tangent tangent a novel, “surprise-me”, recommendation algorithm kensuke onuma : sony...
TRANSCRIPT
TANGENTTANGENTA Novel, “Surprise-me”, Recommendation Algorithm
Kensuke Onuma : Sony CorporationHanghang Tong : Carnegie Mellon Univ.
Christos Faloutsos : Carnegie Mellon Univ.
2
Motivation
Go off on a ‘TANGENT’ !
Movies
KevinKevinLizLiz
TimTim
MarkMark
JessicaJessica
MaryMary
JohnJohn
RachelRachel
BobBobMikeMike
TomTom
Broadening users’ horizon
More chance to increase sales of items
3
What we want are …
user
movie
comedy fans horror fans
Conventional recommendationalgorithms’ answer TANGENT’s answer
A
target user(= query node)
4
Outline• Motivation• Problem definition• Algorithm• Experiments• Conclusion
5
Graphs for recommendation[bipartite graph]
John Mike
A B C D
Mark Rachel Tom Mary
E F G H
1V
2V
E
),( EVG : weighted based on rating
V E: users and movies 21,VVV
6
Problem definition of TANGENTGiven: - An edge-weighted undirected graph with adjacency matrix - The set of query nodes
Given: - An edge-weighted undirected graph with adjacency matrix - The set of query nodes
Find: - A node that satisfy following conditions. (1) Close enough to (2) Possessing high potential to reach other nodes
Find: - A node that satisfy following conditions. (1) Close enough to (2) Possessing high potential to reach other nodes
AG
kiiqQ 1)(
Q
QG
user
movie
7
Outline• Motivation• Problem definition• Algorithm• Experiments• Conclusion
8
Outline of TANGENT algorithm
1. Calculate relevance score of each node to 2. Calculate bridging score of each node3. Compute the TANGENT score
by merging two criteria above
Qr
Q
b
Qt
Quser
movie
9
[Step 1] Relevance scoreRandom walk with restart [Pan+ KDD ’04]
1
2
3
4 5
6
7
8
9
querynode
node
1 0.577
2 0.132
3 0.123
4 0.123
5 0.036
6 0.001
7 0.006
8 0.001
9 0.001
ir ,1
R nrrr
21
Various Scalable Solution [Tong ’06] - OnTheFly - B_Lin - NB_Lin - BB_Lin (for bipartitle graph)
Various Scalable Solution [Tong ’06] - OnTheFly - B_Lin - NB_Lin - BB_Lin (for bipartitle graph)
10
[Step 2] Bridging score (Intuition)
1
2 34
5
7 6
12
34
5
7
6
7,77,67,57,47,37,2
6,76,66,56,46,36,2
5,75,65,55,45,35,2
4,74,64,54,44,34,2
3,73,63,53,43,33,2
2,72,62,52,42,32,2
1
rrrrrr
rrrrrr
rrrrrr
rrrrrr
rrrrrr
rrrrrr
R
7,77,67,57,47,37,2
6,76,66,56,46,36,2
5,75,65,55,45,35,2
4,74,64,54,44,34,2
3,73,63,53,43,33,2
2,72,62,52,42,32,2
1
rrrrrr
rrrrrr
rrrrrr
rrrrrr
rrrrrr
rrrrrr
R
a node in a group a node between groups
~0
~0
1b 1bsmall large
11
[Step 2] Bridging score (Detail)
),,,,,mean(
1
6,44,66,22,64,22,41 rrrrrrb
1
2 3
4
neighbors
6,66,46,2
4,64,44,2
2,62,42,2
rrr
rrr
rrr
1R
1Sr
12
[Step 3] TANGENT score
Si
jQjjQjQ r
rbrt
ˆ,
,,
A. Simple multiplication. (not linear combination, not skyline query, )
user
movie
query
relevance scoreto query nodes
relevance scoreamong neighbors
13
Example
1
2
3
4 5
6
7
8
9
node
1 0.577 8.579 4.949
2 0.132 8.579 1.129
3 0.123 11.085 1.362
4 0.123 11.085 1.362
5 0.036 20.789 0.755
6 0.001 7.967 0.010
7 0.006 12.847 0.074
8 0.001 7.967 0.010
9 0.001 7.967 0.010
querynode
ibir ,1 it ,1
Group 1 Group 2
14
Outline• Motivation• Problem definition• Algorithm• Experiments
– Synthetic data– Real data
• MovieLens (user-movie)• DBLP (author-paper)
• Conclusionon our paper
15
Synthetic data[bipartite graph]
1 2 3 4 5 6 7 8 9
10 11 1213 14 15 16 17 18 19 20 21 22 23
24 25 26
query No.1 in TANGENT
node 1 node 16
node 5 node 20
