Stereotype Modeling for Problem-Solving Performance Predictions in MOOCs and ...
Social information Access2012
1. Social Information Access
Peter Brusilovsky
with Rosta Farzan, Jaewook Ahn, Sharon
Hsiao, Denis Parra, Michael Yudelson,
Chirayu Wongchokprasitti, Sherry Sahebi
School of Information Sciences
University of Pittsburgh
http://www.sis.pitt.edu/~peterb
3. The New Web: the Web of People
http://www.veryweb.it/?page_id=27
4. Web 2.0: Fast Start, Broad Spread
• Term was introduced following the first O'Reilly
Media Web 2.0 conference in 2004
• By September 2005, a Google search for Web 2.0
returned more than 9.5 million results
• In 2012 similar search returned over 2 billion
results
http://datamining.typepad.com/data_mining/2005/12/the_rise_and_ri.html
7. Key Elements
• The Users’ Web • User as a first-class
• Collective participant,
Intelligence: contributor, author
Wisdom of Crowds
• The power of the
user
• Applications
powered by user
community
• Stigmergy
http://www.masternewmedia.org/news/2006/12/01/social_bookmarking_services_and_tools.ht
14. The Other Side of the Social Web
User content
User interaction
Which wisdom of crowds?
15. Social Information Access
Methods for organizing users’
past interaction with an
information system (known as
explicit and implicit feedback),
in order to provide better access
to information to the future
users of the system
16. Critical Questions
• What kind of past interaction to take into
account?
• How to process it to produce “wisdom of
crowds” ?
• In which context to reveal it to end users?
• How to make wisdom of crowds useful in
this context?
17. Social Information Access: Contexts
Social Navigation
– Social support of user browsing
Social Recommendation (Collaborative Filtering)
– Proactive information access
Social Search
– Social support of search
Social Visualization
– Social support for visualization-based access to information
Social Bookmarking
– Access to bookmarked/shared information facilitated with tags
18. Social Navigation: The Motivation
• Natural tendency of people to follow
each other
Making use of “direct” and “indirect
cues about the activities of others
Following trails
Footsteps in sand or snow
Worn-out carpet
Using dogears and annotations
Giving direction or guidance
• Navigation driven by the actions
from one or more “advice
providers”
19. The Lost Interaction History
What is the difference between walking in a
real world and browsing the Web?
– Footprints
– Worn-out carpet
– People presence
What is the difference between buying and
borrowing a book?
– Notes in the margins
– Highlights & underlines
– Dog-eared pages
– Opens more easily to more used places
20. Edit Wear and Read Wear (1992)
The pioneer idea of
asynchronous indirect social
navigation
Developed for collaborating
writing and editing
Indicated read/edited
places in a large document
21. Footprints (1997)
Wexelblat & Maes, 1997
Allowing users to create
history-rich objects
Providing history-rich
navigation in complex
information space
Showing what
percentage of users
have followed each link
22. SN in Information Space:The History
History-enriched environments
– Edit Wear and Read Wear (1992)
– Social navigation systems
• Footprints, Juggler, Kalas
Collaborative filtering
– Manual push and pull
• Tapestry, LN Recommender
– Modern automatic CF recommender systems
Social bookmarking
– Collaborative tagging systems
Social Search
23. Social Navigation in Information Space
Synchronous Direct
Communication in real time Direct communication
Asynchronous between people
Using the Interaction of past Indirect
users Relying on user presence
and traces of user behavior
Synchronous Asynchronous
Recommenders
Direct Chats
Q/A Systems
Indirect History-enriched
Presence of other people
environments
25. Amazon: Asynchronous, Indirect
Traces of viewing and purchasing decisions is a valuable collective wisdom!
•Compare with an Amazon
review: “the remake of this
movie is horrible, I recommend
to watch the original version
instead”
26. CourseAgent: Direct, Asynchronous
• Adaptive community-based course planning system
–Provides social navigation through visual cues
http://halley.exp.sis.pitt.edu/courseagent/
28. Generating Social Navigation
Overall workload
Averaging over all ratings of the community
Overall Relevance
Average does not work
Irrelevant to many but very relevant to one
Goal-centered algorithm
16 rules
29
29. Trade-offs for Direct Approach
• Reasonably reliable
• Feedback directly provided
• No need to deduce and guess
• Explicit feedback is hard to obtain
• Takes time to provide and requires
commitment
• “One out of a hundred”
• Social system, which extensively relies on
explicit feedback need either large
community of users or special approaches
to motivate direct contributions
31. The Intrinsic Motivation Works
• Career Planning was not advertised and was not
noticed and used by half of the students
• Contribution of experimental users who did not use
Career planning (experimental group I) is close to
control group
• Significant increase of all contributions for those who
had and used Career planning (experimental group
II)
32
32. More about CourseAgent
Farzan, R. and Brusilovsky, P. (2006) Social
navigation support in a course recommendation
system. In: V. Wade, H. Ashman and B. Smyth
(eds.) Proceedings of 4th International
Conference on Adaptive Hypermedia and
Adaptive Web-Based Systems (AH'2006), Dublin,
Ireland, June 21-23, 2006, Springer Verlag, pp.
91-100.
Farzan, R. and Brusilovsky, P. (2011)
Encouraging User Participation in a Course
Recommender System: An Impact on User
Behavior. Computers in Human Behavior 27 (1),
276-284.
33. Knowledge Sea II: Indirect, Asynchronous
•Social Navigation to support course readings
35. Trade-off for indirect approach
• Feedback is easy to get
• Users provide feedback simply by navigating and doing
other regular actions
• It works quite well
• Most useful pages tend to rise as socially important
• Social navigation cues attract users
• Indirect feedback might not be reliable
• A click or other action in the interface is a small
commitment, may be a result of error
• “Tar pits”
• Main challenge of systems based on indirect
approach: increase the reliability of indirect
feedback
• Better processing of unreliable events (time, scrolling)
• Use more reliable events (cf. browsing vs. purchase)
36. Knowledge See II: Beyond clicks
• Make better use of existing feedback
• Switched from click-based calculation of user
traffic to time based
• Time and patterns can provide more reliable
evidence
• Added annotation-based social navigation
• Annotations are more reliable
• Users are eager to provide annotations and
even categorize them into positive/regular
37. Spatial Annotation Interface
A Spatial Annotation Interface adds social
navigation on the page level
Staking a space
Commenting
BooksOnline'08 38
38. Page-level Navigation Support
Visual Cues - annotation background and border
Background Style
•Background filling
Ownership
•Background color
Owner’s attitude
Border style
•Border color
Positiveness
•Border thickness
# of comments
•Border stroke
Public or personal
BooksOnline'08 39
39. Annotation-based SN does work
• Usage
• With additional navigation
support map-based and
browsing-based access
emerged as the primary
access way
• Effect on navigation
• Significant increase of link
following (pro-rated
normalized access)
• Impact
• Annotation leads students
to valuable pages
40. Back to Motivation Issue
Annotations are explicit actions used for implicit feedback and as
with all explicit actions, it come with motivation problems.
BooksOnline'08
41. More on KS-II and AnnotatEd
Farzan, R. and Brusilovsky, P. (2005) Social navigation
support through annotation-based group modeling. In: L.
Ardissono, P. Brna and A. Mitrovic (eds.) Proceedings of
10th International User Modeling Conference, Berlin, July
24-29, 2005, Springer Verlag, pp. 463-472
Farzan, R. and Brusilovsky, P. (2008) AnnotatEd: A social
navigation and annotation service for web-based
educational resources. New Review in Hypermedia and
Multimedia 14 (1), 3-32.
Brusilovsky, P. and Kim, J. (2009) Enhancing Electronic
Books with Spatial Annotation and Social Navigation
Support. In: Proceedings of the 5th International
Conference on Universal Digital Library (ICUDL 2009),
Pittsburgh, PA, November 6-8, 2009
42. What is Social Search?
- Social Information Access in Search
context
- A set of techniques focusing on:
• collecting, processing, and organizing
traces of users’ past interactions
• applying this “community wisdom” in
order to improve search-based
access to information
43. Variables Defining Social Search
Which users?
• Creators
• Consumers
What kind of interaction is considered?
• Browsing
• Searching
• Annotation
• Tagging
What kind of search process improvement?
• Off-line performance improvement of search engines
• On-line user assistance
44. The Case of Google PageRank
Which users?
Which activity?
http://www.labnol.org/internet/google-pagerank-drop-stop-worrying/4835/
What is affected?
How it is affected?
How it improves search?
45. How Search Could be Changed?
Let’s classify potential impact by stages
Before search During search After search
46. Search Engines: Improve Finding
Use social data to expand document index
(document expansion)
What we can get from page authors?
Anchor text provided on a link to the page
What we can get from searchers?
Page selection in response to the query (Scholer,
2002)
Query sequences (Amitay, 2005)
What we can get from page visitors?
Page annotations (Dmitriev et al., 2006)
Page tags (Yanbe, 2007)
47. Search Engines: Improve Ranking
What we can get from page authors?
Links (Page Rank)
What we can get from searchers?
Page selection in response to the query
(DirectHit)
What we can get from page visitors beyond
seatch context?
Page visit count
Page tags (Yanbe, 2007; Bao, 2007)
Page annotations
Combined approaches
PageRate (Zhu, 2001), (Agichtein, 2006)
48. Using Social Wisdom Before Search
Can be done by both search engines and
external interfaces
Query checking - now standard
Suggesting improved/related queries
Example: query networks (Glance, 2001)
Automatic query refinement and query
expansion
Using past queries and query sequences - what the user is
really looking for (Fitzpatrick, 1997; Billerbeck, 2003;
Huang, 2003)
Using anchors (Kraft, 2004)
Using annotations, tags
49. Using Social Wisdom After Search
Better ranking, link promotion
• Link re-ordering using social wisdom (based on
the result selection traces by earlier searchers)
Suggesting additional results
• Suggest results (or sites!) found by earlier
searchers
Providing social annotations
• Link popularity, past link selection by socially
connected users
50. Challenges of Social Search
• Matching similar users
• Number of page hits is not reliable (DirectHit failure)
• Using “everyone” social data is a bad idea – need not
good pages overall, but those that match a query
• Even matching with users who issue the same query is
not reliable enough – same query, very different goals!
• Reliability of social feedback
• A click on a result link is not a reliable evidence of
quality and relevance
• Need to do a wise mining of search sessions and
sequences
• Fusing query relevance and social wisdom
• Single ranking is not the best way to express two
dimensions of relevance
51. AntWorld: Quest-Based Approach
– Quests establish similarities between users
– Relevance between documents and quests is provided
by explicit feedback
52. Quest Approach to Social Search
Evaluation of Quest approach: SERF (Jung,
2004)
– Results with recommendations were shown on over 40%
searches.
– In about 40% of cases the users clicked and 71.6% of
these clicks were on recommended links! If only Google
results are shown users clicked in only 24.4% of cases
– The length of the session is significantly shorter (1.6 vs
2.2) when recommendations are shown
– Ratings of the first visited document are higher if it was
recommended (so, appeal and quality both better)
55. Other Ways to Increase Reliability
• Moving from single query to query sequences
• What the user selected at the end
• Moving from page recommendation to site recommendation
White, R., Bilenko, M., and Cucerzan, S. (2007) Studying the use of popular
destinations to enhance web search interaction. In: SIGIR '07, Amsterdam, The
Netherlands, July 23 - 27, 2007, ACM Press, pp. 159-166
56. Social Search with Visual Cues
Query relevance and social relevance shown separately: rank/annotation
Similarity score
General annotation
Question
Document with high traffic (higher rank) Praise
Negative
Document with positive annotation
(higher rank) Positive
57. Annotation-Based Search: Impact
Acceptance
– Users noticed and applied social visual cues
• Frequency of usage - viewed more documents per query
with social visual cues
– Users agreed with the need for social search
• Survey results
Performance
– Social Visual Cues are taken into account for
navigation
• Social Navigation cues are twice as more influential in
affecting user navigation decision than high rank
– Social visual Cues provide higher prediction for
page quality that high rank
More information
– Ahn, J.-w., Farzan, R., and Brusilovsky, P. (2006)
Social search in the context of social navigation.
Journal of the Korean Society for Information
Management 23 (2), 147-165.
58. SIA Challenges across Contexts
• Increasing reliability of indirect sources
• Time spent reading vs. simple click
• Query sequences vs. simple result access
• Adding more reliable evidences of
relevance/quality/interests
• Annotation vs. browsing
• Purchasing/downloading vs. viewing
• May add the problem of motivation!
