phd defense: dynamic generation of personalized hybrid recommender systems
TRANSCRIPT
Dynamic Generation of Personalized Hybrid Recommender Systems
Simon DoomsPublic PhD Presentation
December 19, 2014. Belgium.
News Movies
RestaurantsHotelsBooks
TV shows
Clothes
Apps Laptops
CultureBars Recipes
Comic books Perfume
Conferences
Websites
GamesColors
There’s too much of everything
Spotify’s music catalog contains 20 million songs
Every minute 100 hours of video uploaded to
YouTube
Every 5 minutes a new book on Amazon
We can’t have it all
News Movies
RestaurantsHotelsBooks
TV shows
Clothes
Apps Laptops
CultureBars Recipes
Comic books Perfume
Conferences
Websites
GamesColors
The Solution?
Recommender Systems
Content-based FilteringItemAttributeKNNFactorWiseMatrixFactorization
MatrixfactorizationItemKNN
SigmoidCombinedAsymmetricFactorModel
SigmoidItemAsymmetricFactorModel
SigmoidUserAsymmetricFactorModel GlobalAverage
ItemAverage
SVDPlusPlus
TimeAwareBaselineWithFrequencies
SlopeOneUserKNN
BiPolarSlopeOne
LatentFeatureLogLinearModel
SVD
Collaborative FilteringBiasedMatrixFactorization
Random Items
SigmoidSVDPlusPlus
CoClusteringUserItemBaseline
TimeAwareBaseline
Probability-based Extended Profile Filtering
Recommender Systems
Content-based FilteringItemAttributeKNNFactorWiseMatrixFactorization
MatrixfactorizationItemKNN
LatentFeatureLogLinearModel
SVD
Collaborative FilteringBiasedMatrixFactorization
Random Items
SigmoidSVDPlusPlusTimeAwareBaseline
Probability-based Extended Profile Filtering
AanbevelingssystemenWhat’s the challenge?
Content-based FilteringItemAttributeKNNFactorWiseMatrixFactorization
MatrixfactorizationItemKNN
LatentFeatureLogLinearModel
SVD
Collaborative FilteringBiasedMatrixFactorization
Random Items
SigmoidSVDPlusPlusTimeAwareBaseline
Probability-based Extended Profile Filtering
How to select the best system for a given context, user or domain?
Goal
Dynamic Generation of Personalized Hybrid Recommender Systems
The title
Goal
Dynamic Generation of Personalized Hybrid Recommender Systems
Goal
Dynamic Generation of Personalized Hybrid Recommender Systems
Goal
Dynamic Generation of Personalized Hybrid Recommender Systems
Goal
Dynamic Generation of Personalized Hybrid Recommender Systems
Goal
We Need Data
Products
Preferences
Data
MovieLens Netflix
(Old) Data is available
Data
MovieLens Netflix
1994 1995 1997
(Old) Data is available
Data
Old Data Old Recommendations
(Old) Data is available
but…
Data
+
Searching Recent Data
Data Searching Recent Data
Data
“I rated #IMDb”
Found it!
Data
Collected during 1 year, 9 months
320 000 ratings
MovieTweetings Dataset
30 000 users
20 000 movies
Found it!
Goal
Goal
+ + + +
Personalized
Personalized
Optimization Problem
Data
Optimization
Training EvaluateCombine
Data
Optimization
Training EvaluateCombine
Data
Optimization
Training EvaluateCombine
Data
Optimization
Training Evaluate
+
Combine
Repeat
Data
Optimization
Training Evaluate
+ = 25
= 50
= 75
Adapt
Combine
Optimize
Fold
dat
aset
s
Slow (hours)
All
dat
a
Fast (seconds)
Full Model
Optimize
Fold
dat
aset
s
Slow (hours)
All
dat
a
Fast (seconds)
New Ratings: no re-training required
Optimization Results
It works
… in theory
Real-life evaluation?
Goal
User Evaluation
Google Chrome Extension
Movie Recommendations
Recent Movies
Interaction
Recent Movies
Movie Recommendations
Interaction
Recent Movies
Movie Recommendations
Interaction
Recent Movies
Evaluation
Explicit
Movie Recommendations
Interaction
Recent Movies
Evaluation
Explicit
ImplicitClick Tracking & Logging
Movie Recommendations
Movie Recommendations: Results
Interactive users use the system more often
All users are different
Explicit & Implicit evaluation was positive
Goal Achieved
User Interface Design Experiments
But wait … there’s more!
High-Performance Calculation Research
User Interface Design Experiments
But wait … there’s more!
High-Performance Calculation Research
User Interface Design Experiments
But wait … there’s more!
High-Performance Calculation Research
User Interface Design Experiments
But wait … there’s more!
Rating systems, online experiments
High-Performance Calculation Research
User Interface Design Experiments
But wait … there’s more!
Conclusion
A dynamic, personalized hybrid recommender system works
All users are unique
Interactive systems are the future
Conclusion
You might also like…
My PhD dissertation for more details:
Google Scholar for my publications
Slideshare for my presentations and posters
Github for my public code and data
http://bit.ly/simonphd
https://github.com/sidooms
That’s all folks!
More questions? Contact me!
https://twitter.com/sidooms
simon [dot] dooms [at gmail dot] com
Dynamic Generation of Personalized Hybrid Recommender Systems
Simon DoomsPublic PhD Presentation
December 19, 2014. Belgium.