A Social Recommendation Framework Based on Multiscale Continuous Conditional Random Fields CRF Xin X

1
A Social Recommendation Framework Based on
Multi-scale Continuous Conditional Random Fields
(CRF)Xin Xin

• 2009-8-27

2
Outline

• Motivation
• Problem Formulation
• Our Approach
• Single-scale continuous CRF
• Multi-scale continuous CRF
• Features
• Algorithm
• Experiments
• Conclusion

3
Motivation-Recommendation
In this example, there are four users, denoted
by ul and seven items, denoted by im. rlm is
rating record by ul to im (e.g., scale from 1 to
5, higher value means better satisfaction). The
collaborative filtering (CF) algorithms predict
values of unrated user-item pairs, denoted as
ylm, and suggest top ranked items as
recommendations.
4
Motivation-Traditional Methods

• Content
• Melville 02
• Memory
• User-based
• Breese 98, Jin 04
• Item-based
• Deshpande 04, Sarwar 01
• Hybrid
• Wang 06, Ma 07
• Model
• Hofmann 04, Si 03, Eerika 09
• State-of-the-art methods
• Memory-based
• Ma 07
• Model-based
• Zhu 08

• New Features in Social Recommendation
• Relational Dependency
• Multiple Information

5
Limitation1Lack of Relational Dependency within
Predictions
i1
i2
i3
i4
i5
i6
i7
u1
How to model the relational dependency within
predictions in recommendation algorithms?
u2
u3
u4

• Similarity Fusion (Wang 06)
• Difficult to measure similarity between y35 and
  y43.
• Cannot guarantee the nearness between y35 and y33
  (or y35 and y45 )
• EMDP (Ma 07)
• Error propagation

6
Limitation2Difficult to integrate features into
an unified approach
How to integrate various attribute and relational
features into an unified model?
Content information of items
Profile information of users
Trust relationship

• Classification Methods (Melville 2002)
• Ratings cannot be predicted
• Linear Integration (Ma 2007)
• Difficult to calculate the feature function
  weights

7
Motivation Multi-scale Continuous CRF

• Relational dependency within predictions can be
  modeled by the Markov property
• Feature function weights are globally optimized

single-scale
multi-scale
8
Motivation Our Contributions

• We formulate the problem of social
  recommendations and propose a Multi-scale
  Continuous CRF approach as a framework.
• We propose a MCMC-based optimization algorithm in
  both training and inference processes.

9
Outline

• Motivation
• Problem Formulation
• Our Approach
• Single-scale continuous CRF
• Multi-scale continuous CRF
• Features
• Algorithm
• Experiments
• Conclusion

10
Problem Formulation
Let X denote observations which can be existing
rating records, trust information, similarities
between different users/items, etc. Let Y denote
predictions with ylm denoting the prediction of
item m by user l.
Traditional Recommendation
Social Recommendation
11
Outline

• Motivation
• Problem Formulation
• Our Approach
• Single-scale continuous CRF
• Multi-scale continuous CRF
• Features
• Algorithm
• Experiments
• Conclusion

12
Single-scale CRF
feature example
Avg. Rate for item m
Similarity between item m and item n
13
Multi-scale CRF
feature example
Avg. Rate for user l
Similarity between item m and item n
trust between user l and user j
14
Features
State features
Avg. rate of variant occupation
Avg. rate of variant ages and gender
Avg. rate of variant genres
Edge features
similarity among items
similarity among users
trust relation
15
Algorithm-Train
Optimization Object
Gradient
Gibbs Sampling
16
Algorithm-Inference
Maximum Object
Simulated Annealing
17
Outline

• Motivation
• Problem Formulation
• Our Approach
• Single-scale continuous CRF
• Multi-scale continuous CRF
• Features
• Algorithm
• Experiments
• Conclusion

18
Experiment-Targets

• How about the overall performance of our proposed
  approach comparing with traditional and
  state-of-the-art CF methods?
• How does the relational dependency in predictions
  affect the accuracy of recommendation results?
• How do the features we combined from previous
  work affect the recommendation results?
• How about the computing complexity of MCCRF?

19
Experiment-Setup

• Datasets
• MovieLens
• Epinions
• Metrics
• MAE
• RMSE
• Baselines
• EPCC combination of UPCC and IPCC (memory)
• Aspect Model (AM) classical latent method
  (model)
• Fusion directly find similar users similar
  items
• EMDP two rounds prediction

20
Experiment-Overall Performance
21
Experiment-Dependency Effectiveness
MovieLens
Epinions
22
Experiment-Feature Effectiveness
MovieLens
Epinions
23
Experiment-Complexity

• Sampling times is the key factor
• Iteration times
• Temperature schema
• In our experiments O(mn)

Epinions
24
Experiments-Cluster in Epinions

• To run the whole data will take too much memory
  in large datasets like Epinions
• Xue et al 05 propose to employ cluster methods
  to solve this problem
• The figures show the impact of cluster size in
  Epinions using K-means cluster methods

25
Outline

• Motivation
• Problem Formulation
• Our Approach
• Single-scale continuous CRF
• Multi-scale continuous CRF
• Features
• Algorithm
• Experiments
• Conclusion

26
Conclusion

• We propose a novel model MCCRF as a framework for
  social recommendation.
• We propose a MCMC-based method for training and
  inference.
• Experimental verification.

27

• Thanks
• xxin_at_cse.cuhk.edu.hk