Hybrid Event Recommendation using Linked Data and User Diversity
1. Hybrid Event Recommendation
using
Linked Data and User Diversity
Houda Khrouf and Raphaël Troncy
{khrouf, troncy}@eurecom.fr
Eurecom, Sophia Antipolis, France
The 7th ACM Recommender Systems Conference
Oct 12-16, 2013 Hong Kong
2. Outline
Event Recommendation
Collaborative Filtering
Content-based
RDF Modeling and Similarity computation
User Interest Detection
Hybrid Approach
Evaluation and Conclusion
10/15/2013
7th ACM Recommender Systems 2013, Hong Kong
2
3. Events on the web
Millions of active users
Thousands of events per day
Highly diverse content
Recommender Systems?
10/15/2013
7th ACM Recommender Systems 2013, Hong Kong
3
4. What do users think?
Seen on Last.Fm
10/15/2013
7th ACM Recommender Systems 2013, Hong Kong
4
5. Is this event interesting?
Decision
Time
Places
EVENTS
Decision factors (depends on type)
• Where is it? (Location)
• Who’s going? (Participants)
Attendees
Tags/Topics
• When is it? (Time)
• What is it? (Content)
• Who is involved? (players)
10/15/2013
7th ACM Recommender Systems 2013, Hong Kong
Performers
5
6. Collaborative Filtering (CF)
Predict the event the user will attend
based on the attendance of other
like minded users
Events are transient items
inducing a very sparse user
attendance matrix (sparsity 99%)
similar
Best choice to reflect the social
dimension, but:
Apart from the social information,
there is no explicit consideration
of the other factors
10/15/2013
7th ACM Recommender Systems 2013, Hong Kong
6
7. Content-based Recommendation (CB)
Recommend new events that match the user profile based on
their descriptions
Event context:
-
Location (geo-coordinates, city…)
-
Time
-
Topics/Tags
-
Performers (genres, tags…)
Events are entities with attributes and relational attributes (links)
to other entities
Events similarity depends on the similarity of related entities
10/15/2013
7th ACM Recommender Systems 2013, Hong Kong
7
8. User Profile
The user profile is based on past attended events
Topical Diversity: real-world events range from large festivals to
small concerts and social gatherings
A user might be interested in some specific topics/performers
during the event
We need to alleviate the profile
diversity and detect the user’s
interests
10/15/2013
7th ACM Recommender Systems 2013, Hong Kong
8
9. Approach and Contributions
Events similarity
Structured RDF event model
Similarity in Linked Data
Data enrichment with DBpedia
User interests detection using LDA (Latent Dirichlet
Allocation)
Hybrid recommendation (CF+CB)
10/15/2013
7th ACM Recommender Systems 2013, Hong Kong
9
10. LODE Ontology
LODE is a minimal model that encapsulates the factual properties of events: What,
Where, When and Who.
URL: http://linkedevents.org/ontology
10/15/2013
7th ACM Recommender Systems 2013, Hong Kong
10
11. subjects
Linked Data in a Tensor Space
objects
For each property p, and for each object op [1]
𝑾𝑾 𝒐𝒐,𝒆𝒆 = 𝒇𝒇 𝒐𝒐,𝒆𝒆 ∗ 𝒍𝒍 𝒍𝒍 𝒍𝒍
𝒑𝒑
𝒑𝒑
|𝑬𝑬|
|𝑬𝑬 𝒐𝒐,𝒑𝒑 |
[1] T. Di Noia et al. Linked open data to support content-based recommender systems. In 8th
International Conference on Semantic Systems, Graz, Austria, 2012.
