2. Образец заголовкаMotivation
• Modern consumers are inundated with choices!
– Which movie should I see?
– What web page has the information I need?
– Where should I order food from?
– Which places should I visit on my holiday?
– Which book should I buy?
• Recommender systems help the decision process by suggesting items that a user might value based on
the space of possible items – this has become a major aspect of user experience
• Collaborative filtering is the de facto mechanism for recommendation
– The recommender compares user’s ratings to other users to find similar users, and recommends
items
• When a new user comes along, the system knows nothing about the user! – this is called the cold start
problem
The winners of the Netflix Prize. Source: NYT
6. Образец заголовкаSelecting the right items to rate
Main Idea
• New users are asked to rate items from a seed set
• Seed set previously chosen using criteria such as popularity, contention
and coverage
• However, these criteria are arbitrary and different combinations might be
required for optimal predictions
• Seed set items selected in a principled manner by minimizing a cost
measure
Technique
• Given a user preference prediction algorithm A, and a cost measure F, use
a greedy algorithm to iteratively add items which minimize the cost
measure
• Specifically, A is the factorized item-item model, and cost measure is RMSE
Golbandi, N., Koren, Y. and Lempel, R. 2010. On Bootstrapping Recommender Systems. Proceedings of the 19th ACM International Conference on Information
and Knowledge Management(New York, NY, USA, 2010), 1805–1808.
RMSE =
s X
(u,i)2T
(rui ˆrui)2/(|T|)
11. Образец заголовка
Functional matrix factorization for
cold start recommendation
Main Idea
• Initiate several trials when a new user visits the site, and at each trial ask the user to provide their opinion on a
seed item
• In Golbandi et al, WSDM 2011, a decision tree is fit to the users’ ratings – this technique extends the method by
integrating decision tree construction into the matrix factorization framework, and the authors call this method
Functional Matrix Factorization
Technique
• Compared to traditional matrix factorization, no information about the user profile is available to predict
ratings
• The user’s answers on P questions are mapped to a user’s profile using a function, and the goal is to learn both
the item profiles and the user profiles
• Alternative optimization is used for calculating the item profiles given the mapping function, and then the
mapping function is obtained by fitting a decision tree, given the item profiles
• To prevent overfitting, hierarchical regularization is used, where the coefficient of a node is shrunk towards its
parent
Zhou, K., Yang, S.-H. and Zha, H. 2011. Functional Matrix Factorizations for Cold-start Recommendation. Proceedings of the 34th International ACM SIGIR Conference on
Research and Development in Information Retrieval (New York, NY, USA, 2011), 315–324.
12. Образец заголовка
Functional matrix factorization for cold
start recommendation (contd.)
Results
• This method is compared with Tree (Golbandi et al., WSDM
2011) and TreeU (learns decision tree using the Tree algorithm
and matrix factorization through two separate steps)
• Performance improves as number of interview questions
increases for all three methods
• fMF performed better than Tree and Tree U on all three
datasets – MovieLens, EachMovie, and Netflix
• The authors studied the impact of non-responses, and model
parameters
• Compared to the warm-start matrix factorization technique,
this method worked poorly since it was constrained to ask a
few questions – however the performance is comparable if the
depth of the tree is allowed to be large
13. Образец заголовка
Representative based rating
elicitation
Main Idea
• Select representatives – those users/items whose linear combinations of profiles
accurately approximate other users’/items’ profiles
Technique
• Propose representative based matrix factorization – ratings matrix has a rank-k
approximation: Y = CX, C consists of columns of Y, X is the loading matrix with free
parameters
• The dimensionality of the column space of Y is reduced using a rank-k SVD, and the
basis vectors (representatives) are chosen using the maximal volume concept
• The loading matrix X is then obtained using regularized least squares method
• Folding in method is used to efficiently compute ratings for a new user – by asking him
to rate k representative items
Liu, N.N., Meng, X., Liu, C. and Yang, Q. 2011. Wisdom of the Better Few: Cold Start Recommendation via Representative Based Rating Elicitation. Proceedings of the Fifth
ACM Conference on Recommender Systems (New York, NY, USA, 2011), 37–44.
17. Образец заголовка
Interactive Collaborative Filtering
(ICF)
Paper: Zhao et al., CIKM 2013
Main Idea
• Addresses cold start problem
• Probabilistic matrix factorization
Technique Recommendation-feed back loop
• Uses an objective function to select an item to recommend next:
– Maximum expected reward from target user (+ve feedback)
• Two methods of items selection –
– Sampling
– Confidence Bound
Zhao,X., Zhang, W. and Wang, J. 2013. Interactive collaborative filtering. Proceedings of the 22nd ACM international conference on Conference on information & knowledge
management (New York, NY, USA, 2013), 1411–14.
18. Образец заголовкаICF (contd.)
