Collaborative Filtering & Recommender Systems

Download Report

Transcript Collaborative Filtering & Recommender Systems

KMS & Collaborative Filtering
• Why CF in KMS?
• CF is the first type of application to leverage
tacit knowledge
• People-centric view of data
• Preferences matter
- Implicit
- Explicit
• Are people just data points?
- Neo-Taylorism
- Efficiency over Quality for data collection
Community Centered CF
•
•
•
•
•
What is a community?
Helping people find new information
Mapping community (prefs?)
Rating Web pages
Recommended Web pages
- Measuring recommendation quantity?
- Measuring recommendation use
• Constant status
Community CF
• “Personal relationships are not necessary”
• What does this miss?
• If you knew about the user, would that help
with thte cold start problem?
• Advisors
• Ratings
- Population wide
- Advisors
- Weighted sum
• How would an organization use this?
Recommender Systems
• Broader term than CF, may not be explicitly
collaborating
• We get recommendations every day
• Types of recommendations
- Implicit
- Explicit
• Properties of recommendations
- Identity
- Experts
• Use of recommendations
- Aggregation from data
- Leveraging naturally occurring factors
Recommendation Issues
• How do you get people to cooperate?
• How good can the recommendations be?
- Find things you’d never find?
- Step savings, information navigation
• Volume of recommendations vs. number of
recommendable items?
• How accurate can the recommendations be?
- Initially
- Overall
- Over time
• What about changing interests?
Social (Filtering) Issues
•
•
•
•
Who controls the sharing?
Who controls the controls?
“Give to get” systems
Anonymity vs. Community
- Community of “friends”
- People as data points
• Free riders
• Logrolling and Over-rating
Social Filtering
• Very dependent on the society (types of users)
• Very dependent on the information (Web pages,
books, restaurants)
• PageRank becomes PersonRank?
- Matching your interests and then using it as a filter for both
other people and other items?
• Person, Document & Time
• Extracting Implicit ratings
- Reading time, # of accesses
- Own, rent, borrow
- Amount paid vs. avg cost, time to market
Information Filtering & IR
• How about filtering, without the
collaboration?
- Individual preferences
- Implicit and Explicit
• Text is analyzed
- Feature extraction
- Recall & precision measures
• Vector space identified
• Relevance Feedback
- Matched with user or rating
- Attributes are matched or added to queries
Two sides of the same coin?
•
•
•
•
Filtering is removing data, IR is finding data
Dynamic datasets
Profile-based - preferences
Repeated use of the system, long term
interests
• Precision & Recall of profiles, not info?
• Different needs & motivations
• Less interactive than (Web) IR?
Tapestry
• First system to apply these ideas among a
group of people
• Email & mailing lists - more email?
• Technical proof of concept architecture
• Queries - strengths & weaknesses
• Annotations are key, the extra information
adds to the document for searching and
knowing who has already read it
• How well do you have to know the other
readers/raters?
• Difficult to use queries, small document set
Active Collaborative Filtering
• Getting the community involved
• “people looking for information should be able to
make use of what others have already found and
evaluated” p1
• Allow people to send “pointers” to information to
colleagues.
- Link to the information with additional contextual
information/comments by sharer
- Keeps recommendation local
- Is supported by social norms (reputation, status)
• Lotus Notes
• How is it used?
- Sharing
- Information “digests”
- Looking at other’s bookshelves
Active CF Advantages
• Active Intent by the person who finds and evaluates a
document to purposefully share with particular
people
- Comments are specifically for your friends & co-workers,
perhaps specific people
- What about for your own re-retrieval or notes?
• Better as people find more documents, less of a
measure of popularity
• Leverages Gatekeeper behavior
• Uses “forms, views and databases” (like a list of
bookmarks) then adding macros to automate or filter
• Online lists, email & private or group databases
• Good for smaller groups and smaller document sets
Fab
•
•
•
•
Beyond “black box” content
Combining recommendations & content
Tastes in the past & future likes
Identifies “emerging interests”
-
•
Profiles of content analysis compared
-
•
•
•
•
Group awareness
Communication (feedback)
Users’ own profile can recommend
Relation between users can recommend
User profile = multiple interests
Content profile = static interest
Both may change
Items are continually presented to users
PHOAKS
•
•
•
•
Wider group of people (anyone?)
Usenet news (more text)
Link mining for Web resources
What counts as a recommendation?
- More than one mention?
- Positive & negative?
• Fair and balanced for a Community
• How do you rank resources?
- Weights
- Topics
Referral Web
• Leverage the informal network in an org
- Finding help & finding helper context
• Using a referral chain to get expert help
- Determining expertise by association
- Getting help by chain of association
• Creates referral network automatically
- How about asking?
- Neither way is always accurate
• Uses existing networks, not help building new ones
- Find a friend of a friend
- Can be applied to anything people in the group are interested
in
• Makes relationships visible
Social Affordance & Implicit
• How can you not use ratings to learn?
• Read wear, clicks, dwell time, chatter
• Not all resources are as identifiable
- Granular- Web pages
- Items - commercial products
• Web is a shared informaiton space without
much sharing
• How do incent people to contribute?
- Social norms
- Rewards
Context for Implicit Ratings
-
•
•
•
•
Who
When
What
How (discovery)
Web Browsing
RSS Reading
Blog posting
Newsgroup- listserv use
Contexts for Explicit Ratings
•
•
•
•
•
Movies
Books
(Junk) mail
eBay transactions
Other content
Active CF
•
•
•
•
•
Classic paper issues
Leveraging what others do
Finding what is already found?
Take advantage of universal publishing
How about filtering, without the
collaboration?
- Individual preferences
- Implicit and Explicit
• Is “wisdom” being accumulated?
Sharing References
•
•
•
•
Pointers
Packages of Information
General flexibility
Private and Public resources and ratings
Future Issues in Collaboration
• It may be more interesting to find a like mind
than a resource recommendation
- Social Networking
- Ad hoc group discussions
• Allowing users control over their profile of
interests
- Over time
- Privacy
- Difficult to capture interests
• Working with diverse content or user interests
• Visualization of recommendations & areas