Recommender Systems Ray Larson

1
Recommender Systems Ray Larson Warren
SackIS202 Information Organization and
RetrievalFall 2001UC Berkeley, SIMS lecture
author Warren Sack
2
Last Time

• Guest Lecture
• Abbe Don on Information Architecture
• (1) Guides
• (2) We Make Memories
• (3) don.com

3
Approach to User Interface Design
points of view
Information Architecture
politics of information
scenarios
Slide by Abbe Don
4
Issues

• Understand the relationships between information
  architecture, interaction design and media
  design.
• Examine how organizational structures and
  politics affect information architecture and
  thereby the overall design process and the final
  user interface.
• Re-enforce the importance of needs assessment,
  user scenarios, user requirements, and clear
  product definitions, business goals, etc.

Slide by Abbe Don
5
Guides Revised Characters

• 3 Content Characters in period dress
• Settler Woman
• Frontiersman
• Native American
• Always present in the interface gestures
  revealed level of interest
• Recommended all media types based on point of
  view algorithm with weighted terms
• Added point of view video stories for each
  character based on diaries and oral histories
• 1 System Character in contemporary dress
• Provided context sensitive help
• Recommended all media types based on emergent
  browsing pattern of the user

Slide by Abbe Don
6
Last Last Time

• Interfaces for Information Retrieval
• What is HCI?
• Interfaces for IR using the standard model of IR
• Interfaces for IR using new models of IR and/or
  different models of interaction

7
The standard interaction model for information
access

• (1) start with an information need
• (2) select a system and collections to search on
• (3) formulate a query
• (4) send the query to the system
• (5) receive the results
• (6) scan, evaluate, and interpret the results
• (7) stop, or
• (8) reformulate the query and go to step 4

8
HCI Interface questions using the standard model
of IR

• Where does a user start? Faced with a large set
  of collections, how can a user choose one to
  begin with?
• How will a user formulate a query?
• How will a user scan, evaluate, and interpret the
  results?
• How can a user reformulate a query?

9
Interface design Is it always the HCI way or the
highway?

• No, there are other ways to design interfaces,
  including using methods from
• Art
• Architecture
• Sociology
• Anthropology
• Narrative theory
• Geography

10
Information Access Is the standard IR model
always the model?

• No, other models have been proposed and explored
  including
• Berrypicking (Bates, 1989)
• Sensemaking (Russell et al., 1993)
• Orienteering (ODay and Jeffries, 1993)
• Intermediaries (Maglio and Barrett, 1996)
• Social Navigation (Dourish and Chalmers, 1994)
• Agents (e.g., Maes, 1992)
• And dont forget experiments like (Blair and
  Maron, 1985)

11
Relevance is not just topic, but also

• Recency
• Novelty
• Quality
• Availability
• Authority (Wang, ASIS 1997, 34, 162-173)
• Utility (Cooper, JASIS 24 87-100, 1973)

12
Today

• Recommender systems (see also collaborative
  filtering, social filtering, social navigation)
• Example systems Amazon.com, GroupLens, Referral
  Web, Phoaks, GroupLens, Fab
• How does it work? An Example Algorithm
• Generalizations of the recommender systems idea
  e.g., Social Navigation

13
The Basic Idea

• The basic idea of collaborative filtering is
  people recommending items to one another.
  Terveen et al., 1997

14
(No Transcript)
15
(No Transcript)
16
Amazon.com

• How might one visualize Amazons people who buy
  this book also buy feature?
• Examples from IS296a-2 Social Information Spaces
• www.sims.berkeley.edu/courses/is296a-2/f01/assignm
  ents.html
• Vivien Petras visualization www.sims.berkeley.ed
  u/vivienp/presentations/is296/ass1nonfiction.html
  

17
Social Networkscan beComputer-based Networks
(e.g., cross-indexed elements in a database)

• Cf., Barry Wellman, Computer Networks As Social
  Networks, www.sciencemag.org,
• Science, vol. 293,
• 14 September 2001

