View
221
Download
0
Embed Size (px)
Citation preview
An Investigation into Guest Movement in the Smart PartyJason Stoops ([email protected])
Faculty advisor: Dr. Peter Reiher
Outline
Project Introduction Key metrics and values Mobility Models, Methods of Testing Results Analysis
What is the Smart Party?
Ubiquitous computing application Someone hosts a gathering Guests bring wireless-enabled devices Devices in the same room cooperate to
select and supply media to be played Songs played in a room represent tastes
of guests present in that room
Project Motivation
Are there ways to move between rooms in the party that can lead to greater satisfaction in terms of music heard?
Can we ultimately recommend a room for the user?
What other interesting tidbits about the Smart Party can we come up with along the way?
Smart Party Simulation Program
Basis for evaluating mobility models (rules of movement).
Real preference data from Last.FM is used. Random subsets of users and songs chosen
Many parties with same conditions are run with different subsets to gather statistics about the party.
Initial challenge: extend existing simulation to support multiple rooms.
Metrics
Satisfaction: based on 0-5 “star” rating Rating determined by play count Exponential scale: k-star rating = 2k satisfaction 0-star rating = 0 satisfaction (song unknown)
Fairness: distribution of satisfaction Gini Coefficient – usually used for measuring
distribution of wealth in a population. In Smart Party, wealth = satisfaction. Ratio between 0 to 1, lower is more fair.
Key values
History Length Number of previously heard songs the user device will
track. Used to evaluate satisfaction with current room
Satisfaction Threshold Used as a guide for when guest should consider
moving. If average satisfaction over last history-length songs
falls below sat-threshold, guest considers moving.
Mobility Models Tested
No movement Random movement Threshold-based random movement Threshold-based to least crowded room Threshold-based, population weighted Threshold-based, highest satisfaction
Test Procedure
Round 1: Broad testing to find good values for history length and satisfaction threshold for each model. (25 iterations)
Round 2: In-depth evaluation of model performance using values found above. (150 iterations)
Ratio of six guests per room maintained
Round 1 Results
Model History Length Threshold
No Movement n/a n/a
Random n/a n/a
Threshold Random 4 1
Threshold Least Crowded
4 1
Threshold Random, Population Weighted
5 0.5
Threshold Highest Satisfaction
2 2.25
Round 2: Satisfaction Overview
18 guests / 3 rooms30 guests / 5 rooms
60 guests / 10 rooms90 guests / 15 rooms
0
50
100
150
200
250
Median Overall Satisfaction
25th, 75th quartiles shown
NOMOVE
THRESHOLD LEAST CROWDED
THRESHOLD RANDOM POP WEIGHTED
RANDOM
THRESHOLD RANDOM
THRESHOLD HIGHEST SATSa
tisfa
ctio
n
Round 2: Fairness Overview
18 guests / 3 rooms30 guests / 5 rooms
60 guests / 10 rooms90 guests / 15 rooms
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Median Overall Fairness
25th, 75th quartiles shown, lower is better
NOMOVE
THRESHOLD LEAST CROWDED
THRESHOLD RANDOM POP WEIGHTED
RANDOM
THRESHOLD RANDOM
THRESHOLD HIGHEST SAT
Fa
irn
es
s
Topics for Analysis
Moving is better than not moving Party stabilization? Initial room seeking Population-based models perform poorly Satisfaction-based model performs well
Moving Versus Not Moving
Movement “stirs” party, making previously unavailable songs accessible
Songs users have in common changes with movement, depleted slower. NOMOVE RANDOM
0
20
40
60
80
100
120
140
160
180
200
Random vs. No Move, Median Overall Satisfaction
25th, 75th quartiles shown
18 Guests / 3 Rooms
30 Guests / 5 Rooms
60 Guests / 10 Rooms
Sa
tis
factio
n
Party stabilization?
Do users find “ideal rooms” and stop moving?
No! Some movement is always occurring.
Cause: Preferences are not static, they evolve over time.
0 5 10 15 20 25 30 35
0
0.1
0 .2
0 .3
0 .4
0 .5
0 .6
0 .7
0 .8
0 .9
R o o m C hang e s p e r G ue s t o ve r T im eThres hold Highes t S at, 30 gues ts / 5 room s
Round
Mov
emen
t P
roba
bilit
y
Initial room seeking
90% of guests move after round 1
Guests have some information to go on after one song plays.
Guests that like the first song in a room likely have other songs in common.
0 5 10 15 20 25 30 35
0
0.1
0 .2
0 .3
0 .4
0 .5
0 .6
0 .7
0 .8
0 .9
1
R oo m C hanges pe r G ues t ove r T im eThresho ld Highest S at, 60 guests / 10 rooms
Round
Mov
emen
t P
roba
bilit
y
Initial room seeking, cont.
In satisfaction-based model, peak is in round 2 All other models peak in round 1.
0 5 10 15 20 25 30 35
0
1
2
3
4
5
6
7
8
Round-by -round S at is fac t ion
60 Gues ts in 10 Rooms
N O M O VE
R AN D O M
T H R ESH O LD R AN D O M
T H R ESH O LD H IG H EST SAT
R ound
Sa
tis
fac
tio
n
Population-based models
Worse than choosing a room at random!
Weighted model performed better as weighting approached being truly random.
However, still better than not moving at all.
18 guests / 3 rooms 30 guests / 5 rooms
0
20
40
60
80
100
120
140
160
180
200
Median Overall Satisfaction
25th, 75th quartiles shown
NOMOVE THRESHOLD LEAST CROWDED
THRESHOLD RANDOM POP WEIGHTED
RANDOM
Sa
tis
factio
n
Satisfaction based model
Informed movement better than random movement.
Greater advantage as more rooms are added.
Short history length (two songs) used since history goes “stale”.
18 guests / 3 rooms30 guests / 5 rooms
60 guests / 10 rooms90 guests / 15 rooms
0
50
100
150
200
250
Median Overall Satisfaction
25th, 75th quartiles shown
NOMOVE RANDOM THRESHOLD RANDOM
THRESHOLD HIGHEST SAT
Sa
tis
factio
n
Conclusion
Room recommendations are a feasible addition to the Smart Party User Device Application.
Recommendations based on songs played are more valuable than those based on room populations.
Movement is a key part of the Smart Party.
Acknowledgements
At the UCLA Laboratory for Advanced Systems Research: Dr. Peter Reiher Kevin Eustice Venkatraman Ramakrishna Nam Nguyen
For putting together the UCLA CS Undergraduate Research Program Dr. Amit Sahai Vipul Goyal
References
Eustice, Kevin; Ramakrishna, V.; Nguyen, Nam; Reiher, Peter, "The Smart Party: A Personalized Location-Aware Multimedia Experience," Consumer Communications and Networking Conference, 2008. CCNC 2008. 5th IEEE , vol., no., pp.873-877, 10-12 Jan. 2008
Kevin Eustice, Leonard Kleinrock, Shane Markstrum, Gerald Popek, Venkatraman Ramakrishna, Peter Reiher . Enabling Secure Ubiquitous Interactions, In the proceedings of the 1st International Workshop on Middleware for Pervasive and Ad-Hoc Computing (Co-located with Middleware 2003), 17 June 2003 in Rio de Janeiro, Brazil.
Gini, Corrado (1912). "Variabilità e mutabilità" Reprinted in Memorie di metodologica statistica (Ed. Pizetti E, Salvemini, T). Rome: Libreria Eredi Virgilio Veschi (1955).
Audioscrobbler. Web Services described at http://www.audioscrobbler.net/data/webservices/