Upload
marco-fossati
View
633
Download
1
Tags:
Embed Size (px)
DESCRIPTION
Talk at SeRSy workshop, co-located at ISWC 2012, Boston, U.S.
Citation preview
Semantic Network-driven News Recommender
Systems A Celebrity Gossip Use Case
Marco Fossati
claudio giuliano
Giovanni Tummarello
Web of Data Unit
Fondazione Bruno Kessler
Trento, Italy
Post-click news recommendation
2
Typical approaches ª Issues
² Data sparsity ² Implicit user profile
interpretation ² Lack of
recommendation explanation
3
ª Collaborative Filtering
² User profile
ª Content-based
² Keyword-driven
ª Issues ² Lack of
recommendation explanation
² “More of the same”
Challenge Provide interesting recommendations
to an anonymous user via large scale structured knowledge bases
4
Proposed approach
Entity linking
Lindsay Lohan
Dina Lohan
Michael Lohan
Entity list Source article
Extract types + properties
Fully described entity set
Pre-built recommenders
Lindsay Lohan: [actress, dated, American, legal problems]
…
dated dead
celebrities legal
problems …
5
Candidate object
entities
W. Walderrama
Leo Di Caprio
Britney Spears
Johnny Depp
Candidate
recommended articles
• Article A • Article B • Article C
• Article X • Article Y • Article Z
1. Article B
2. Article Z
3. Article X
Ranking and winner
Triggerable
recommenders
Legal problems
Dated
A semantic recommender
SELECT articles that mention entities in
relation R with X !
Candidate Articles
Entity-linked Corpus
1 2
Triple Store
Source article entity (X)
• Article 1 • Article 3 • …
• Article 2 • Article 4 • …
<X, R, Y>
<X, R, Z>
6
Specific Explanations
Lindsay dated
Lindsay dated
Y
Z
Explanation template
“X dated Y. See what Y did” !
SELECT articles that mention Y who dated X !
(X)
Evaluation
ª Online with real users
ª Crowdsourcing
7
“Which is the recommendation that best attracts your attention?”
Objectives
ª Recommendation strategies competition ² Ours (hybrid) ² Baseline (LSA+BOW)
² Fake (random)
ª Specific explanation
ª Simplified specific explanation
ª No specific explanation
8
Lindsay dated Leo. !Read more about him !
Read more about !who dated Lindsay !
Related stories !selected for you !
Job unit example
9
Results
♣ indicates statistical significance difference between the baseline and our method, with p<0.001
10
Objective Fake % Baseline % Ours %
Specific explanation 3.33 23.33 73.33♣
Simplified specific explanation 5.88 41.17 52.94
Without specific explanation 13.63 37.5 48.86
ª 810 judgments
ª 36.6 $
ª 10 jobs
ª 10 units/job
Discussion
ª Significant difference with specific explanations ² Disappears while decreasing the
specific explanation complexity
ª Results seem comparable with the baseline even in absence of explanation
11
Future work
ª Methodologies for ² Generic semantic recommenders
building ² Natural language specific
explanations
ª More experiments on the recommendation quality
12
Conclusion
ª Our approach enables ² Rich explanations ² Diverse/unusual recommendations
ª Evaluation shows ² No explanation à comparable results ² With explanation à significantly better
13