Improve General Contextual SLIM Recommendation Algorithms By Factorizing Contexts

1. Improve General Contextual SLIM Recommendation
Algorithms By Factorizing Contexts
Yong Zheng, PhDc
Center for Web Intelligence
DePaul University, Chicago, USA
Student Research Competition
2nd Round@ACM SAC 2015
Salamanca, Spain, April 15

2. Overview
• Background (Recommender System and Context-aware RS)
• Research Problems and Motivations
• Solutions by Factorizing Context
• Experimental Results
• Conclusion and Future Work

3. Background
• Recommender Systems
• Context-aware Recommender Systems

4. Intro. Recommender Systems
• Recommender System (RS)
Recommendation

5. Intro. Recommender Systems
• Recommender System (RS)
Social RS (Twitter) Tagging RS (Flickr)

6. Intro. Recommender Systems
• Recommender System (RS)

7. Intro. Recommender Systems
• Typical Data Set in RS: Rating-Based Data Set
Usually, it is a 2D rating matrix: User × Item -> Ratings
Task: For a user, RS provide a list of suggested items to him/her

8. Context-aware RS (CARS)
• Example of Context-aware Recommender
What is context? “any information that can be used to characterize
the situation of an entity” by Abowd et al. in 1999
Companion
Example of Contexts in different domains:
 Food: time (noon, night), occasion (business lunch, family dinner)
 Movie: time (weekend, weekday), location (home, cinema), etc
 Music: time (morning, evening), activity (study, sports, party), etc
 Book: a book as a gift for kids or mother, etc

9. Context-aware RS (CARS)
• Traditional RS: Users × Items  Ratings
• Context-aware RS: Users × Items × Contexts Ratings
Task: CARS provide a list of suggests to <user, contexts>
Recommendation cannot live alone without considering contexts.
E.g., choose a romantic movie with partner, but comic with kids.

10. Research Problems
• Contribution of This Work
• Research Problem and Motivations

11. Contribution of This Work
Before moving on, it is necessary to introduce the
contribution of this work: Basically, the work in this
paper is an improvement over a context-aware
recommendation algorithm which was our previous
work published in ACM CIKM 2014 and ACM RecSys
2014 conferences.
To further introduce the following work:
1). Introduce the General Contextual Sparse LInear
Method (GCSLIM) published in ACM CIKM 2014
2). Introduce the drawbacks in GCSLIM and
research problems in this work

12. Intro. GCSLIM
General Contextual Sparse LInear Method (GCSLIM)
published in ACM CIKM 2014
Recommendation task: given user U1, and a context
situation {weekday, home, sister}, the system should
provide a list of recommended movies to U1.
How to generate such a recommendation list?
User Movie Time Location Companion Rating
U1 M1 Weekend Home Girlfriend 4
U2 M2 Weekday Home Girlfriend 5
U3 M3 Weekday Cinema Sister 4
U1 M2 Weekday Home Sister N/A

13. Intro. GCSLIM
Task: Recommendation to <U1, Ctx2>
Step1: we extract U1’s contextual rating on other movies
from P, e.g., R (U1, t1, Ctx1)
Step 2: This rating will be converted into R (U1, t1, Ctx2) by
adding the rating deviation Dev(Ctx1, Ctx2)
Step 3: Right now, we have estimated rating R (U1, t1,
Ctx2) , we multiply it with coefficient W (t1, t2)
S(U1, t2, Ctx2) = an aggregation of the term below:
[R(U1, t1, Ctx1) + Dev (Ctx1, Ctx2)] × W(t1, t2)

14. Intro. GCSLIM
S(U1, t2, Ctx2) = an aggregation of the term below:
[R(U1, t1, Ctx1) + Dev (Ctx1, Ctx2)] × W(t1, t2)
Ctx1 = {weekend, home, girlfriend}
Ctx2 = {weekday, home, sister}
Dev (Ctx1, Ctx2) =
Dev (weekend, weekday) + Dev (girlfriend, sister)
User Movie Time Location Companion Rating
U1 M1 Weekend Home Girlfriend 4
U2 M2 Weekday Home Girlfriend 5
U3 M3 Weekday Cinema Sister 4
U1 M2 Weekday Home Sister N/A

15. Intro. GCSLIM
How good is the GCSLIM?????????
In our previous work published in ACM CIKM 2014, it
has been demonstrated that GCSLIM outperforms the
state-of-the-art context-aware recommendation
algorithms within multiple data sets.
In other words, GCSLIM is the BEST context-aware
recommendation algorithm based on our empirical
experimental evaluations in 2014.

