RecSysTV 2016 - Time Series Effects For TV Recommendations

Recommendations for the television platform behave quite differently than the ones for video on demand platforms. A key difference is the dependency on television schedule and shows’ seasonality. In this talk, we will explore the time series effects that can be observed in the shows cyclicality, and more interestingly the effects of this cyclicality on users’ behavior. We’ll also go over some approaches to incorporate the time component onto different kinds of recommendations at Verizon IPTV

Transcript

1. TIME SERIES EFFECTS
   FOR TV RECOMMENDATIONS
   Diana Hu
   Verizon Labs
   RecSysTV, September 2016
   

   View Slide

2. HELLO!
   ○  Currently @ Verizon Labs
   ○  Formerly @ Intel Labs
   ○  Large Scale Machine Learning & Computer Vision
   ○  Scala & Spark since 2014
   You can find me at:
   @sdianahu
   

   View Slide

3. DATA @ IPTV VERIZON
   ○ Millions of subscribers
   ○ Live TV playbacks
   ○ DVR recordings
   ○ DVR playbacks
   ○ Time sensitive content by TV Schedule
   ○ Hundreds of channels: Long tail
   

   View Slide

4. TV EXPERIENCE
   

   View Slide

5. SOME HISTORY
   first second
   http://www.brucesallan.com/2012/10/13/evolution-technology-television-back-day/
   http://avdesigns.com/blog/design-your-own-at-home-sports-bar-with-a-multiscreen-tv-system/
   

   View Slide

6. CREATURES OF HABIT
   

   View Slide

7. CYCLICAL PATTERNS
   ○ Prime time
   ○ Weekend binge
   ○ Sunday night football
   ○ Basketball season
   ○ Effect of Sitcoms
   ○ Cartoons
   ○ News
   

   View Slide

8. TIME OF DAY VOLUME
   

   View Slide

9. Time Segment Day Segment
   6 MONTH TIME SERIES
   

   View Slide

10. GENRE EFFECTS
    

    View Slide

11. MAKING SENSE OF TIME
    Besides pretty visualizations…
    

    View Slide

12. NAÏVE APPROACH
    

    View Slide

13. FEATURE ENGINEERING
    ○ Day parting
    ○ Weekday parting
    ○ Volatility spikes
    ○ Intraday volatility
    ○ Rank across time segments for:
    □ Channels
    □ Genres
    □ Programs
    

    View Slide

14. UNLOCKING PREDICTIONS
    

    View Slide

15. TIME AS A CONTEXT
    Enhancing recommendations
    

    View Slide

16. SOME APPROACHES
    

    View Slide

17. USE CASES
    ○ Predicting viewership
    ○ Personalization
    ○ Re-Ranking
    ○ Adapting to TV content “shelf life”
    ○ Seasonality
    ○ Trends
    

    View Slide

18. ACKNOWLEDGEMENTS
    ○ Luis M. Sanchez
    ○ Humberto Corona
    

    View Slide

19. THANKS!
    Any questions?
    You can find me at:
    @sdianahu
    diana.hu@verizon.com
    Also, we are hiring!
    

    View Slide