Adaptive Personalization
for Mobile Content Delivery
Daniel Billsus, Craig Evans,
Raymond Klefstad, Michael J. Pazzani
[email protected], [email protected],
[email protected], [email protected]
Department of Information and Computer Science
University of California
Irvine, CA 92612
How do you fit a newspaper
into a cell phone?
Wireless Web vs. Wired Web
Most Wireless Web applications interact
with a simplified browser with simplified
markup languages
Smaller Screen Sizes
Limited Input Capabilities
Slower Network Connections
Higher Network Latencies
Less Reliable Network Connections
More Expensive Network Connections
Less Memory
Less Processing Power
Qualcomm Brew
1.
Programming language for cell phone applications
(C or C++ with cell phone API and libraries)
Many capabilities not present in browsers
2.
•
•
•
•
3.
Persistent file system
Alarms
Access to servers via http but no predefined browser or
markup language
Use wireless downloads and installs application
Launching in US with Verizon, Alltel, US Cellular plus
Japan, Brazil, Korea, etc.
However, still quite limited
4.
•
Limited Program space, file system
Adaptive News Browser
Capabilities
Interactive news reading: selecting sections
(e.g., sports) and headlines
Batch downloading a user specified number
of articles at a user specified time into cache
Uses off-peak minutes
Uses minutes more efficiently
Reduces latency when selecting headlines or
sections
Content available when wireless isn’t
Personalization
Reorders articles based on learned user
preferences
Determines which articles to download
Personalization Constraints
Learns quickly about new users
Smoothly transitions from general to
personalized display
Adapts quickly as interests change
Requires little or no work by user
Requires no manual tagging of content
Leverages existing editorial judgment so
important novel events are shown to all
Supports search, diversity,related content
High performance & scalable
Mobile Content
One Size Fits All (SMALL)
Personalization Middleware
Recommendation
r
e
t
r
i
e
v
a
l
short-term
long-term
diversity
editorial rankings
previous sessions
Interaction layer
search
history
related items
Batch download
Data Management
user models
content cache
t
r
a
n
s
c
o
d
i
n
g
Personalized Content
User Customization
Less than 2% of the users do it
User Configuration:
Inaccurate and hard to use
Check boxes
Keywords
Too coarse-grained
Ambiguous: Metallica Concert at Verizon Wireless
Amphitheater Sold Out
Filters vs. prioritizes
Requires regular web access
Requires constant maintenance
Adaptive Personalization
Adaptive Personalization II
Intelligent Search
Adaptive Personalization Solution
Solution: Hybrid Model to Sort Articles
without classification
Rank read (and skipped) stories from
behavior
Predict rankings of unseen
Short-term: Similarity-based
Long-term: Probabilistic
Editorial or Marketing input: Exponentially
decaying bonus
Variety by Similarity: Penalty for being too
similar to other recommended article
Representation and Similarity
Lawmakers Fine-Tuning Energy Plan
SACRAMENTO, Calif. -- With California's energy reserves remaining all but
depleted, lawmakers prepared to work through the weekend fine-tuning a
plan Gov. Gray Davis says will put the state in the power business for "a
long time to come."
The proposal involves partially taking over California's two largest utilities and
signing long-term contracts of up to 10 years to buy power from
wholesalers…
util-0.339 power-0.329 megawatt-0.309 electr-0.217 energi-0.206 caiso-0.192 california-0.181 florio-0.176 bui0.156 debt-0.128 lawmak-0.128 state-0.122 wholesal-0.119 partial-0.106 consum-0.105 tune-0.104 alert0.103 scroung-0.096 bottorff-0.096 iso-0.093 advoc-0.09 testi-0.088 bailout-0.088 crisi-0.085 amid-0.084
price-0.083 long-0.082 bond-0.081 plan-0.081 term-0.08 grid-0.078 reserv-0.077 blackout-0.076 bid-0.076
market-0.074 fine-0.073 deregul-0.07 spiral-0.068 deplet-0.068 liar-0.066 edison-0.065 contract-0.063 condit0.062 largest-0.061 rate-0.06 takeov-0.059 stock-0.059 michel-0.059 offici-0.058 audit-0.057 billion-0.056
apolog-0.056 auction-0.055 costli-0.055 rip-0.055 shed-0.055 drain-0.055 cost-0.054 skeptic-0.053 anymor0.053 announc-0.052 craft-0.052 pai-0.051 hour-0.05 take-0.05 super-0.049 howard-0.049 midnight-0.049
dai-0.048 percent-0.048 desper-0.048 flow-0.047 fridai-0.047 sacramento-0.046 sundai-0.046 grai-0.046
unabl-0.045 issu-0.044 set-0.044 shut-0.043 open-0.042 reveal-0.042 mexico-0.042 facil-0.04 tight-0.04
bowl-0.04 calif-0.039 pacif-0.039 expect-0.039 option-0.039 extend-0.039 consecut-0.039 conserv-0.038 roll0.038 davi-0.038 blame-0.037 bar-0.037 purchas-0.037 credit-0.036 revenu-0.036 stage-0.035 tom-0.035
custom-0.035 grant-0.035 hundr-0.035 fan-0.035 work-0.035 amount-0.035 reduc-0.034 call-0.034 weekend0.
Short Term model
tf t,d
w(t,d) 
i
sim ( X , D) 
tf
 N 
log 

