Upload
phonecom
View
293
Download
1
Embed Size (px)
DESCRIPTION
Citation preview
Role-Based Contextual Recommendation
Cheng Zeng, Jian Wang, Liang Hong(Wuhan University)Jilei Tian, Xiaogang Yang(Nokia)
Nokia and Whu
Background
Nokia and Whu
Background
Cannot compete the amount of applications/services,Serving on demand is more important
Nokia and Whu
Background
Search Engine solves the problem of information explosion.Facing the vast and various applications/services (Traditional Software, Widget, REST/SOAP-WS, RSS…), what can we do?
The popularization of smart phones brings opportunities for exploiting personalized recommendation based on rich context information and mobile social networks.
What you think is what you get
Nokia and Whu
Background
On bus Media Player
Walking RadioMeeting Switch to Vibration E-book
Cooking Shopping AD
Nokia and Whu
Framework
Nokia and Whu
Role and Trust
Collaborative filtering (CF) is widely used in recommendation
systems. However, the weakness of CF involves in cold-start,
sparseness of useful information, and internal attacks. Trust-based
recommendation is proposed to improve the question.
Current trust -based recommendation approaches only consider
user’s trust statements and the similarity of user’s rating history,
which oversimplify the trust relationship.
Nokia and Whu
Role and Trust
“Role” has been mentioned in data security and software
engineering domain. Compared with these work, we adopt roles in
intelligent recommendation system with different connotation.
The advantage of using role: A user generally trusts one/several aspects(roles) of another user’s
interests but not all, namely conditional trust; Role takes common knowledge among different users into account and is
a high-level abstract of a group of mobile users with certain similarity, Interest reasoning among roles is more efficient than among individuals;
Role can distinguish the statements changing for a user based on context information and provide more accurately recommending.
Transfer learning (Cross-domain)
Nokia and Whu
Advanced User Profile
AUP is built in RDF, which describes user's static and dynamic information: <time, scene, behavior>
AUP supports shopping, music domain currently and will extend for AD and other domains.
User's habit, preference and behavior were learnt by data mining approach with the logged data collected from personal mobile devices
Nokia and Whu
Advanced User Profile
Nokia and Whu
Role Mining
Nokia and Whu
Role MiningPotential role mining algorithm
c1 c2 c3 c4 c5
u1 b1 b1 b1 0 0
u2 0 0 0 b2 b2
u3 b1 b1 b1 b2 b2
u4 b3 b3 0 0 0
u5 b3 b3 0 b2 b2(1) (2)
user
context
(1) (user, context, behavior) table. (2) Mine roles by clustering users’ behavior.(3) Get three potential roles. (4) Get role-context-behavior table
behavior
(3)
(4)
role
c1 c2 c3 c4 c5
u1 b1 b1 b1 0 0
u2 0 0 0 b2 b2
u3 b1 b1 b1 b2 b2
u4 b3 b3 0 0 0
u5 b3 b3 0 b2 b2
r1 r2 r3
u1 1 0 0
u2 0 1 0
u3 1 1 0
u4 0 0 1
u5 0 1 1
c1 c2 c3 c4 c5
r1 b1
b1
b1
0 0
r2 0 0 0 b2
b2
r3 b3
b3
0 0 0
Nokia and Whu
Role Mining
We create a potential role tree based on FCA(Formal Concept Analysis) and map between the potential role tree and the manually created role ontology. 45%
68%
81%
72%
Nokia and Whu
Role Mining Experiment
Data Set
Random data
Questionnaire (230)
Flickr
Nokia and Whu
Role Mining Experiment
Performance test results
#Context=20, #Behavior=20 #User=200, #Behavior=16 #User=100, #Context=10
M1: our methodM2: traditional FCA method
Nokia and Whu
Trust Mining
We have three ways to calculate trust between users for service
recommending:
Only trust preference fragments of those users when they play same
role with myself
Trust those users who have similar role set
Synthesize two ways above
We consider role relations, user’s dependency to role, and the
weight of each role.
Build trust network among users to provide trust propagation.
Nokia and Whu
Trust Mining
User1 …… .Usern
User1
……
Usern
Problem:How to decrease calculating times?
Nokia and Whu
Trust Mining
A B D E
A B C E F
User W
User X
{A8, B7, C5, D4, E3, F3}Role Set: Calculate weight of each role With IDF and rank
( ),
( )
ii j
j
W aa w x a w x
W a
8 7 30.6
8 7 5 4 3 3
0.63 0.32 0.14
0.69 0.42 0.23 ……..
W(0.63) X(0.69) … …. ….
W(0.32) X(0.42) Y(0.68) …. ….
… … ..
ABC
Nokia and Whu
Trust Mining Experiment
Data Set:
Epinion (have trust relationship information)
Verify the trust is higher between those users with similar role set
Nokia and Whu
Trust Mining Experiment
Performance test:
Similarity between friends : 0.37 -> 0.51
No role with role
Nokia and Whu
Trust Mining Experiment
MAE&RMSE test comparing with traditional approach:
Thank You !