Seal of Good Local Governance (SGLG) 2024Final.pptx
A Framework for Learning and Analyzing Hybrid Recommenders based on Heterogeneous Semantic Data
1. Competence Center Information Retrieval & Machine Learning
A Framework for Learning and Analyzing Hybrid
Recommenders based on Heterogeneous Semantic Data
Andreas Lommatzsch, Benjamin Kille, Sahin Albayrak
27. Mai 2013 OAIR 2013 – Session 5 - Recommendation
3. Problem Description
► Data structure
(user, item, preference)
(Anna, Iron Man 3, 4 stars)
…
► Approach: Collaborative
Filtering
3OAIR 2013 – Session 5 - Recommendation
► Recommendation Task
Collection of items
User preferences
Find most relevant items
in collection according to
the target user‘s
preferences
27. Mai 2013
► Cold-Start Problem: Little to no preferences of new users are
available
► Solution: Hybridisation with Content-Based Filtering
► Semantic Recommenders using semantic data
4. Semantic Recommenders
427. Mai 2013 OAIR 2013 – Session 5 - Recommendation
Movie Actors Directors Genres …
Robert Downey Jr.
Gwyneth Paltrow
Don Cheadle
Guy Pierce
Ben Kingsley
…
Shane Black Action
Adventure
Science Fiction
…
Source:http://www.imdb.com/title/tt1300854/
► Data Representation:
(item, attribute (entity relationship set = instance), exist?)
(Iron Man 3, actor=Robert Downey Jr., true)
(Iron Man 3, actor=Keanu Reeves, false)
► Strategy: recommend items with similar/overlapping
attributes
6. Research Questions
627. Mai 2013 OAIR 2013 – Session 5 - Recommendation
► How to combine different entity relationship sets?
► Should we apply dimensionality reduction techniques to
reduce existing noise (model-based vs. memory-based
recommenders)?
► In what way do we need to scale data which is typically binary?
7. Combination Strategies
727. Mai 2013 OAIR 2013 – Session 5 - Recommendation
► Block matrices
► Agent Ensembles:
Each agent models an individual entity relationship set
Recommendation of agents are subsequently assembled
11. Conclusions
1127. Mai 2013 OAIR 2013 – Session 5 - Recommendation
► Agent ensembles obtain superior MAP compared to block
matrices (combination strategy). This effect depends on their
ability to consider differences in between entity relationship
sets. In contrast, block matrices treat all entity relationship sets
equally.
► Scaling can both improve and spoil recommendation quality.
► We observed superior MAP for model-based recommenders
compared to memory-based (dimensionality reduction).
► Semantic recommenders can be applied to any data that can be
represented as triples.
► More and more semantic data sources become available the
emerging and success of semantic recommenders will likely
continue.
12. Future Work
1227. Mai 2013 OAIR 2013 – Session 5 - Recommendation
► Applying the approach to other domains (music, news,
products, etc.)
► Analyze hybridization with collaborative filtering
► Evaluate learning methods for weighting agents
► Investigate what data characteristics (density, entropy,
connectedness, etc.) should be considered when combining
various recommenders in an ensemble
13. Announcement: NRS 2013
1327. Mai 2013 Challenges in Cross-Domain News Article Recommendations
► International News Recommender Systems Workshop and Challenge
► In conjunction with ACM RecSys 2013
IMPORTANT DATES
July 1, 2013 paper submission deadline
July 1, 2013 data set release
August 15, 2013 on-line challenge kick-off
HIGHLIGHTS
Access to a real recommender system
Real-time requirements
Big Data
Cross-domain
Implicit feedback
Website: https://sites.google.com/site/newsrec2013/home
Twitter: @NRSws2013