16. Drawbacks in GCSLIM
Ctx1 = {weekend, home, girlfriend}
Ctx2 = {weekday, home, sister}
Dev (Ctx1, Ctx2) =
Dev (weekend, weekday) + Dev (girlfriend, sister)
--------------------------------------------------------
Deviations is measured in pairs, e.g. Dev (weekend,
weekday) , which may result in sparsity problem.
For example:
In training set, we learned Dev<weekend,
weekday> and Dev <weekday, holiday>
But in testing set, we need Dev <weekend, holiday>
Which was NOT learned in the algorithm

17. Research Problem
Problem: how to alleviate the sparsity problem?
Sparsity Problem:
In training set, we learned Dev<weekend,
weekday> and Dev <weekday, holiday>
But in testing set, we need Dev <weekend, holiday>
Which was NOT learned in the algorithm

18. Solutions and Experimental Results
• Solution: Factorizing Context
• Experimental Results

19. Solution
Problem:
In training set, we learned Dev<weekend, weekday> and
Dev <weekday, holiday>
But in testing set, we need Dev <weekend, holiday>
Which was NOT learned in the algorithm
----------------------------------------------------------------
Solution:
We represent each context condition by a vector, e.g.,
Weekend = <0.1, 0, 0.03, 0.4, 1>
Weekday = <0.2, 0.3, 1.2, 0, 0.1>
Holiday = <1.22, 0.1, 0, 0.2, 2.3>
Dev (c1, c2) = Euclidean distance (Vector1, Vector2)
Note: other distance metrics may also be applied!!

20. Solution
Problem:
In training set, we learned Dev<weekend, weekday> and
Dev <weekday, holiday>
But in testing set, we need Dev <weekend, holiday>
Which was NOT learned in the algorithm
------------------------------------------------------------------
Solution:
Training: Dev<weekend, weekday> Dev <weekday, holiday>
Testing: Dev<weekend, holiday>
In training, we already learned those two pairs of deviations, where
their corresponding vectors have been learned!
In testing, we can directly use the distance of the vectors “weekend”
and “holiday” to compute the deviations!!  Alleviate the sparsity!!

21. Experimental Results
This approach has been evaluated over multiple data sets:
Due to limited space in the SRC paper, we just present our results
based on the restaurant data, where there are two contextual
dimensions: Time (weekend, weekday), Location (school, home, work)
We use two metrics: Precision, and Mean Average Precision (MAP)
Precision  measure how accurate the recommendation list is by
evaluating the hit-ratio
MAP  additionally measure the ranking positions in addition to
Precision

22. Experimental Results
We used two baselines
from our previous work
published in RecSys
and CIKM 2014.
We did not include other
baselines, since GCSLIM
was already proved as
the best one in our
previous work.
The improvement is 16%
on precision and 8% on
MAP for this data set.

23. It is gonna end…
• Conclusions
• Future Work

24. Conclusions
 Factorizing context is able to alleviate the sparsity problem in the
GCSLIM algorithm
Future Work
 Even if this solution can alleviate the sparsity problem, but it cannot
fully solve it when the data is too sparse, especially when it comes
to the cold-start problems: no knowledge about user/item/context.
Stay tuned ..
 Actually, the idea of factorizing context is also applicable to other
algorithms to alleviate the sparsity problem, since this problem is a
general one in this domain.
 We have a paper "Integrating Context Similarity with Sparse Linear
Recommendation Model" accepted by the UMAP 2015 which is the
premier conference in user modeling and personalization, where we
reused the approach of factorizing context.

25. Acknowledgement
Thanks to the ACM SIGAPP providing travel support;
Thanks to Microsoft Research providing travel support;
Thanks to the organizers of ACM SAC and the SRC program!

26. Thank You!
Center for Web Intelligence, DePaul University, Chicago, IL USA