measurement is proposed by my colleague Justin which can be used to estimate how public a place is. . He will give a detailed explanation of the measurement in his presentation tomorrow in the session of Location Sharing II. And also due to the interests of time, I will only go through two of the influencing factors. If you’re interested in the other two, welcome to read my paper.
This figure consists of two different measurements, hence it has two y axis.First let’s focus on the y axis on the left hand side. It is corresponding to the red and blue solid line. It shows the usage of semantic and geographic naming methods in percentage. ……closer relationship, share more common knowledge, more semantic information… The y axis on the right hand side is for the green dashed line. It represents the average granularity shared with different social group. Closer relationship, feel more comfortable of sharing more detailed information.
Important attributes that influence people’s place naming methods (left y-axis) and place naming granularity (right y-axis), vertical bar indicated the 95% confidence intervals. (a) Sharing with different social groups; (b) Comfort level of sharing (c) Recipient's familiarity (d) Place entropy. The total percentage of semantic and geographic naming exceeds 100%, since some place names contains both of them (i.e. Hybrid).
Important attributes that influence people’s place naming methods (left y-axis) and place naming granularity (right y-axis), vertical bar indicated the 95% confidence intervals. (a) Sharing with different social groups; (b) Comfort level of sharing (c) Recipient's familiarity (d) Place entropy. The total percentage of semantic and geographic naming exceeds 100%, since some place names contains both of them (i.e. Hybrid).
When viewer is not familiar with the location at all, the usage of semantic place naming is very high, since geographically information probably don’t matter that much. For example, I will never tell my mum I’m at 5000 forbesave, since she know nothing about forbesave at all. When the familiarity increase, we found the usage of geographic naming methods increases. It might be interpreted like the viewer knows the area but might not be familiar with the exact location. Hence the geographic information with aproximate street-level granularity can also convey the location information effectively.When the familiarity become even higher, more common knowledge are shared between the two parties. Therefore, once again, the semantic place naming method greatly dominate the usage.
Important attributes that influence people’s place naming methods (left y-axis) and place naming granularity (right y-axis), vertical bar indicated the 95% confidence intervals. (a) Sharing with different social groups; (b) Comfort level of sharing (c) Recipient's familiarity (d) Place entropy. The total percentage of semantic and geographic naming exceeds 100%, since some place names contains both of them (i.e. Hybrid).
After we finish analyzing the influencing factor, we wonder whether we can utilize these factors to predict in what way ppl would refer to different places. Hence we build predictive models on people’s place naming methods.
Top left: Average accuracy (%) of the top 3 algorithms (STD in parentheses) in predicting top level categories {semantic, geographic, hybrid}, Sub-class {personal, functional, business name, address, landmark}, and Granularity {state, city, region, street, building, room} labels. J48 performs the best.
We are definitely not satisfy with this prediction accuracy. If we make the study last longer, can we get better results. We break down the dataset in to days, varies the number of days included in the modeling from 2 days to 14 days. We surpricely found that, after 8 days, the prediction accuracy tend to plateau. Hence we believe that 2 weeks duration is enough for the purpose of our study. It also suggests that human being usually have weekly routine in their lives.
Then can we do better?Currently we learn the model from random ppl, what if we learn from ppl that is similar to ourselves. We use the exit survey results to estimate the similarity by calculating kappa value among participants. We found that when we learn from ppl similar, we can boost our prediction accuracies to …… in predicting…
We realize that. There are some limitation in our study, e.g.
Say about the trend observations specifically. To sum up.