In figures: IBFF - Incremental Item Based w/ Fading Factors UBSW 78 - Non-incremental User Based w/ Sliding Window (length: 78 sessions) UBFF 0.99 - Incremental User Based w/ Fading Factors UBFF 1 - Incremental User Based without Fading

1. Incremental Item-based
Collaborative Filtering
João Marques da Silva
Palco Workshop - May 13, 2009

2. Item Similarity
Clã Xutos Gift DaWeasel
Ana 1 1 0 0
Miguel 1 1 1 0
Ivo 0 1 0 1
Paula 0 0 1 0
Joana 1 0 0 0
Take columns as vectors:
v =1,1,0 ,0 ,1
Clã and v =0,1 ,0 ,1,0
Gift
Similarity between Clã and Gift (cosine measure):
v . v
Clã Gift
sim Clã , Gift =cos  v , v =
Clã Gift ≃0.16
∥v ∥∗∥v ∥
Clã Gift
2

3. Similarity Matrix
S matrix
MxM, with M = nº of items
Clã Xutos Gift DaWeasel
Clã 1
Xutos ... 1
Gift 0.16 ... 1
DaWeasel 0 ... 0 1
How do we keep S up-to-date?
 Rebuild S at each new session:
O(m2n) for m items and n users.
 Incrementally update S with session data:
O(km) for k items in session.
3

4. Algorithm
Cosine measure for binary ratings:
#  I ∩J 
cos  ,  =
i j I , J are the sets of users that rated items i , j
 # I × # J
A cache matrix Int stores #(I ∩ J) for all item pairs (i,j):
Inti,j = #(I ∩ J)
Inti,i = #I
For each new session:
 Increment Inti,j by 1 for each item pair (i,j) in session

For each item in session update corresponding row/col in S:
Int i , .
S i ,.=
 Int i , i ×  Int . , . 4

5. Forgetting
 Usage and content change!
 News content quickly becomes obsolete
 Music/Movies/Books - popularity is often volatile
 How can CF adapt to change?
 Forget older data
 Two methods: sliding windows and fading
factors
5

6. Forgetting: Sliding Windows
Sliding Windows
window
Session weight
length
Data in window Current Session
Session index
Good for non-incremental:
Rebuild S with data in window. 6

7. Forgetting: Fading Factors
Fading Factors
Session weight
Current Session
Session index
Good for incremental. Before updating S:
S = αS , 0 < α < 1
α=1 is the non-fading factor
7

8. Implementation
 Implementation in R
 Code available from previous work (C. Miranda)
 Adapt algorithms to use forgetting mechanisms
 Improvements: sparse matrix handling
 Limitations with R: speed
8

9. Experiments
 Aims
 Forgetting – is it useful?
 Sliding windows vs fading factors
 Item-based better than user-based?
 Evaluation method
 All-but-one protocol (training, test and hidden sets)
 Artificial disturbances in datasets
 Accuracy: precision/recall (binary ratings)
9

10. Experiments: datasets
2 sequential datasets:
Dataset Origin # sessions # items
PALCO* Palco principal 725 1285
ART* Artificial 1500 4
PALCO: Listened tracks in Palcoprincipal
ART: dataset with abrupt change
{a,b,c} → {a,b,d} at session 500
10

11. Results (so far)
 Matrix update time
 Update time < Rebuild time
 Item-based better for #users > #items
 PALCO: user-based performs better
 Non-incremental good with small windows
 Recommendation time
 Item-based is faster
 Recovery from drifts

ART: α<1 recovers faster than α=1 (as expected)

PALCO: α=1 still better even with 90% drift! 11

12. Accuracy IBFF w/ ART
12

13. Accuracy UBSW, UBFF w/ PALCO
13

14. Issues
 Forgetting
 Not good for PALCO dataset?
 Good with ART dataset, but not realistic
 Other datasets (ex: news)?
 Long term effects → larger scale experiments
 Better hardware - on the way
 Other implementations (Java, C, SQL…)
 Palcoprincipal
 More items than users!
 Item-based possibly better for artist recommendations.
14

15. Thank you!
15