Presentation from ECML-PKDD 2017 of joint work with my supervisors Helge Langseth and Heri Ramampiaro, about a Poisson factorization model for recommendations that integrates topic models in the items set and social connections between users. code: https://github.com/zehsilva/poissonmf_cs preprint version: https://inajourney.files.wordpress.com/2012/11/poissonmf_cs.pdf

1. Content-Based Social
Recommendation with Poisson
Matrix Factorization
Eliezer de Souza da Silva
PhD student
Department of Computer Science, NTNU
Joint work with Helge Langseth and Heri Ramampiaro
ECML-PKDD 2017

2. 2
Introduction
● Basic Problem: Recommendation of items to users given user interaction
with some items
User 1 User 2 User 3
Item 1 Item 2 Item 3 Item 4

3. 3
Challenges and opportunities in RS
Research
User 1 User 2 User 3
Item 1 Item 2 Item 3 Item 4 Item 5
Topics

4. 4
Challenges and opportunities in RS
Research
User 1 User 2 User 3
Item 1 Item 2 Item 3 Item 4 Item 5
Topics
User 4
User
Social
Network

5. 5
Challenges and opportunities in RS
Research
● Incorporate
○ Social network analysis tools and methods
○ Content analysis (topic models,
sentiment/intent/mood)
○ New rich contextual information
■ location, activity, user intent/goal, etc.

6. 6
Joint modelling of user social
network and item topic content
● User social network
○ Homophily
○ Item exposure positively influenced by peers
(positive “peer-pressure”)
● Item content analysis
○ Enrich items latent factors with topic model
○ Cold start items
○ Preferences can be influenced by topics

7. 7
[Topics, Words]
[Topics, Users]
[Items, Topics]
[Items, Words]
[Items, Users]
Observed
Latent

8. 8
Poisson Matrix Factorization with Content and Social
trust information (PoissonMF-CS)

9. 9
Items Topic Model

10. 10
User preference and social factors

11. 11
Poisson Matrix Factorization with Content and Social
trust information (PoissonMF-CS)

12. 12
Inference
• Batch variational inference:
• Conjugate model with auxiliary variable “tricky” for each
Poisson likelihood term:
• Running time for each iteration depends on the sparse
observations:
– O(K(obs_W + obs_R + obs_S + U + D + W ))

13. 13
Item Recommendations
• Top-M items for each user:
– Approximate expected value of user-item matrix for each
unseen item for ranking
Rud User preferences
Shared item topic
intensity
Item topic offset
Weighted sum of social
network neighbors
interactions with item

14. 14
Application
Artist recommendation (Last-fm dataset):
• User-artist interactions counts
• User-user social network
• Artist-tags counts
Dataset size:
– 1892 users, 17632 artists, 11946 tags
– 25434 user–user connections, 92834 user–items interactions,
and 186479 user–tag–items entries.

15. 15
Results
● Avg. Recall Metric:
● Compare with previous work:
○ Collaborative Topic Regression (CTR)
○ Collaborative Topic Regression with Social Matrix Factorization
(CTR-SMF)
○ Collaborative topic Poisson factorization (CTPF)
○ Social Poisson Factorization (SPF)

16. 16
Results
PoissonMF-CS (K =10) and
Gaussian-based models
PoissonMF-CS (K =10) and other Poisson
factorization models

17. 17
Results

18. 18
Conclusion
• Model including social and topic information in Poisson
matrix factorization using coupled latent factors
• Inference is computationally efficient with variational
inference
• Future work:
– Non-negative relational learning
– Non-parametric extensions
– Scalable inference (SVI)

19. 19
Questions?
https://github.com/zehsilva/poissonmf_cs

20. 20
Content-based Social Poisson Factorization for recommendation