 df t 
t ,d
i
2


tX
tX
 N 

log
 df ti 


2
w(t , X ) w(t , D)
w(t , X )
2

tX
w(t , D)
2
Profiles and informative words
Health
patients (16) study (11) drug (11) doctors (10) food (9) breast (9) surgery (8) cancer (8) health
(7) blood (7) women (7) center (7) vitamin (6) graedon (6) exercise (6) cell (6) kids (6) eating
(6) body (6) products (6) person (5) procedures (5) nutrition (5) risk (5) pill (5) weight (5)
disease (5) care (5) hospital (5) program (5) life (5) diet (5) supplement (5) loss (5)
performance (4) bone (4) site (4) surgeon (4) service (4) book (4) tissue (4) anesthesia (4)
children (4) meals (4) calcium (4) mestel (4) treated (4) athletes (4) feel (4) blvd (4) correct
(4) injuries (4) nurses (4) reserve (4) donned (4) older (3) section
Wall Street
share (203) percent (192) cents (151) point (136) billion (113) quarter (112) earnings (105)
million (101) stock (99) oils (88) sales (83) tires (80) bank (74) company (73) rate (67) yen
(64) pound (64) fell (62) price (62) rose (62) ford (60) december (57) bushel (57) soybeans
(54) treasury (54) firestone (53) chicago (53) discount (53) euro (52) yield (50) tokyo (50)
airlines (49) close (47) bond (47) wheat (47) recalled (45) vehicles (41) loss (41) corn (40)
store (40) japan (38) european (37) food (37) workers (36) revenue (35) london (34) plant
(33) cars (33) barrel (32) phone (31) year (31) deal (31) merger (30) union (30) crude (28)
percentage (28) profit (28) service (28) settlement (28) trade (28) technology (28) growth
(27) insurance (27)
Bayesian Text Classification
|d |
Pr(c j | d )  Pr(c j )   Pr( wdi | c j )
i 1
ˆ r( w | c ) 
P
i
j
1   N ( wi , d k )
d k
|V |
| V |   N ( wt , d k )
d k t 1
Increased Reading per usage
150
125
100
75
50
25
0
Original
Order
Adaptive
Order
After looking at 3 or more screens of headlines, users read 43% more of
the personally selected news stories; clearly showing AIS's ability to
dramatically increase stickiness of a wireless web application
Readership and Stickiness
150
125
100
75
50
25
0
Static
Personalized
After 6 weeks, 20% more users keep using wireless system when
personalized
Benefit
Speed to Effectiveness
100
90
80
70
60
50
40
30
20
10
0
1
2
3
4
5
6
7
8
9
10
Number of Sessions
Initially, AIS is as effective as a static system in finding relevant content. After only
one usage, the benefits of AdaptiveInfo's Intelligent Wireless Specific
Personalization are clear; after three sessions even more so; and, after 10
sessions the full benefits of Adaptive Personalization are realized
Probability a Story is Read
P ro b a b ility a t le a s t o n e s to ry is re a d
0 .9
0 .8
0 .7
0 .6
0 .5
0 .4
0 .3
0 .2
0 .1
0
1
2
3
4
5
6
7
P o s itio n o f S to ry
8
9
Ada ptiv e
S ta tic
40% probability a user will read one of the top 4 stories selected by an editor,
but a 64% chance they'll read one of the top 4 personalized stories - the AIS
user is 60% more likely to select a story than a non-AIS user
Similarity & Interest
p(select | cosine)
0.5
0.4
0.3
0.2
0.1
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
cosine
1
Adaptive Personalization Benefits
Consumer
Saves time to find personally valuable information
Enhances the user experience
Carrier
Enhance wireless user’s experience and carrier’s
revenue
Increased usage of premium services
Increased retention of customers
Evaluating the Hybrid User Model
70
70
60
50
F1
60
40
50
F1
30
20
40
1
2
3
Training Sessions
Hybrid
30
20
1
2
3
Training Sessions
Short-Term
Long-Term
Descargar

Document