RecSys’15: Context-Aware Event Recommendation in Event-based Social Networks
Recommendation system paper challenge (7/50)
Why this paper?
RecSys’15 and top popular recommendation system paper.
What problem do they solve?
Event-based recommendation system (meetup), which usually suffer from col-start problem.
What will I solve this problem?
I will apply XGBoost model, which includes user interests and the location between the user and the event and the similar user attending. Considering content-based features into the model.
What others consider long-term user value?
Most researchers focus on the traditional recommendation system which has rich interaction history of user and items. However, in this system, most events are new-item.
What is the baseline models?
MP (Most popular): It ranks events in descending order of the number of positive RSVPs that they received.
BPR-MF: The Bayesian Personalized Ranking is a state-of-the-art matrix factorization-based algorithm for top-n item recommendations.
BPR-NET: The overall idea is to use two kinds of social networks as regularization terms of a BPR-MF model, namely, the social network based on the shared groups of users, and the social network inferred from the co-attended events.
What model do they propose?
They provide MCML model (Multi-Contextual Learning to Rank Events), which includes U user, E event, T temporal preference of user, G group, C text, D distance.
Here, group is the the group holding the event. Some groups will continually hold new events.
Group based similarity function
They apply BPR model for L function and this formula can share U and E variables together.
Content-aware: user embedding
The basic idea for model user embedding is applying the historical event the user RSVP.
For each event, they apply bag-of-words to map the event to a vector and then aggregated these vectors with time decay function.
Location-aware: (distance between user to the event)
They utilize kernel-based density estimation approach to model the mobility patterns of individual users as distributions of geographic distances between the attended events.
For example, some users prefer participate near events; while other don’t.
The basic idea is to measure the distance between the new event and the old events user participated before.
Time-aware: (distance between user to the event)
Different users have different temporal preference.
It is easy to represent user as the 2D vector, averaging the temporal zone of all events the user participated.
Learning to Rank Events with Contextual Features
With all context-aware similarity given user and event, we can normalized all of them and get the score (h) and then we apply Learning to Rank.
How to setup the training, testing data?
The Result
Other related blogs:
Trust-aware recommender systems
Performance of recommender algorithms on top-n recommendation tasks
A Matrix Factorization Technique with Trust Propagation for Recommendation in Social Networks
Hidden Factors and Hidden Topics: Understanding Rating Dimensions with Review Text
Beyond Clicks: Dwell Time for Personalization
Best paper in RecSys:
https://recsys.acm.org/best-papers/
My Website: