Active Learning in Collaborative Filtering Recommender Systems : a Survey
- 1. Active Learning in Collaborative
Filtering RSs: a Survey
Mehdi Elahi
Francesco Ricci
Neil Rubens
August 2014
Munich, Germany
1
Corresponding journal article:
Elahi, Mehdi, Francesco Ricci, and Neil Rubens. "A survey of active learning in
collaborative filtering recommender systems." Computer Science Review (2016).
- 3. Introduction
¤ Recommender Systems are tools that support users
decision making by suggesting products that are
interesting to them.
¤ Collaborative Filtering: A technique used to predict
unknown ratings exploiting ratings given by users to
items.
3
3
4
2
5
3
?
- 4. Cold Start Problem
¤ New User Problem: when
a new user has no rating
it is impossible to predict
her ratings.
4
3
4
2
5
?
?
?
3
?
2
5
?
3
?
¤ New item problem: when
a new item is added to
the catalogue and none
has rated this item it will
never be recommended.
- 5. Active Learning for Collaborative
Filtering
¤ Active Learning:
¤ Requests and try to collect more ratings from the
users before offering recommendations.
5
- 6. Which Items should be chosen?
¤ Not all the ratings are equally useful, i.e.,
equally bring information to the system.
¤ To minimize the user rating effort only some
of them should be requested and acquired.
6
- 8. Definition of AL Strategy
¤ An active learning strategy for a Collaborative
Filtering is a set of rules to choose the best items
for the users to rate.
8
- 9. How an AL Strategy works
Item Score
1 151
2 44
3 7
4 1
5 42
6 34
7 9
8 55
9 20
… …
N 12
System computes the
scores for all the
items that can be
scored (according to
a strategy)
9
- 10. How an AL Strategy works
Top 10
items
Score
1 151
8 55
43 54
11 50
2 44
5 42
6 34
22 33
75 29
13 25
The system selects
the top 10 items
and presents them
to the simulated
user
10
- 11. How an AL Strategy works
The items that are
rated are added to
the train set
Rated
items
1
2
5
75
13
11
- 12. Classifying AL Strategies
A. Personalization: addresses the what extent the
personalization is performed when selecting the list of
candidate items for the users to rate
¤ Two Classes of Strategies:
12
Non-personalized: are
those that ask all the
users to rate the same list
of items
Personalized: ask different
users to rate different
items – the best for each
user.
- 13. Classifying AL Strategies
13
Personalization Dimension
Corresponding journal article:
Elahi, Mehdi, Francesco Ricci, and Neil Rubens. "A survey
of active learning in collaborative filtering recommender
systems." Computer Science Review (2016).
- 14. Classifying AL Strategies
B. Hybridization: whether the strategy takes into
account a single heuristic (criterion) for selecting the
items or combines several heuristics
¤ Two Classes of Strategies:
14
Single-heuristic: are
those that implement a
unique item selection
rule.
Combined-heuristic strategies
hybridize single-heuristic
strategies by aggregating
and combining a number of
strategies.
- 15. Classifying AL Strategies
15
Hybridization Dimension
Corresponding journal article:
Elahi, Mehdi, Francesco Ricci, and Neil Rubens. "A survey
of active learning in collaborative filtering recommender
systems." Computer Science Review (2016).
- 18. Example of Non-Personalized AL
¤ Single Heuristics:
¤ Popularity: scores an item according to the frequency of
its ratings and then chooses the highest scored items
(Carenini, 2003)
¤ Entropy: scores each item with the entropy of its ratings
and then chooses the highest scored items (Rashid, 2002
and 2008)
¤ Combined Heuristics:
¤ log(Popularity)*Entropy: combines the popularity and
entropy scores and then chooses the highest scored
items (Rashid, 2002 and 2008)
18
- 21. Example of Personalized AL
¤ Single Heuristics:
¤ Decision Tree Based: uses a decision tree whose nodes,
represents groups of users. Each node divides the users into
three groups based on their ratings: Lovers, Haters, and
Unknowns. Starting from the root node, a new user is proposed
to rate a sequence of items, until she reaches one of the leaf
nodes (Golbandi, 2011)
¤ Binary Prediction: scores an item according to the prediction
of its ratings (using transformed matrix of user-item) and then
chooses the highest scored items (Elahi, 2011)
¤ Combined Heuristics:
¤ Combined with Voting: scores an item according to the
votes given by a committee of different strategies and
then chooses the highest scored items (Elahi, 2011)
21
- 22. Pros and Cons
22
+ simple, fast, no training, serves users
with no rating, good for early stage
- less accurate, same items for all users
+ fast, benefits of multiple strategies
- flaws of multiple strategies, difficulty of
combining properly
+ accurate, different items for different
users, higher prob. of collecting ratings,
good for late stage
- complex, slow, needs training, cannot
serve users with no rating
+ accurate, great adaptivity to
condition of the system
- more complex, slowest
- 23. Conclusion
¤ We provided a comprehensive review of the state-of-
the-art on active learning in collaborative filtering
recommender systems
¤ We have classified a wide range of active learning
techniques, called Strategies, along the two
dimensions:
¤ how personalized these techniques are
¤ how many different item selection criteria (heuristics)
are considered by these strategies in their rating
elicitation process.
23
- 24. Future Works
2424
¤ To survey works that have been done in AL
for other types of recommender systems,
such as content-based and context-aware.
¤ To analyze active learning techniques based
on their applicability to specific application
domains.