• Sampling
– user & feature vectors : these are random variables
• updated using Thompson sampling
– user vector is more sensitive to new ratings
• Confidence Bound
– fixed feature vectors
– upper bound confidence (item with the highest upper bound is
chosen) & SVD (singular vector decomposition) are used to estimate
function
Results:
• Proposed algorithms outperformed the baselines including greedy
algorithm, active learning and interview approaches
• Also, more exploration is needed when longer term satisfaction is
targeted
19. Образец заголовкаChoice Based Preference Elicitation
Paper:Loepp et al., CHI 2014
Main Idea
• Interactive user control with automatic recommender techniques
• Extracts latent factors from a matrix of user ratings and generates dialogs
in which the user iteratively chooses between two sets of sample items
Technique
• Uses user-item matrix and derive latent item features from it.
• Latent factors are computed using Matrix Factorization techniques
• Movies are arranged in an n-dimensional vector space according to their
feature values
• Position of target user is incrementally updated to find items having
similar latent feature characteristics
Loepp, B., Hussein, T. and Ziegler, J. 2014. Choice-based Preference Elicitation for Collaborative Filtering Recommender Systems. Proceedings of the SIGCHI Conference on
Human Factors in Computing Systems (New York, NY, USA,2014), 3085–3094.
20. Образец заголовка
Choice Based Preference Elicitation
(contd.)
• In each step:
– two sets of movies are
shown (dialog)
– user vector (features) is
updated
– movies with shortest
distance from user vector are
shown Two movie sets that differ in a single factor [7]
Results
• This approach gave significantly better results than the other methods in
15 (out of 24) parameter comparisons
• It gives good result even for the cold start problem
Loepp, B., Hussein, T. and Ziegler, J. 2014. Choice-based Preference Elicitation for Collaborative Filtering Recommender Systems. Proceedings of the SIGCHI Conference on
Human Factors in Computing Systems (New York, NY, USA,2014), 3085–3094.
23. Образец заголовка
Group of Items for Preference
Elicitation (contd.)
• Pseudo rating profile of user is created
– discover users having highest rating
for movies in cluster
– calculate average rating
• Ratings are generated based on pseudo profile
Results
• Only top 10 recommendations were considered
• Baseline is rating 15 movies
• Time taken by users is less than half of
the median time taken by the baselining process
• Users are more satisfied from top N
recommendations.
Survey results about recommendation quality, comparing group-
based process with the base-
line rate 15 process.
Chang, S., Harper, F.M. and Terveen, L. 2015. Using Groups of Items for Preference Elicitation in Recommender Systems. Proceedings of the 18th ACM Conference on
Computer Supported Cooperative Work & Social Computing (New York,NY, USA,2015), 1258–1269
30. Образец заголовкаReferences
1. Golbandi, N., Koren, Y. and Lempel, R. 2010. On Bootstrapping Recommender Systems. Proceedings of the 19th ACM
International Conference on Information and KnowledgeManagement(New York, NY, USA, 2010),1805–1808.
2. Golbandi, N., Koren, Y. and Lempel, R. 2011. Adaptive Bootstrapping of Recommender Systems Using Decision Trees.
Proceedings of the Fourth ACM International Conference on Web Search and Data Mining (New York, NY, USA, 2011), 595–
604.
3. Zhou, K., Yang, S.-H. and Zha, H. 2011. Functional Matrix Factorizations for Cold-start Recommendation. Proceedings of the
34th International ACM SIGIR Conference on Research and Development in Information Retrieval (New York, NY, USA, 2011),
315–324.
4. Liu, N.N., Meng, X., Liu, C. and Yang, Q. 2011. Wisdom of the Better Few: Cold Start Recommendation via Representative
Based Rating Elicitation. Proceedings of the Fifth ACM Conference on Recommender Systems (New York, NY, USA, 2011), 37–
44.
5. Sun, M., Li, F., Lee, J., Zhou, K., Lebanon, G. and Zha, H. 2013. Learning Multiple-question Decision Trees for Cold-start
Recommendation. Proceedings of the Sixth ACM International Conference on Web Search and Data Mining (New York, NY,
USA, 2013),445–454.
6. Zhao, X., Zhang, W. and Wang, J. 2013. Interactive collaborative filtering. Proceedings of the 22nd ACM international
conference on Conference on information & knowledgemanagement(New York, NY, USA, 2013),1411–14.
7. Loepp, B., Hussein, T. and Ziegler, J. 2014. Choice-based Preference Elicitation for Collaborative Filtering Recommender
Systems. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (New York, NY, USA, 2014), 3085–
3094.
8. Chang, S., Harper, F.M. and Terveen, L. 2015. Using Groups of Items for Preference Elicitation in Recommender Systems.
Proceedings of the 18th ACM Conference on Computer Supported Cooperative Work & Social Computing (New York, NY, USA,
2015),1258–1269
9. Blédaité, L. and Ricci, F. 2015. Pairwise Preferences Elicitation and Exploitation for Conversational Collaborative Filtering.
Proceedings of the 26th ACMConference on Hypertext & Social Media (New York, NY, USA, 2015),231–236.
10. Neidhardt, J., Schuster, R., Seyfang, L. and Werthner, H. 2014. Eliciting the Users’ Unknown Preferences. Proceedings of the
8th ACMConference on Recommender Systems (New York, NY, USA, 2014),309–312.