2. Abstract
• The amount of data in the digital universe is estimated to hit 1.2
Zettabytes (1 billion terabytes) during 2010.
• These data quantities make discovering relevant information a difficult
task.
• Recommender Systems are an integral tool for assisting users in
information discovery.
• By combining wisdom of crowds, content, user profiles, etc.
Recommender Systems find relevant data for us.
“We are leaving the age of information and entering the age of recommendation”
Chris Anderson, The Long Tail
3/18/2022 Talis 2
4. Introduction
• IMDb, one of the first online recommender systems, turned 20 on October
17th 2010.
• Ever since, recommender systems have, through relatively simple
techniques, produced adequately good results
• Is adequately good good enough?
– How can recommender systems be improved?
– What do we need to improve them?
3/18/2022 Talis 4
5. Recommender System Types
Introduction
• Semantic recommenders – explicit information
– Content
– Keywords
– Genre
– etc.
• Social recommenders – implicit information (collaborative filtering)
– Item-based user-user similarities, i.e. which users like similar things
– Content-ignorant
• Hybrid recommenders
– Combinations of content- and CF-based
• Context-aware recommenders
– Aware of the current situation
3/18/2022 Talis 5
7. Social recommenders
Most common recommender
systems approach use
Collaborative Filtering
How does collaborative filtering
work?
• Calculates similarities between all users
• Finds users similar to you
• Fills in your ”gaps” based on similar
users, usually by a k-nearest neighbor
algorithm
3/18/2022 Talis 7
Recommend a book for user C
8. Social recommenders
Most common recommender
systems approach use
Collaborative Filtering
How does collaborative filtering
work?
• Calculates similarities between all users
• Finds users similar to you
• Fills in your ”gaps” based on similar
users, usually by a k-nearest neighbor
algorithm
3/18/2022 Talis 8
Recommend a book for user C
9. Social recommenders
Most common recommender
systems approach use
Collaborative Filtering
How does collaborative filtering
work?
• Calculates similarities between all users
• Finds users similar to you
• Fills in your ”gaps” based on similar
users, usually by a k-nearest neighbor
algorithm
3/18/2022 Talis 9
Recommend a book for user C
10. Hybrid models
Hybrid recommender systems
combine semantic recommenders
with collaborative filtering ones.
3/18/2022 Talis 10
Recommend a book for user C
11. Hybrid models
Hybrid recommender systems
combine semantic recommenders
with collaborative filtering ones.
3/18/2022 Talis 11
Recommend a book for user C
13. What is context?
Context-awareness in RecSys
”Any information that can be used to
characterise the situation of entities”,
Dey 2001
1. Item context
• Seasonal (Christmas, Oscar’s)
• Relation (movie sequel, director, actor)
2. User context
• Surroundings (weather, location)
• Company (alone, with friends)
• Mood/emotions
• any user related factor
3/18/2022 Talis 13
14. Why Context?
Context-awareness in RecSys
3/18/2022 Talis 14
+
• Filters relevant information
• Ad hoc recommendations
• Aware of changes
-
• What is context?
• Where do we find it?
15. Applying Context-awareness
Current state of the art research
presents two types of context-
awareness:
• Context-aware collaborative
filtering
– Performs standard CF on virtual,
contextual, items or users
– Benefits: simple
– Drawbacks: statically defined context
3/18/2022 Talis 15
16. Applying Context-awareness
Current state of the art research
presents two types of context-
awareness:
• Context-aware collaborative
filtering
– Performs standard CF on virtual,
contextual, items or users
– Benefits: simple
– Drawbacks: statically defined context
• Tensor factorization for context-
awareness
– Models the data as a tensor
– Applies higiher-order factorization
techniques (HoSVD, PARAFAC,
HyPLSA, etc) to model context in a
latent space
– Benefits: no prior context
identification necessary
– Drawbacks: adds complexity
3/18/2022 Talis 16
17. My work
3/18/2022 Talis 17
Semantic recommenders
Social recommenders
Context-aware recommenders
18. Where does this fit at Talis?
• Library data
– Loan events – CF
– Book meta data – semantic recommenders
– Time of loan event – context-awareness
3/18/2022 Talis 18
19. Distributed higher order
recommender system
• Use matrix factorization techniques
to make a tensor factorization
approximation in MapReduce
• By matricizing the tensor, standard
matrix factorization approaches can
be run in parallel
• What is matrix factorization?
– Decomposition of a matrix into its
building blocks (SVD example)
• A = UΣVT where A is the matrix, Σ is a
diagonal matrix and U and V are unitary
matrices.
• By only taking the k first diagonal values in
Σ and multiplying the resulting matrix
back with U and V we obtain a k ranked
approximation of the initial A matrix
3/18/2022 Talis 19
book
user