Collaborative Filtering & Recommender Systems

Report
Collaborative Filtering and
Recommender Systems
Brian Lewis
INF 385Q
Knowledge Management Systems
November 10, 2005
Presentation Outline



Collaborative filtering and recommender systems defined
Novel example
Readings - overview & key concepts





Glance, Arregui & Dardenne (1997)
Konstan, Miller, et al. (1997)
Proctor & McKinlay (1997)
Conclusions
References
2
Collaborative Filtering defined

"Based on the premise that people looking for
information should be able to make use of what
others have already found and evaluated." (Maltz
& Ehrlich, 1995)

"Technique for dealing with overload in
information environments" (Procter & McKinlay,
1997)
3
Recommender systems defined

Systems that evaluate quality based on the
preferences of others with a similar point of view
4
Hobo symbols from http://www.slackaction.com/signroll.htm
5
Hobo symbols as RS?




Specific to a community
Implicit and explicit signs
Filtered through encoding
Cold-start problem?
6
Compare to today

Recommend

Don't recommend
7
Glance, Arregui & Dardenne (1997)

Knowledge Pump




Designed for use with an electronic repository
Document management and recommendation
Community-centered collaborative filtering
Characteristics


Social filtering
Content-based filtering
8
Glance, Arregui & Dardenne (1997)

User-item matrix of ratings
9
Konstan, Miller, et al. (1997)

GroupLens





Pilot study - Usenet news
Rating system
Integrate into an existing system/existing users
Use existing applications - open architecture
Characteristics




High volume / high turnover
High noise information resource
Sparse set of ratings
Predictive utility cost/benefit
10
Konstan, Miller, et al. (1997)

Predictive utility



Risk - costs of misses and
false positives
Benefit - values of hits and
correct rejections
Usenet has high predictive utility



High volume
Value of correct rejection is high
Risk of a miss is low
11
Konstan, Miller, et al. (1997)

Challenges

Ratings sparsity





"first-rater" problem
Partition articles into clusters
Capture implicit ratings
Filter bots
Performance challenges


System architecture
Composite users
12
Proctor & McKinlay (1997)


Social Affordances and Implicit Ratings
How implicit approaches might be improved




Sources of rating and recommendation data
Context of ratings and recommendations
Real and virtual groups
Privacy and accessibility
13
Proctor & McKinlay (1997)

Characteristics

Explicit ratings systems



Reader ratings based approach is expensive
How do you deal with trust issues?
Implicit ratings systems


Free to users
How do you capture context?
14
Proctor & McKinlay (1997)

Social Affordances


"…making the potential for social (inter)action visible."
How can activities be made visible? (explicitly)



Web bookmarks
Sharable annotations
How can activities be made visible? (implicitly)



Copy browsing behavior of experts (virtual groups)
Documents context in a group of documents (discourse
analysis)
Temporal coherence
15
Proctor & McKinlay (1997)

Extracting implicit ratings from web behavior


Virtual group proxies
Proxy cache analysis








Nominal rating
Frequency
Sequential accountability
Distributional accountability
Sources
Topical coherence
Temporal coherence
Privacy Issues
16
Conclusions

Many different issues



Context dependent



Diverse domains / communities
Diverse content needs
Nature of information
Predictive utility
Very creative solutions to draw from
17
References

Glance, N., Arregui, D., & Dardenne, M. (1997). Knowledge Pump:
Community-centered collaborative filtering. 5th DELOS workshop on
filtering and collaborative filtering, Budapest, Hungary.

Konstan, J., Miller, B., Maltz, D., Herlocker, J., Gordon, L. and Riedl, J.
(1997), Applying collaborative filtering to usenet news, Communication
of the ACM, 40(3), 77-87.

Maltz, D. and Ehrlick, K. (1995). Pointing the way: active collaborative
filtering. CHI '95, ACM Press.

Procter, R. and A. McKinley (1997). Social affordances and implicit
ratings for social filtering on the Web. DELOS workshop on
collaborative filtering, Budapest, Hungary.
18
Questions
Questions live here
19

similar documents