node 12 node 20
16
Real data [MovieLens]User Preference (rating 5)- A Nightmare on Elm Street (1984) (Horror)- The Shining (1980) (Horror)- Jaws (1975) (Action, Horror)
Rank Title Genre
1 The Silence of the Lamb (1991) Dr, Thr
2 Psycho (1960) Hor, Rom, Thr
3 Pulp Fiction (1994) Cr, Dr
4 An American Werewolf in London (1981)
Hor
5 Natural Born Killers (1994) Ac, Thr
6 Carrie (1976) Hor
7 Alien (1979) Ac, Hor, SF, Thr
8 Twelve Monkeys (1995) Dr, SF
9 Evil Dead II (1987) Ac, Ad, Com, Hor
10 Scream (1996) Hor, Thr
15 Star Wars (1977) Ac,Adv,Rom,SF,War
17 Fargo (1996) Cr, Dr, Thr
22 The Godfather (1972) Ac, Cr, Dr
45 Contact (1997) Dr, SF
Rank Title Genre
1 The Silence of the Lambs (1991) Dr, Thr
2 Scream (1996) Hor, Thr
3 Pulp Fiction (1994) Cr, Dr
4 Star Wars (1977) Ac, Adv, Rom, SF, War
5 Fargo (1996) Cr, Dr, Thr
6 Twelve Monkeys (1995) Dr, SF
7 Psycho (1960) Hor, Rom, Thr
8 The Godfather (1972) Ac, Cr, Dr
9 Contact (1997) Dr, SF
10 Alien (1979) Ac, Hor, SF, Thr
13 An American Werewolf in London (1981)
Hor
12 Natural Born Killers (1994) Ac, Thr
16 Carrie (1976) Hor
23 Evil Dead II (1987) Ac, Ad, Com, Hor
Ranked list by relevance score Ranked list by TANGENT score
943 users1682 movies55375 ratings
17
Rank Title Genre
1 The Flintstones (1994) Ch,Com
2 Spy Hard (1996) Com
3 Oliver & Company (1988) Ani,Chi
4 Jack (1996) Com,Dr
5 Son in Law (1993) Com
6 Ace Ventura: When Nature Calls (1995)
Com
7 Renaissance Man (1994) Com,Dr,War
8 Pocahontas (1995) Ani,Chi,Mus,Rom
9 Corrina, Corrina (1994) Com,Dr,Rom
10 Beverly Hillbillies, The (1993) Com
11 Princess Bride, The (1987) Ac,Adv,Com,Rom
15 Monty Python and the Holy Grail (1974)
Com
21 Empire Strikes Back, The (1980) Ac,Adv,Dr
26 Raiders of the Lost Ark (1981) Ac,Adv
29 Return of the Jedi (1983) Ac,Adv,Rom,SF,War
32 Star Wars (1977) Ac,Adv,Rom,SF,War
42 Toy Story (1995) Ani,Chi,Com
53 Men in Black (1997) Com,Dr
Rank Title Genre
1 Star Wars (1977) Ac,Adv,Rom,SF,War
2 Return of the Jedi (1983) Ac,Adv,Rom,SF,War
3 The Princess Bride (1987) Ac,Adv,Com,Rom
4 Toy Story (1995) Ani,Chi,Com
5 Monty Python and the Holy Grail (1974)
Com
6 Spy Hard (1996) Com
7 Raiders of the Lost Ark (1981) Ac,Adv
8 Empire Strikes Back, The (1980) Ac,Adv,Dr
9 Jack (1996) Com,Dr
10 Men in Black (1997) Ac,Adv,Com,SF
25 Ace Ventura: When Nature Calls (1995)
Com
27 Corrina, Corrina (1994) Com,Dr,Rom
35 Son in Law (1993) Com
42 Oliver & Company (1988) Ani,Chi
43 Renaissance Man (1994) Com,Dr,War
52 Pocahontas (1995) Ani,Chi,Mus,Rom
166 The Beverly Hillbillies (1993) Com
1439 The Flintstones (1994) Ch,Com
relevance score TANGENT score
User Preference (rating 5)- Robin Hood: Men in Tights (1993) (Comedy)- Young Frankenstein (1974) (Comedy, Horror)- Naked Gun 33 1/3: The Final Insult (1994) (Comedy)- Fatal Instinct (1993) (Comedy)
18
Outline• Motivation• Problem definition• Algorithm• Experiments• Conclusion
19
Conclusion• Definition of a novel recommendation problem
– “how to make a recommendation that broadens the horizons of the user?”
– [Approach]* close to the user preferences * have high connectivity to other groups
• Design of algorithm– “Relevance score” X “Bridging score”– Effective & Efficient
• Experiments– synthetic dataset– real dataset
20
Thank youKensuke [email protected]
Hanghang [email protected]
Christos [email protected]
Poster tonight !19:30 – 22:00
at Hôtel de Ville
Code availablehttp://www.cs.cmu.edu/~kensuke/