Yong Zheng, Mayur Agnani, Mili Singh. “Identifying Grey Sheep Users By The Distribution of User Similarities In Collaborative Filtering”. Proceedings of The 6th ACM Conference on Research in Information Technology (RIIT), Rochester, NY, USA, October, 2017

1. Yong Zheng, Mayur Agnani, Mili Singh
School of Applied Technology
Illinois Institute of Technology
Chicago, IL, 60616, USA
Identifying Grey Sheep Users By The Distribution of User
Similarities In Collaborative Filtering

2. Agenda
• Background: Recommender Systems
• Grey Sheep Users In Collaborative Filtering
• Methodology and Solutions
• Experimental Results
• Conclusions and Future Work
2

3. Agenda
• Background: Recommender Systems
• Grey Sheep Users In Collaborative Filtering
• Methodology and Solutions
• Experimental Results
• Conclusions and Future Work
3

4. Information Overload
4

5. Alleviating Information Overload
5
Information Extraction and Filtering
Query
Information
Retrieval
Recommender
Systems

6. Recommender System (RS)
• RS: item recommendations tailored to user tastes
6

7. Recommender Systems
7
E-Commerce: Amazon.com, eBay.com, BestBuy, NewEgg

8. Recommender Systems
8
Online Streaming: Netflix, Pandora, Spotify, Youtube, etc

9. Recommender Systems
9
Social Media: Facebook, Twitter, Weibo, etc

10. How it works
10
Red
Mars
Juras-
sic
Park
Lost
World
2001
Found
ation
Differ-
ence
Engine
Recommender
Systems
User
Profile
Neuro-
mancer
2010
Recommendations

11. Traditional Recommendation Algorithms
11
Content-Based Recommendation Algorithms
The user will be recommended items similar to the ones the
user preferred in the past, such as book/movie recsys
Collaborative Filtering Based Recommendation Algorithms
The user will be recommended items that people with similar
tastes and preferences liked in the past, e.g., movie recsys
Hybrid Recommendation Algorithms
Combine content-based and collaborative filtering based
algorithms to produce item recommendations.

12. Collaborative Filtering
1212

13. Collaborative Filtering: Algorithms
13
User-Based KNN Collaborative Filtering (UBCF)
Assumption: a user u’s rating on item t is similar to
other users’ rating on item t, while this group of
similar users is called user K-nearest neighbor
Pirates of the
Caribbean 4
Kung Fu Panda 2 Harry Potter 6 Harry Potter 7
U1 4 4 1 2
U2 3 4 2 1
U3 2 2 4 4
U4 4 4 1 ?

14. Collaborative Filtering: Algorithms
14
User-Based KNN Collaborative Filtering (UBCF)
Pirates of the
Caribbean 4
Kung Fu Panda 2 Harry Potter 6 Harry Potter 7
U1 4 4 1 2
U2 3 4 2 1
U3 2 2 4 4
U4 4 4 1 ?
a = the target user
i = the target item
N = user neighborhood
u = a user neighbor in N

15. Collaborative Filtering: Algorithms
15
Popular Challenges in Collaborative Filtering
Data sparsity problems
Cold-start users or items
Grey-sheep users
Incorporate content into collaborative filtering
….

16. Agenda
• Background: Recommender Systems
• Grey Sheep Users In Collaborative Filtering
• Methodology and Solutions
• Experimental Results
• Conclusions and Future Work
16

17. Grey Sheep Users
17
Definition 1 by Mark Claypool, et al., 1999
A group of users who neither agree nor disagree with any group
of users. Therefore, they will not benefit from the user-based
collaborative filtering technique
Definition 2 by John McCrae, et al., 2004
White Sheep Users: high correlations with other users
Black Sheep Users: very few or no correlating users
Grey Sheep Users: unusual tastes, low correlations with others

18. Research Problem: Identifying Grey Sheep Users
18
Collaborative Filtering Other Algorithms

19. Existing Approaches
19
There are two existing Approaches
Clustering Technique by Ghazanfar, et al., 2011
Distribution of User Ratings by Gras, et al., 2016
They were developed based on the 1st definition
We propose a novel approach based on 2nd definition
White Sheep Users: high correlations with other users
Black Sheep Users: very few or no correlating users
Grey Sheep Users: unusual tastes, low correlations with others
Note: Black Sheep User refers to the user whom we do not have enough knowledge for.

