TechLens:

1
TechLens  Exploring the Use of Recommenders to
Support Users of Digital Libraries

• Joseph A. Konstan, Nishikant Kapoor, Sean M.
  McNee, John T. Butler
• GroupLens Research Project and University
  Libraries
• University of Minnesota

2
Introduction

• Challenges and Opportunities
• large digital collections of uneven quality and
  scope
• continuing trend towards out-of-library usage of
  library collections
• extensive collections of metadata
• citations and other linkage data (published and
  personally collected)
• venue data
• expectations of personal service
• increased prevalence of personalization

3
Recommenders

• Tools to help identify worthwhile stuff
• Filtering interfaces
• E-mail filters, clipping services
• Schedulable current awareness searches
• Recommendation interfaces
• Suggestion lists, top-n, offers and promotions
• Prediction interfaces
• Evaluate candidates, predicted ratings

4
Amazon.com
5
Wine.com Seeking
6
Cdnow album advisor
7
CDNow Album advisor recommendations
8
Classic CF
C.F. Engine
Ratings
Correlations
9
Submit Ratings
ratings
C.F. Engine
Ratings
Correlations
10
Store Ratings
C.F. Engine
ratings
Ratings
Correlations
11
Compute
C.F. Engine
pairwise corr.
Ratings
Correlations
12
Request Recommendations
C.F. Engine
request
Ratings
Correlations
13
Identify Neighbors
C.F. Engine
find good
Ratings
Correlations
Neighborhood
14
Select Items Predict Ratings
C.F. Engine
predictions
recommendations
Ratings
Correlations
Neighborhood
15
Understanding the Computation
16
Understanding the Computation
17
Understanding the Computation
18
Understanding the Computation
19
Understanding the Computation
20
Understanding the Computation

21
Understanding the Computation

22
First Steps

• Established that citation web data can be used to
  effectively rate/recommend research papers
• Developed and evaluated a demonstration
  recommender to recommend additional citations for
  an existing paper (using its references)
• original demo used CiteSeer
• this version uses ACM digital library

23
DL Recs
C.F. Engine
Ratings
Correlations
24
DL Recs
Votes
C.F. Engine
Ratings
Correlations
25
DL Recs
Votes
C.F. Engine
Ratings
Correlations
26
DL Recs
Votes
C.F. Engine
Ratings
Correlations
27
DL Recs
Votes
C.F. Engine
Request
Ratings
Correlations
28
DL Recs
Votes
C.F. Engine
Request
Ratings
Correlations
29
DL Recs
Recommendations
Votes
C.F. Engine
Request
Ratings
Correlations
30
Demonstration 1

• Steps
• Select user
• Select paper
• Select algorithm
• See recommendations

31
What We Found

• Results published in McNee et al. (CSCW 2002)
• Yes, we can make recommendations this way!
• offline analysis showed that best algorithms
  could find half of recommendable withheld
  references in top 10, ¾ in top 40 recs
• online experiments showed best algorithms gave
  recommendations more than half of which were
  relevant, and more than half of which were novel
• Users like it!
• more than half of users felt useful (1/4 to 1/3
  said not)
• 1-2 good recs out of 5 seemed sufficient for use
• Different algorithms have different uses
• Further exploration in Torres et al. (JCDL 2004)

32
Phase II

• Shifted our focus to ACM Digital Library
• Greater exploration of user tasks
• awareness services
• keeping track of a community
• More automation
• find own bibliography from citations
• find collaborators
• Thinking about researchers desktop

33
Demonstration 2

• Steps
• identify self
• see automated collections of citations and
  collaborators
• show how to use collections for recommendation

34
Moving Forward

• Collaboration
• Computer Scientists (HCI, recommenders)
• Librarians (field work, domain expertise,
  real-life service deployment)
• Research methods
• Offline data gathering and feasibility studies
• Online pilots and controlled experiments
• Online field studies (including random-assignment
  studies)

35
Whats Next?

• Short-Term Efforts
• Task-specific recommendation
• Understanding personal bibliographies
• Privacy issues
• Longer-Term Efforts
• Toolkits to support librarians and other power
  users
• Exploring the shape of disciplines
• Rights issues

36
Task-Specific Recommendations

• Many different user needs
• awareness in area of expertise
• find specific work in area of expertise
• explore peripheral or new area
• find people with relevant expertise
• reviewers, program committees, collaborators
• reading list for students, newcomers
• individuals or groups
• Different algorithms fulfill different needs

37
(No Transcript)
38
Personal Bibliographies

• Working with RefWorks to explore bibliographies
  maintained by library users
• how resolvable is personally-managed
  bibliographic data?
• where does data come from (import/type) and is
  there sufficient quality control?
• depth and span of bibliographies
• suitability for recommenders

39
Privacy Issues

• Anything involving personal bibliographies,
  library usage is extremely sensitive
• what can we do with minimal personal data (e.g.,
  explicit queries)?
• can we identify particularly sensitive cases?
• can we de-personalize data for collaborative
  applications?
• for what benefits will users give informed
  consent to use private data?
• feasibility/efficacy of ratings in library domain

40
The Toolkit

• What would it take to support complex requests?
• Help me assemble a collection of the 20 papers in
  molecular biology that have been most influential
  in other sciences
• Help me assemble a committee of leading humanists
  who together span a collection of fields and have
  collaborated with most of the leaders of those
  fields
• A new dimension of service for expert librarian

41
Describe a Discipline

• Can we build automated tools to
• identify the most important conferences and
  journals for a field?
• identify the most important papers?
• seminal work from other fields
• seminal work that established this field
• new work of particular influence
• identify trends in topic?
• identify hubs of activity?

42
Rights Issues

• Not our core expertise, but
• rights issues are critical, particularly
• use of metadata, including abstracts
• possible future use of reviews
• also important to understand and educate authors
  on future uses of their work
• everything from rating systems to plagiarism
  detection

43
Discussion

• Issues of your choice, or
• privacy issues are these a show-stopper?
• will these tools change the nature of
  scholarship? is it already changing?
• can I cite each member of the program committee?
• what will it take to demonstrate the value of
  such tools?
• pragmatic issues of interoperability

44
Our Thanks

• GroupLens Research Group
• U of M Libraries
• NEC Research, ACM, RefWorks
• NSF Grants DGE 95-54517, IIS 96-13960, IIS
  97-34442, IIS 99-78717, and IIS 01-02229 (and we
  hope more to come!)
• All the colleagues whove given us feedback along
  the way
• Our research subjects/users

45
TechLens  Exploring the Use of Recommenders to
Support Users of Digital Libraries

• Joseph A. Konstan, Nishikant Kapoor, Sean M.
  McNee, John T. Butler
• GroupLens Research Project and University
  Libraries
• University of Minnesota