3. About Products
Founded – June 13 2012 CAS / DRM
Fully owned by France
Telecom Personalized Content
Discovery platform that
350 employees
recommends
Offices: France, US, Hong- the right content
Kong, Israel
100+ customers worldwide TV Everywhere
Flexible middleware
platform delivering a full
array of IPTV & OTT
services – including live TV,
VOD and PVR,
across multiple screens
5. 3 main ways to discover content
Search Recommendations Exploration
I know what The service knows I’d like to explore
I am looking for what I am looking for with my personal guide
6. 3 main ways to discover content
Search Recommendations Exploration
Auto-complete Collaborative Personal zone
filtering
Auto-suggest Trends
NLP - semantic
Social-aided Games
Social
Personalized Gossip
Popular/Trending
search Deals
Lists (experts,
Smart filters operator, users) Friends
10. How it generally Works ?
Filtering
Engines
Advanced
semantic
Usage Discovery
data Collaborative Manager
filtering
Event Registration
User
Profiles
Middleware
(CMS, CRM)
14. 75% of what people watch is from
some sort of recommendation.
Netflix blog April 2012
35% of Amazon sales are due to
recommendations
Venturebeat - 2006
Nice numbers but no single way to measure success…
18. Lesson 1 –
have a dedicated group of real users for tests
15%
15% satisfaction gain in 1 week by
adding bots and tuning thresholds
and filtering in collaborative-filtering
19. Lesson 2 –
avoiding the rotten
apple is more important
than getting the perfect
ones
We’ve seen dramatic jumps in
satisfaction when pruning bad
results
- Time slices
- Thresholds
- Exclude rules
- Adequate “system warming”
- External guides (e.g. popularity)
20. Lesson 3 –
Multiple engines (i.e. engines blend ) help overcoming the
filter- bubble effect
Overall satisfaction Rating
Semantic (1 – 10) ( 1- 5)
Single
Network 6.01 2.72
Dual
Social network 8.00 3.14
Change 33% ~15%
Research: The Effect of Dual Networks
Prof. J. Goldenberg, Dr. G Ostreicher and S Reichman - Feb. 2011
21. Lesson 4 –
Collaborative filtering is popularity biased,
users prefer novelty
The highest correlation found in early experiments is
between novelty and purchase decision (0.55)
Which one to recommend if
score is the same ?
Avatar ?
The Pianist ?
Research: Negotiation agents n=50 x 2 (US, India)
Prof. S. Kraus and Dr. A. Hasidim - Apr. 2012 (not yet published)
22. Lesson 5 – OMG it speaks French
Fear Peur
abhorrence, agitation, angst, anxiety, aversi
on,
awe, chickenheartedness, cold feet,cold sw
eat, concern, consternation, cowardice,cree
ps, despair, discomposure, dismay,disquiet
panique, phobie, frayeur, appréhension, frisson,
ude, distress, doubt, dread,faintheartednes
épouvante, crainte, alarme, émotion, affolement
s, foreboding, fright, funk, horror,jitters, mis
giving, nightmare, panic, phobia,presentime
nt, qualm, recreancy, reverence,revulsion, s
care, suspicion, terror, timidity,trembling, tre
mor, trepidation, unease, uneasiness, worry
Change in words statistics, change of sources
, change in amount of reviews, change of
vocabulary, correlation to English data is not
always clear
23. Lesson 6 – Cold start could get really cold….
System bootstrap User cold start Content cold start
Add users Non-personal Non-personal
Add values Implicit evidence Implicit evidence
Hybrid methods Questionnaire Aggregated data
25. Lesson 8 – Privacy matters
Explicit / Implicit
- Channels
- Genres
Profile visibility - Actors
- Devices
- Subscriptions
- Black / White
lists
Still relevant
- Semantic without
history
Enough value - Popularity
when not opted-in - Trends
- Lists
- General CF
26. Other points to consider
Provide reason
Recommended because 3 of
your friends liked it
Channel zapping
VOD order
Program recording
VOD content end
Setting a reminder
VOD rating
Launches of
Add to wish list
COMPASS Collect valuable indirect evidence
Show movie
Explicit input
preview
Exclude content
Channel zapping
Search terms
Morning
Evening
Noon
Night
Handle “time” with care
27. It’s an on going process
It’s getting better every month
It’s a lot of fun
metadata and responsiveness to different business models, yields almost unimaginable reservoir from which substantial business value can be derived.As a start, an operator can review different ways to reach for content and how these ways can be optimized towards different audiences. For example, in Orca’s Content Discovery system - COMPASS – 4 main approaches are provided to discover content, recommendations, search, social TV and exploration. By utilizing these approaches, it’s now easier to understand how your audience find and consume content. Furthermore, these discovery methods are further supported by different discovery engines: Collaborative Filtering, Semantic Recommendations, Users’ Profiles, Social TV, Operator Promotions, Popularity Based and External Critics, which provide insights into which engine serves which audience.
metadata and responsiveness to different business models, yields almost unimaginable reservoir from which substantial business value can be derived.As a start, an operator can review different ways to reach for content and how these ways can be optimized towards different audiences. For example, in Orca’s Content Discovery system - COMPASS – 4 main approaches are provided to discover content, recommendations, search, social TV and exploration. By utilizing these approaches, it’s now easier to understand how your audience find and consume content. Furthermore, these discovery methods are further supported by different discovery engines: Collaborative Filtering, Semantic Recommendations, Users’ Profiles, Social TV, Operator Promotions, Popularity Based and External Critics, which provide insights into which engine serves which audience.
Research done with Prof. J. Goldenberg (Hebrew University- Jerusalem) and Dr. G Ostreicher and S Reichman from Tel Aviv university
Research done with Prof. J. Goldenberg (Hebrew University- Jerusalem) and Dr. G Ostreicher and S Reichman from Tel Aviv university
Research done with Prof. J. Goldenberg (Hebrew University- Jerusalem) and Dr. G Ostreicher and S Reichman from Tel Aviv university
Research done with Prof. J. Goldenberg (Hebrew University- Jerusalem) and Dr. G Ostreicher and S Reichman from Tel Aviv university
Research done with Prof. J. Goldenberg (Hebrew University- Jerusalem) and Dr. G Ostreicher and S Reichman from Tel Aviv university
Research done with Prof. J. Goldenberg (Hebrew University- Jerusalem) and Dr. G Ostreicher and S Reichman from Tel Aviv university
Research done with Prof. J. Goldenberg (Hebrew University- Jerusalem) and Dr. G Ostreicher and S Reichman from Tel Aviv university