1. CONTEXT-AWARE SIMILARITIES
WITHIN THE FACTORIZATION
FRAMEWORK
Balázs Hidasi
Domonkos Tikk
CARR WORKSHOP, 5TH FEBRUARY
2013, ROME

2. OUTLINE
 Background & scope
 Levels of CA similarity
 Experiments
 Future work

3. BACKGROUND
Implicit
feedback Context
I2I
Factorization
recommendation
This work

4. IMPLICIT FEEDBACK
 User preference not coded explicitely in data
 E.g. purchase history
 Presence  preference assumed
 Absence  ???
 Binary „preference” matrix
 Zeroes should be considered
 Optimize for
 Weighted RMSE
 Partial ordering (ranking)

5. CONTEXT
 Can mean anything
 Here: event context
 User U has an event
 on Item I
 while the context is C
 E.g.: time, weather, mood of U, freshness of I, etc.
 In the experiments:
 Seasonality (time of the day or time of the week)
 Time period: week / day

6. FACTORIZATION I
 Preference data can be organized into matrix
 Size of dimensions high
 Data is sparse
 Approximate this matrix by the product of two
low rank matrices
 Each item and user has a feature vector
 Predicted preference is the scalar product of the
item
appropriate vectors
user r
T
 r Uu Ii
u ,i
 Here we optimize for wRMSE (implicit case)
 Learning features with ALS

7. FACTORIZATION II.
 Context introduced  additional context
dimension
 Matrix  tensor (table of records)
item
 Models for preference prediction:
 Elementwise product model user r
T
WeightedU u  I i product
r 1 ( scalar  C c )
 u ,i , c

 Pairwise model
T T T
 ContextU u I i
ru , i , c dependent user/item I i C c
U u Cc bias

context
context
item
+ +
user user item r

8. ITEM-TO-ITEM RECOMMENDATIONS
 Items similar to the current item
 E.g.: user cold start, related items, etc.
 Approaches: association rules, similarity between
item consumption vectors, etc.
 In the factorization framework:
 Similarity between the feature vectors
Scalar product:s i , j
T
 Ii I j
T
Ii Ij
 Cosine similarity: s i , j
Ii 2
Ij
2

9. SCOPE OF THIS PROJECT
 Examination wether item similarities can be
improved
 using context-aware learning or prediction
 compared to the basic feature based solution
 Motivation:
 If factorization models are used anyways, it would be
good to use them for I2I recommendations as well
 Out of scope:
 Comparision with other approaches (e.g. association
rules)

10. CONTEXT-AWARE SIMILARITIES: LEVEL1
 The form of computing similarity remains
T
 si, j Ii Ij Ii
T
Ij
si, j
Ii 2
Ij
2
 Similarity is NOT CA
 Context-aware learning is used
 Assumption: item models will be more accurate
 Reasons: during the learning context is modelled
separately
 Elementwise product model
 Context related effects coded in the context feature vector
 Pairwise model
 Context related biases removed

11. CONTEXT-AWARE SIMILARITIES: LEVEL2
 Incorporating the context features
 Elementwise product model
 Similarities reweighted by the context feature
 Assumption: will be sensitive to the quality of the
context
T
 si , j ,c
T
1 Ii  Cc  I j 1 Ii  Cc  I j
s i , j ,c
 Pairwise model Ii 2
Ij
2
 Context dependent promotions/demotions for the
participating items
 Assumption: minor improvements to the basic
similarity I j I i T C c I j T C c
s i , j ,c Ii
T


12. CONTEXT-AWARE SIMILARITIES: LEVEL2
NORMALIZATION
 Normalization of context vector
 Only for cosine similarity
 Elementwise product model:
 Makes no difference in the ordering
 Recommendation in a given context to a given user
 Pairwise model
 Might affect results
 Controls the importance of item
promotions/demotions
I I
T
I C
T
I
T
C
i j i c j c
s i , j ,c
Ii Ij Ii 2
Ij
2 2 2
T T T
Ii Ij Ii Cc Ij Cc
s i , j ,c
Ii Ij Ii 2
Cc 2
Ij Cc
2 2 2 2

13. EXPERIMENTS - SETUP
 Four implicit dataset
 LastFM 1K – music
 TV1, TV2 – IPTV
 Grocery – online grocery shopping
 Context: seasonality
 Both with manually and automatically determined
time bands
 Evaluation: recommend similar items to the
users’ previous item
 Recall@20
 MAP@20
 Coverage@20

14. EXPERIMENTS - RESULTS
Improvement for Grocery Improvement for LastFM
35.00% 200.00% 183.91%
29.40%
30.00%
150.00%
25.00% 21.40%
20.00% Recall 100.00%
Recall
15.00% 12.90% Coverage
Coverage
10.00% 50.00%
6.11% 6.79% 6.14% 17.99%19.72%
5.14% 5.83%
3.91% 3.53% 5.27%
5.00%
0.87% 0.00%
0.00% -7.92% -1.62% -3.18%
-18.26% -21.96%
-50.00%
Improvement for TV1 Improvement for TV2
900.00% 797.51% 20.00%
800.00% 0.91% 0.79%
0.00% 0.59%
700.00%
-0.23% -3.23%
600.00% -20.00% -8.08% -8.13% -10.40%
500.00% Recall Recall
400.00% -40.00%
Coverage -37.24% Coverage
300.00%
-60.00%
200.00%
7.34% 46.40%
100.00% 0.69% 8.31%
1.77% 6.77% 5.89% 4.86% -80.00%
0.00% -82.32%
-3.53% -100.00%
-100.00%
 From left to right: L1 elementwise, L1 pairwise, L2 elementwise, L2 pairwise, L2 pairwise (norm)

15. EXPERIMENTS - CONCLUSION
 Context awareness generally helps
 Impromevent greatly depends on method and
context quality
 All but the elementwise level2 method:
 Minor improvements
 Tolerant of context quality
 Elementwise product level2:
 Possibly huge improvements
 Or huge decrease in recommendations
 Depends on the context/problem

16. (POSSIBLE) FUTURE WORK
 Experimentation with different contexts
 Different similarity between feature vectors
 Predetermination wether context is useful for
 User bias
 Item bias
 Reweighting
 Predetermination of context quality
 Different evaluation methods
 E.g. recommend to session

17. THANKS FOR THE
ATTENTION!
For more of my recommender systems related research visit my website:
http://www.hidasi.eu