Temporal recommendation on graphs via long and short-term

Temporal recommendation on graphs via long- and short-term preference fusion

Liang Xiangxlvector@gmail.com

Main Content

• Temporal Recommendation– Long/short term preference

• Bipartite Graph Model– Session Graph Model– Path Fusion Algorithm

Related Works

• Neighborhood Model [Ding CIKM05]– Users future preference is mainly dependent on

their recent behavior• Latent Factor Model [Koren KDD09]– User bias shifting– Item bias shifting– User preference shifting– Seasonal effects

Our Contribution

• Temporal Recommendation on Graph Model– Implicit feedback data

• Combine Long/short term interest together

Graph Model Temporal Recommendation

Long/Short Term Preference

Short-term PreferenceLong-term Preference

Long/Short Term Preference

• Long term preference– Personal preference– Do not change frequently– Last for long period

• Short term preference– Influenced by social event– Change frequently– May be become long term preference

Session Graph Model

Session Graph Model

A

B

a

b

c

(A,a,1) (A,c,2)(B,b,1) (B,c,2)

A

B

a

b

c

A:1

A:2

B:1

B:2

Bipartite Graph Model Session Graph Model

Session Node

User Node

Item Node

Session Graph Model

Session Node

User Node

Item Node

1

1

1

1

( )

(1 )

i

u

uT

v v

v v v

v v

Ranking and Recommendation

Path Fusion Ranking

• Two nodes in a graph have large similarity if:– There are many paths between two nodes;– These paths have short length;– Most of these paths do not contains nodes with

large out degree.

[YouTube WWW2008]

Path Fusion Ranking

A

B

a

b

c

11

1

( ) ( , )( )

| ( ) |

Ni i i

i i

v w v vweight P

out v

( , ')

( , ') ( )P path v v

d v v weight P

( ) ( , ) ( ) ( , ) ( ) ( , )( , , , )

| 2 | | 2 | | 2 |

A w A c c w c B B w B bweight A c B b

Path Fusion Ranking1. Implement by Breath-First-Search2. Fast and low space complexity

a) Its speed dependents on graph sparsity;

b) It can be speed up by randomly select edges;

c) Do not need to store user-user or item-item similarity matrix

3. Easy to do incremental updatea) New data can insert into graph

directly;b) After graph is updated,

recommendation result will be changed immediately

Experiments

Experiments

Experiments

This model does not work in every system!

Future work

Temporal Effectiveness

Slow Evolution SystemSession Graph Model Perform Good

Fast Evolution SystemSession Graph Model Perform Bad

Temporal Effectiveness

0 10 20 30 40 50 600

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

nytimes youtube wikipediasourceforge blogspot netflix

Solution

• Add Item Session Node

A

B

a

b

c

A

B

a

b

c

A:1

A:2

B:1

B:2

A

B

a

b

c

A:1

A:2

B:1

B:2

a:1

b:1

c:2

(A,a,1) (A,c,2)(B,b,1) (B,c,2)