20. Agenda
• Background: Recommender Systems
• Grey Sheep Users In Collaborative Filtering
• Methodology and Solutions
• Experimental Results
• Conclusions and Future Work
20

21. Proposed Solution
21
We propose a novel approach based on 2nd definition
White Sheep Users: high correlations with other users
Black Sheep Users: very few or no correlating users
Grey Sheep Users: unusual tastes, low correlations with others
The Distribution of User-User Correlations or Similarities

22. Proposed Solution
22
Proposed Solution
Step 1, represent each user as distribution of user correlations
Step 2, select good and bad examples
Step 3, apply outlier detection on selected examples. Grey sheep
users are the intersections of bad examples and identified
outliers
Step4, examine the quality of identified grey sheep users

23. Proposed Solution
23
Step 1, Distribution Representations
We calculate user-user correlations by cosine similarity
Obtain the descriptive statistics of the distribution

24. Proposed Solution
24
Step 2, Example Selection
Good examples: high correlations and left-skewed
Bad examples: low correlations and right-skewed

25. Proposed Solution
25
Step 3, Outlier Detection by Local Outlier Factor (LOF)
LOF helps identify outliers by the local density
Observations with LOF > 1 will be considered as outliers
We set different threshold values to find
the optimal one for identifying grey sheep
users, for example
LOF threshold = 1.0
LOF threshold = 1.1
LOF threshold = 1.2
LOF threshold = ….

26. Proposed Solution
26
Step 4, Examine the quality of identified GS Users
The parameters in our solution
Example Selection
LOF threshold
Neighbor of neighborhood in LOF method
Our goals or examination criteria
To find as many GS users as possible
Recommendation by UBCF should be worse for GS users than non-
GS users

27. Agenda
• Background: Recommender Systems
• Grey Sheep Users In Collaborative Filtering
• Methodology and Solutions
• Experimental Results
• Conclusions and Future Work
27

28. Experimental Setting
• Data: MovieLens 10 million rating data
– 10M ratings
– 72K users
– 10K movies
– Each user has rated at least 20 movies
• Evaluation
– 80% as training, 20% as testing
– Mean absolute error, MAE, to eval rating predictions
28

29. Results
• Impact by the # of neighbors in LOF method
29

30. Results
• Comparison of Recommendation Quality
30

31. Results
• Visualization of GS and Non-GS users
31

32. Agenda
• Background: Recommender Systems
• Grey Sheep Users In Collaborative Filtering
• Methodology and Solutions
• Experimental Results
• Conclusions and Future Work
32

33. Conclusions
• We develop a novel approach to identify GS
users by utilizing the definition related to the
user-user correlations
• Our approach can successfully identify GS users
• Our approach is less complicated than the
existing approaches
33

34. Drawbacks and Future Work
• We did not compare our solution with the two
existing methods
• The user-user correlations may not be reliable
if the rating data is sparse
34

35. Stay Tuned
• Yong Zheng, Mayur Agnani, Mili Singh. “Identification of Grey
Sheep Users By Histogram Intersection In Recommender Systems”.
Proceedings of The 13th International Conference on Advanced
Data Mining and Applications, 2017
– We improve the proposed solution in the RIIT paper
– We better measure user-user correlations
– We compare our solution with the two existing methods and
demonstrate the advantages and effectiveness of our proposed
solution
35

36. Yong Zheng, Mayur Agnani, Mili Singh
School of Applied Technology
Illinois Institute of Technology
Chicago, IL, 60616, USA
Identifying Grey Sheep Users By The Distribution of User
Similarities In Collaborative Filtering