1 g m dg m d i p s ii p s i recommending tv programs on the web between content based retrieval and...
TRANSCRIPT
1
G M D
I P S I
Recommending TV programson the WebBetween content based retrievaland social filtering
Patrick BaudischGMD-IPSI
March 3rd 98
March 8th 98
Credits
Thanks for the award (its almost done...)
ContentsPart 1
About the project Requirements An evolving system Personalization
Part 2 Recommendation and cooperative
aspects
Feedback & Conclusions
About the project A cooperation between GMD-IPSI &
GMD: German national reserach institute for information technology
TV-TODAY German printed TV program guide They sell 1,400,000 copies per two weeks Where are printed guides going when digital
TV and video on demand emerge?
About the project
Goal: Help users in creating their personal TV schedule
None of the German Web-based TV program system gives more recommendation than their printed counterpart
Be more than a prototype, reach thousands of users
Learn German now!
Design criteriaTV programs vs. books and movies TV programs are a stream rather than a
database=> We do not have much time to collect data for recommendations
TV programs are experienced as having a lower value => Require only low user effort
Users have experiences and therefore expectations from printed TV program guides (e.g. TV-TODAY)=> Start with what users expect
Design criteria
System must be easy to learn (WWW) => Do what people expect
Be spectacular (TV-TODAY) => Do what people don´t expect
An evolving system
1
2
3
guestsmembers
FamiliarityBehave like printedTV program guides
RetrievalQuery/Browsing
Personalization & FilteringAdjust permanent settingsProfile, Push service
First time user interface (guest mode)
What’ on tonight
We ran user tests: 40% of first time users plan only for today
Press Start
1
Genre Visualization
Table cells color-coded
List items have colored field
Hue = Genre
e.g. {sports=green, movie=red, ...}
Recommendation Visualization
Color intensity = relevance the darker the more recommended less recommended programs fade to background
color
What means “recommended”? (later slide)
Retrieval: Adjust four parameters
Date interval
Time interval
Channels (predefined set)
Genre
Press Start
2
Genre hierarchy
A Genre is the set of programs that match a descriptor
Deeper genres are more specific
Guides users
Less universal than boolean search
Create account / login
To personalize users need an account
Store user data on server side
Use this data matching users making
recommendations
--
Member user interface
Personalization
Personalize three of four parameters favorite times favorite channels favorite genre (There are no
favorite dates)
3
Personalize times
Click yellow buttons into hour fields
Draw whole rectangles at once (Mac Paint)
Personalize channels
(Applet Demo)
Select the German regional stations that you can receive
Personalize genres
Check favorite genres
Use folder with favorite genres like bookmarks
Click “all favorite genres” to load all at once
Personal schedule (“grocery list”) Select programs, print it out, take it home
RecommendationHow the colors are generated?
Is Social filtering applicable?
Chez. P Wally’s Beef-O Veggio Pizza.H McD’s
Joe D A B DJohn A F D FBrad C D BSue C C CBen A A AEllen F A FEthan D A A
(Diagram by Joe Konstan)
Applicability of the Ringo approach Correlate users by the programs in the grocery
list?
“In/Not in” info from the grocery list is much less informative than 7 ratings scale=> results of correlating people is rather poor
We don´t have unlimited time, only one week.
The database is not stable User A just returned from a 2 week vacation User B is a newbee
Correlate on a standard set of items means extra effort (amazon recommendation center)
correlation?
Applicability of the Grouplens approachGroupLens: Press 1,2, ..,7 to rate and go
to the next article
Joseph A. Konstan says: These ratings require high cognitive costs
=> Rating effort might be too much
Four types of recommendation
A Recommendationsby TV-TODAY
“Size of the audience”
Personal genre profile
Opinion leaders
B
C
D
The editors of TV-TODAY provide ratings for all movies of the day (60 of 1000 programs)
Ratings , , ,
You agree or you don`t
Recommendations by TV-TODAYA
Size of the audience
Use programs in “Grocery list” as a recommendation for other users
We count how often a program occurs in users “grocery lists”
The more the better the rating
Works for all programs not only movies
Will lead your attention to events like“Tour de France”
Not personalized => might not fit your personal interests
B
Initialization of “Size of the audience” Everyday one day is added, one removed
When a program is inserted into the system it is not in anyone´s “grocery list”
=> Initialize ratings from the genre or series
Remove initialization during the week and replace with the real recommendations
“Grocery list” means: “I want to see that”. It does not mean “I like that” (how could I know before I`ve seen it) (and afterwards nobody cares)
Anyway: It works!