10/15/2013
7th ACM Recommender Systems 2013, Hong Kong
11
12. Events Similarity
Similarity according to one property p:
𝐜𝐜𝐜𝐜𝐜𝐜𝐜𝐜 𝐜𝐜 𝐩𝐩 𝐞𝐞 𝟏𝟏 , 𝐞𝐞 𝟐𝟐 =
∑ 𝒐𝒐∈𝑶𝑶
𝒑𝒑
∑ 𝒐𝒐∈𝑶𝑶 𝒘𝒘 𝒐𝒐,𝐞𝐞 𝟏𝟏
𝟐𝟐
𝒘𝒘 𝒐𝒐,𝐞𝐞 𝟏𝟏 ∗ 𝒘𝒘 𝒐𝒐,𝐞𝐞 𝟐𝟐
𝒑𝒑
∗
𝒑𝒑
𝒑𝒑
∑ 𝒐𝒐∈𝑶𝑶 𝒘𝒘 𝒐𝒐,𝐞𝐞 𝟐𝟐
𝟐𝟐
Similarity between two events:
𝒔𝒔𝒔𝒔 𝒔𝒔 𝒆𝒆 𝟏𝟏 , 𝒆𝒆 𝟐𝟐 =
∑ 𝒑𝒑∈𝑷𝑷 𝜶𝜶 𝒑𝒑 𝒄𝒄𝒄𝒄𝒄𝒄𝒄𝒄𝒄𝒄 𝒑𝒑 𝒆𝒆 𝟏𝟏 , 𝒆𝒆 𝟐𝟐
|𝑷𝑷|
Not adapted for discriminant properties associated with
highly sparse adjacency matrix
10/15/2013
7th ACM Recommender Systems 2013, Hong Kong
12
13. Events Similarity
Discriminability
𝑫𝑫𝑫𝑫𝑫𝑫𝑫𝑫 𝒑𝒑 =
𝒐𝒐 𝒕𝒕𝒕𝒕𝒕𝒕𝒕𝒕𝒕𝒕𝒕𝒕 = < 𝒔𝒔, 𝒑𝒑, 𝒐𝒐 > ∈ 𝑮𝑮 |
|𝒕𝒕𝒕𝒕𝒕𝒕𝒕𝒕𝒕𝒕𝒕𝒕 = < 𝒔𝒔, 𝒑𝒑, 𝒐𝒐 > ∈ 𝑮𝑮|
Similarity-based Interpolation
𝑾𝑾
𝒑𝒑
𝒐𝒐 𝟐𝟐 ,𝒆𝒆 =
10/15/2013
|𝑬𝑬|
𝐦𝐦𝐦𝐦𝐦𝐦𝐦𝐦𝐦𝐦𝐦𝐦(𝒐𝒐 𝟐𝟐 , 𝒐𝒐) ∗ 𝒍𝒍 𝒍𝒍 𝒍𝒍
𝒐𝒐∈𝑶𝑶 𝒑𝒑,𝒆𝒆
|𝑬𝑬 𝒐𝒐 𝟐𝟐 ,𝒑𝒑 |
7th ACM Recommender Systems 2013, Hong Kong
o1
p
similar
e
p
Interpolation of a
fictitious link
13
o2
14. Interest Detection
How to detect user interests from diverse event space?
Latent Dirichlet
Topic distribution over
each event (T=30)
Allocation (LDA)
Events
[Blei et al 2003]
Tags
Diversity
score
Attended
events Eu
User Interest
Mean of
the variances
10/15/2013
Distribution
Variance of each
topic
7th ACM Recommender Systems 2013, Hong Kong
14
16. Experiments
Open RDF Dataset (EventMedia)
Visualization: http://eventmedia.eurecom.fr
SPARQL: http://eventmedia.eurecom.fr/sparql
Learning the rank weights:
Linear regression with gradient descent
Genetic Algorithm
Particle Swarm Optimization
Evaluation: training 70% - test 30 %
2.436 events in UK from Last.Fm , 481 active users, 14.748 artists,
897 locations (available on request)
precision/recall of Top-N recommendations
10/15/2013
7th ACM Recommender Systems 2013, Hong Kong
16
18. User Diversity
Score ≈ 1 => strong interest
Score ≈ 0 => cold-start users
Most of users have relatively high interests towards some topics
10/15/2013
7th ACM Recommender Systems 2013, Hong Kong
18
20. CB+CF evaluation
𝛃𝛃 𝐢𝐢 𝐢𝐢 𝐢𝐢 𝐢𝐢 𝐢𝐢𝐢𝐢𝐢𝐢𝐢𝐢 >
𝟒𝟒 × 𝛃𝛃 𝐧𝐧𝐧𝐧−𝐢𝐢 𝐢𝐢 𝐢𝐢 𝐢𝐢 𝐢𝐢𝐢𝐢𝐢𝐢𝐢𝐢
Interest Detection
High influence of social
information in event
recommendation
10/15/2013
7th ACM Recommender Systems 2013, Hong Kong
20
21. Comparison with other approaches
Probability based Extended Profile Filtering (UBExtended): T. D. Pessemie et al. Collaborative
recommendations with content-based filters for cultural activities via a scalable event distribution platform.
Multimedia.Tools Appl., 58(1):167-213, 2012.
10/15/2013
7th ACM Recommender Systems 2013, Hong Kong
21
22. Conclusion
Effectiveness of Semantic Web technologies to steer data retrieval
and processing
Importance of the social information and the user interest model in
event recommendation
Future work:
Other features: popularity, temporal patterns, weather, etc…
Test the system scalability on large datasets using spatial and/or
temporal indexing of user attendance
10/15/2013
7th ACM Recommender Systems 2013, Hong Kong
22