1. Kamal Ali – TiVo, Yahoo Wijnand van Stam, TiVo
TiVo Suggestions: Predicting Viewer Affinity Using Collaborative Filtering
Kamal Ali – TiVo, Yahoo
Wijnand van Stam, TiVo

2. Outline What is “TiVo” ? Why Suggestions?
Collaborative filtering background
TiVo collaborative filtering data cycle
Server-side learning
Previous Work
Contributions

3. Contributions Large fielded system Distributed architecture
Large number (3M) of users
Long-lived interaction w user: >90 t/user
10^8 ratings over 300K shows
Very large in user-hours
Distributed architecture
Server: Throttle-able
Clients do bulk of work
Privacy-preservation
Privacy and distributed goals aligned
No persistent memory of user on server

4. What is “TiVo” ? TiVo = set-top TV box + program-guide service
Pause & rewind live TV
Linux OS
Viewers can rate shows
Suggestions
Q4 1999

5. Why Suggestions? Connect users to shows they’ll like
Predict degree to which viewer will like TV show
Produces ranked list of upcoming shows
Records shows if disk space is available

6. Filtering Background Recommendation Systems
Content-based: use “intrinsic” features such as genre,
cast, director, writers, age, channel-type,…
Combined,
Cascaded
Collaborative filtering:
use other people’s ratings

7. Content isn’t sufficient
Genres are few
Text length is small

8. Data cycle Rated shows in sorted order Thumbs Profile on TiVo
1: Collecting
Feedback:
Thumbs up/dn
Recorded
Rated shows in
sorted order
Thumbs
Profile on TiVo
Client box
5. Use correlations and
Thumbs profile to rate shows
2. TiVo calls server
uploads entire
anonymized profile
Correlation pairs on
client
No persistent state (even in randomized form) on server for each viewer
Random ID generated
for profile and
stored on server
Correlation pairs
<s1,s2,r> on
server
4. Download pairs during
some client-initiated calls
3. Server-
side learning

9. Collaborative Filtering Model
k Nearest Neighbor over other rated correlated shows
Use Pair-wise Pearson correlation
Adjusted correlation for low support
Use weighted linear combination

10. 1. Collecting Feedback Explicit: Implicit: Thumbs up, down: -3 ... +3
User-initiated recording thumbs

11. 2. Privacy and Data Upload
TiVo calls server daily
Entire profile uploaded and given temp id
Server deletes old profiles: sliding window

12. 3.1 Server-side scaling 300,000 unique shows /week
10^11 pairs of shows
3M users
Average of 90 thumbs / user: > 10^8 thumbs (ratings)
Ratings are sparse in the pair space
Don’t need to predict for very unpopular pairs

13. 3.2 Server-side Learning
Building pair-wise item/item correlations on server
Use simple Pearson pair-wise correlation
7 ratings levels per show [-3 … +3]
Only need to maintain 7 * 7 array of counts per pair
Efficient: CPU, memory
Compute r-to-z transform to computer confidence interval
Support-penalized degree of correlation: lower bound of confidence-interval
Distinguishes r = 0.8 for S=10 versus S=1000

14. 3.3 Throttled Server-side Architecture
Log Collector 1
Boxes K
Log Collector m
Boxes 100K(m-1) K m
By-series Counter 1
Series 0..30K
By-series Counter n
Series 30k(n-1)..30kn
1: By-series-pair Counter and Correlations Calc.
P: By-series-pair Counter and Correlations Calc.
Transmit correlation pairs to TiVo Clients

15. 3.4 Server-side throttling
min_single (150)
min_pair (100)
Throttle-able:
More HW available
Increasing TiVo population
Go deeper into distribution tail

16. Details Pearson r Weighted average r-to-z transform (Fisher)
Standard: Lower bound of confidence interval:

17. 4. Download to clients 28K pairs sent to client (320kb)
Correl. between old shows don’t change fast
New Shows: want to do it faster

18. 5. Client-side processing
Ratings must not cause video glitching!
2am: TiVo re-rates all shows
Collab: k-nearest neighbor
Content-based: Naïve Bayes

19. Previous Work User-user or item-item - Sarwar et al Form of model
k-nearest neighbor,
Bayes nets (Breese et al.),
Factor Analysis (Canny)
Similarity/distance function
Pearson (subsumes cosine)
TFIDF corrections (Salton et al.)
User amplification
Combination functions: k-NN, Bayes nets..
Evaluation Criteria: MAE, Spearman rank correl.

20. Contributions Large fielded system Distributed architecture
Large number (3M) of users
Long-lived interaction w user: >90 t/user
10^8 ratings over 300K shows
Very large in user-hours
Distributed architecture
Server: Throttle-able
Clients do actual suggestion calculations
Privacy-preservation
Privacy and distributed goals aligned
No persistent memory of user on server