Personal genre profile
Describe favorite genres in more detail
Based on public recommendation, but users define offsets to adapt ratings to personal needs
Andrea´s personal TV interests She is interested in sports, especially in
basketball, where she does not want to miss a single program.
She wants to be up-to-date about current information without spending too much time on it.
Finally, for recreation, she wants to include some good action movies.
C
Profile
Form based Interface
Define how many programs of this genre to get
Define how personally important these are
Form based Interface
Define for all favorite genres
Initialization:Small is important(Law of Zipf)
Graphical user interface
Grey = cropped
Yellow = selected
Red = important
Drag boxes around
Box sizes reflect number of programs available per week
Evaluation: Number of subjects
No deformation deformation
No tutorial 5 5
tutorial 5 5
(We just got started, ...)
Form based interface:
Graphical Profile Editor:
10 subjects
Comparison of the two interfaces The graphical interface is much more difficult
to learn than the form-based interface
The graphical interface provides more utility and is easier to use than the form-based interface precision graphical overview
=> Provide a form-based interface for first-time users and a graphical interface for frequent users
Learnability: There seems to be a lack of methaphors (Where is Don Gentner?)
Opinion leaders
Allow more individual users to generate recommendation (not only TV-TODAYs editors)
Loren/Phoaks: Not everybody wants to give recommendation, but some do
Take “Grocery lists” of an individual user as your personal source for recommendation(instead of summing all up)
D
Opinion leaders
Opinion leaders are represented as a folder containing their “grocery list”
An opinion leader behaves exactly like a genre
Users can have their favorite opinion leaders
Who benefits...?
Being an opinion leader means no extra work
“Don´t you want to become an opinion leader?”
But: Opinion leaders loose part of their privacy
Let´s reward them for that: Give them program data one week in advance
=> that helps initializing “size of the audience” A free subscription to the printed guide Tell them that it is “cool” to be one
Survival of the fittest: If a new opinion leader applies drop the one with the fewest subscribers
... and who loses?
TV-TODAY editors can be opinion leaders
TV-TODAY didn´t like the idea too much :)
Evaluation of the overall system so farOur group + TV-TODAY people (about 20
users)
Beta test at GMD IPSI with about 30 users
User tests with 10 users for 40 minutes each
Feedback
Orientation is easy, but undo is missing
For some users the system is still too complex (opening folders, buttons to small for elder users)
People liked the „grocery list“That´s good for our recommendation system
Overall it is useful and easy to use
High fun-factor!
„When will you go online?“
Future work
Go online! April 98
Where else can we apply the described techniques: Usenet news, web pages, ...
Be more proactive: Push service, email notification of very important programs
Scott Robertson (digital libraries): Soft pushes
Future work: Cooperative stuff
We do have a “Find similar users” component(based on favorite genres and genre profiles)
Allow users to exchange their profiles
Become an opinion leader for individual users (friends, community)
Recommend genres and opinion leadersThis allows managing a greater number of them
Have specific opinion leaders One that just recommends action movies, ... Keep them inside the genre structure
Profile creation is NOT JUST iterative
Three paths lead to Profiles
1. Creation
2. Outerrefinement cycle
3. Innerrefinement cycle
Producers ofDocuments
Distributors ofDocuments
Distribution andRepresentation
DocumentSurrogates
RegularInformation Interest
Users/Groups withLong-term goals
Representation
Comparisonor Filtering
Modification
Use and/orEvaluation
RetrievedDocuments
Profiles
Conclusions
Traditionally a broadcast medium TV makers broadcast, viewers watch Editors write, readers read Interactive and collaborative concepts are new
here
The system contains a lot of functions, some of which are more complex Users have months to discover all these
functions Until then retrieval is just fine Many users will never push it that far That´s ok!
The END
What do you think?