18
Resnick and Varian, 1997
19
Resnick and Varian, 1997
20
Resnick and Varian, 1997
21
GroupLensKonstan, Miller, Maltz,
Herlocker, Gordon, and Riedl
22
GroupLensKonstan, Miller, Maltz,
Herlocker, Gordon, and Riedl
23
GroupLensKonstan, Miller, Maltz,
Herlocker, Gordon, and Riedl

• Usenet news is a domain with extremely high
  predictive utility.
• High predictive utility implies that any
  accurate prediction
• system will add significant value.
• So then, why do we need a collaborative
  filtering system?
• In general, users do not agree on which articles
  are desirable.

24
Fab Balabanovi and Shoham
25
Fab Balabanovi and Shoham
26
Fab Balabanovi and Shoham
To create a hybrid content-based, collaborative
system, we Balabanovi and Shoham maintain user
profiles based on content analysis, and directly
compare these profiles to determine similar
users for collaborative recommendation. (p. 68)
27
Referral WebKautz, Selman and Shah
28
Referral WebKautz, Selman and Shah
29
Referral WebKautz, Selman and Shah

• Referral Web uses social networks extracted for
  public information
• Sources of the web.
• The current Referral Web system uses the
  co-occurrence
• of names in close proximity in any documents
  publicly
• available on the Web as evidence of social
  connection.
• Such sources include
• Links found on home pages
• Lists of co-authors in technical papers and
  citations of papers
• Exchanges between individuals recorded in news
  archives
• Organization charts (such as for university
  departments)

30
PHOAKSTerveen, Hill, Amento, McDonald, Creter
31
PHOAKSTerveen, Hill, Amento, McDonald, Creter

• PHOAKS works by automatically recognizing,
  tallying,
• and redistributing recommendations of Web
  resources
• mined from Usenet news messages.
• For a mention of a URL to count as a
  recommendation
• it must
• Not be posted to too many news groups
• Not be part of a posters signature or signature
  file
• Not be mentioned in a quotation from another
  message
• Contain word markers that indicate that it is
  being
• Recommended (and not advertised or announced).

32
SiteseerRucker and Polanco
Siteseer utilizes each users bookmarks as an
implicit declaration of interest in the
underlying content, and the users grouping
behavior (such as placement of subjects in
folders) as an indication of semantic coherency
or relevant groupings between subjects. Siteseer
looks at each users folders and bookmarks, and
measures the degree of overlap (such as common
URLs) of each folder with other peoples folders.
33
SiteseerRucker and Polanco
34
How do they work?An Example Algorithm

• Yezdezard Lashkari, Feature Guided Automated
  Collaborative Filtering, Masters Thesis, MIT
  Media Laboratory, 1995.
• Webhound
• Firefly

35
Webhound, Lashkari, 1995All automated
collaborative filtering algorithms use the
following steps to make a recommendation to a
user
36
Webhound, Lashkari, 1995
37
Webhound, Lashkari, 1995
38
Webhound, Lashkari, 1995
39
Webhound, Lashkari, 1995
40
Webhound, Lashkari, 1995
41
Webhound, Lashkari, 1995
42
From Items to PathsChalmers, Rodden Brodbeck,
1998
43
Social Navigation

• From Recommender Systems to the more general
  issue of Social Navigation (Dourish and Chalmers,
  1994)
• The ideas of social navigation build on a more
  general concept that interacting with computers
  can be seen as navigation in information space.
  Whereas traditional HCI sees the person
  outside of the information space, separate from
  it, trying to bridge the gulfs between themselves
  and information, this alternative view of HCI as
  navigation within the space sees people as
  inhabiting and moving thrugh their information
  space. Just as we use social methods to find our
  way through geographical spaces, so we are
  interested in how social methods can be used in
  information spaces.
• (Munro, Hook, Benyon, 1999).