• Basis for user similarity (not “all for all”)
• Co-rating in recommender systems (sparsity!)
• Users with similar goals (CourseAgent)
• Single class in Knowledge Sea II (still topic drift!)
• Quest or community in AntWorld and iSpy
59. More Challenges:Merging the Technologies
• Different branches of SIA have little connections
to each other
• Social navigation use navigation data to assist
navigation
• Social search use search traces to assist future
searchers
• Many opportunities to merge two or more SIA
technologies
• Social Web system with broader SIA
• Use several kinds of user traces to support a specific SIA
technology
• Offer several kinds of SIA
• Earlier work: Social Navigation + Social Search
– ASSIST ACM
– ASSIST YouTube
• Social Navigation + Recommendation
• Adding Social Visualization
60. ASSIST-ACM: Social Search + Nav
Re-ranking result-list Augmenting the links
based on search and based on search and
browsing history browsing history
information information
Farzan, R., et al. (2007) ASSIST: adaptive social support for information space traversal.
In: Proceedings of 18th conference on Hypertext and hypermedia, HT '07,, pp. 199-208
62. Some New Ideas in CoMeT
• Broader set of evidences
• View, annotate, tag, schedule talks, send to
friends, connect to peers
• Declare affiliations (similarity!)
• Join and post links to a set of communities
• Combining in-context (visual cues) and
out-of context (ranking) guidance
• Exploring the power of “top N”
• Powerful, but dangerous!
66. Community vs. Peer-Based NS: E-
learning
• Progressor and Progressor+ projects
• Problem: guide students to most
appropriate educational content –
examples, problems, etc.
• Using reliable indicators of student
progress (problem solving success)
• Provide visualization to better support
guidance
• Explore peer-based and community-based
SNS
72. How Social Guidance Works
Non-adaptive adaptive
Social, adaptive, single content Progressor+
73
Notes de l'éditeur
The most cited reason of DirectHit failure was low query repetition, which made the social data collected by it too sparse to use frequently and reliably. User diversity was another likely contribution: users with different goals and interests may prefer different results returned by the same query. Finally, the proposed approach to link ranking was too easy to abuse by malicious users who wanted to promote their favorite pages.
Three levels of QQ similarity: direct, through document corpus (all documents returned by a query), through user selections (terms in selected documents)
AntWorld introduced the concept of a quest , which is an information goal pursued by a user over a sequence of queries (Fig. 3). The system successfully encouraged its users to describe their quests in natural language and used this description to determine inter-quest similarity. During their search, the users were able to rank search results by their relevance to the original quest (not a query used to obtain this result!). These innovations allowed the system to address to some extent the sparsity and reliability problems. To determine documents, which are socially relevant for a particular quest, the system looked for positively ranked documents in past similar quests. The system assisted the user by adding socially relevant documents to the list of search results and also adding a small ant icon to socially relevant links returned during each search within the quest
Results with recommendations were shown on over 40% searches. In about 40% of cases the users clicked and 71.6% of these clicks were on recommended links! If only Google results are shown users clicked in only 24.4% of cases - so when social recommendations are provided, chance to click is higher. Also the length of the session is significantly shorter (1.6 vs 2.2) when recommendations are shown. Finally, ratings of the first visited documents are higher if it was recommended (so, appeal and quality both better). In more than 2/3 of cases the users really provided expanded search requests - over 1 sentence! However, regardless of social help, the user rate visited documents only in 2/3 of cases.
Progressor+ significantly outperformed QuizJET. Students spent more time per session in Progressor+ than QuizJET. Students spent more time per session in Progressor+ than Progressor. introducing annotated examples to the open social student modeling visualization did not sacrifice the usage in selfassessment quizzes. Providing personalized guidance in open social student modeling interface (Progressor+) was equivalent efficient as non-social open student modeling interface (JavaGuide). More than that, students did spend more time in studying the annotated examples. Whoever works on 1 type of content, more likely to work on the other type (r=0.81, p<.01) *between subject ANOVA, used Bonferroni adjustment, that is the most conservative method.
- Problem solving importance to knowledge acquisition Why not perfect (100%) - knowledge-based and social-based combination indeed brought added value to the system, where the knowledge-based personalization alone did not.
common: general pattern on exam preparation, especially in final exam period