[IUI 2017] Criteria Chains: A Novel Multi-Criteria Recommendation Approach
1. Criteria Chains: A Novel Multi-Criteria
Recommendation Approach
Yong Zheng
Illinois Institute of Technology
Chicago, IL, 60616, USA
ACM Conference on Intelligent User Interfaces
Limassol, Cyprus, March 13-16, 2017
3. Traditional RS: Ratings By Users on Items
3
Red
Mars
Juras-
sic
Park
Lost
World
2001
Found
ation
Differ-
ence
Engine
Recommender
Systems
User
Profile
Neuro-
mancer
2010
Recommendations
6. 6
Multi-Criteria Recommender Systems
• Traditional RS:
• Multi-Criteria RS:
R0 is a user’s overall rating on the item. R1, R2, …, Rk are ratings on item aspects.
7. 7
Multi-Criteria Recommender Systems
Research Problems in Multi-Criteria RS
Step2
Multi-Criteria RatingsStep 1 Step 1
Step 1. Learn from knowledge to predict multi-criteria ratings
Step 2. Aggregate multi-criteria ratings to predict the overall rating.
Linear Regression:
8. 8
Multi-Criteria Recommender Systems
There are two solutions to improve it:
• Improve the predicted multi-criteria ratings
• Better utilize them to estimate the overall rating
The contributions by Criteria Chains:
• Better predict multi-criteria ratings
• Figure out a new way to aggregate these ratings
9. 9
Criteria Chains
Assumptions in Criteria Chains
• Multi-criteria ratings can be viewed as contexts
• Ratings can be predicted in a chain
10. 10
Criteria Chains
Assumptions in Criteria Chains
• Multi-criteria ratings can be viewed as contexts
• Ratings can be predicted in a chain
First, predict U3’s rating on Room
Next, take U3’s rating on room as contexts, User + Item + Room Check-in
Again, take previous predictions as contexts, User + Item + Room + Check-in Service
The prediction process works like a chain: Room Check-in Service
11. 11
Criteria Chains
The sequence of the chain matters
• Random Sequence
• Rank by Lower Prediction Errors
• Rank by Information Gain
12. 12
Criteria Chains
How to predict the final overall rating?
• Criteria Chain: Aggregation Model (CCA)
Linear regression by predicted multi-criteria ratings
• Criteria Chain: Contextual Model (CCC)
Direct prediction by viewing the predicted multi-
criteria ratings as context information
• Criteria-Independent Contextual Model (CIC)
This is a baseline approach. We predicted multi-
criteria ratings independently and use them as
context to predict the final overall rating
13. 13
Experimental Setting
• Data Sets
• Evaluations
– Five-fold Cross Validation
– Rating Prediction: Mean Absolute Error (MAE)
– Top-N Recommendation: Precision, Recall, NDCG
– We use CAMF_C for context-aware recommendations
User Item Rating # of Criteria
TripAdvisor 1,502 14,300 22,130 7
Yahoo!Movie 2,162 3,075 49,351 4
17. 17
Conclusions and Future Work
• Criteria Chains work better than baseline approaches
• Criteria Chains take correlations among multiple
criteria into consideration
• Information Gain is the best way to produce chain
• Using multi-criteria ratings as contexts, CCC, is the
best approach after predictions on multiple ratings
• Future Work: figure out optimal ways to generate the
chain sequence in addition to information gain.
18. Criteria Chains: A Novel Multi-Criteria
Recommendation Approach
Yong Zheng
Illinois Institute of Technology
Chicago, IL, 60616, USA
ACM Conference on Intelligent User Interfaces
Limassol, Cyprus, March 13-16, 2017