This document proposes a new type of recommender system called a context recommender that recommends appropriate contexts (e.g. time, location, companion) for users to consume items. It discusses how context recommenders are different than traditional and context-aware recommenders. It also presents the framework for context recommenders including algorithms using multi-label classification to directly predict contexts. The document reports on experiments comparing these algorithms on several datasets and finds that personalized algorithms outperform non-personalized ones and that certain multi-label classification algorithms like label powerset using support vector machines achieve the best performance.
Processing & Properties of Floor and Wall Tiles.pptx
Context Rec Recommendation Using Multi-Label Classification
1. Context Recommendation
Using Multi-label Classification
Yong Zheng, Bamshad Mobasher, Robin Burke
Center for Web Intelligence, DePaul University, Chicago
IEEE/WIC/ACM Conference on Web Intelligence
Aug 14, Warsaw, Poland
2. Intro – Recommender Systems
• Information Overload Problem IR and RS
• Recommender systems (RS) are the systems being
able to provide recommendations to the end users.
Currently, RS are popular everywhere:
• E-Commerce: Amazon, Ebay, Newegg, etc
• Social networks: Twitter, Facebook, etc
• Movie/Stream: Netflix, Movie Pilot, Youtube, etc
• Music: Pandora, Last.FM, etc
3. Intro – Type of Recommendations
• [Item Recommendations]
• [User Recommendations]
4. Intro – Context-aware Recommender
• Context-aware Recommender Systems (CARS)
CARS is a new type of RS which provide recommendations
by adapting to users’ contextual situations.
• Traditional RS: Users × Items Ratings
• Contextual RS: Users × Items × Contexts Ratings
• Assumptions behind:
1). User may have different preferences in different contexts;
2). Contexts are important in decision-makings.
However, what CARS recommended are still items or users.
Companion
5. New Application: Context Recommender
• In this paper, we propose a new application: context
recommender (CR), which is able to recommend or
suggest appropriate contexts for users to select or
consume items.
• Sample of contexts: time, location, companion, etc
Context Rec
User RecItem Rec
8. Related Work
L. Baltrunas, et al. "Best usage context predictions for
music tracks", The 2nd Workshop on Context-aware
Recommender Systems, ACM RecSys, 2010
This is the only work related to context recommendation,
where the authors tried to provide suggestions of
appropriate contexts to listen to music tracks.
Pros: they proposed three KNN-based classifiers to
suggest appropriate contexts;
Cons: they proposed a specific application, but not a
general delineation of the problem of context
recommendations.
9. Contributions: Context Recommender
• We formally provide the definition of CR
• We propose the formal framework of CR applications
• We discuss the algorithmic paradigms for CR
• We examine the algorithms using multi-label classifications
10. Context Recommender
1. Definitions
Context recommenders are the systems being able to
recommend or suggest appropriate contexts for users to
select or consume items.
Examples:
When is the best season to travel to Poland for user Tom?
Who is the suggested companion to see “Titanic” with Tom?
This book is better to be gifted to Mom or Kids?
2. Research Problems
How to infer the appropriate contexts?
And those contexts should be personalized or not?
Example:
The best season to Poland is always the same for all the
users, or it could be personalized for different specific
users? E.g. Tom likes winter much more than summer,
where most users prefer summers in Poland.
11. Context Recommender
3. Context Recommendations App
• There could be many other types, for example, it could
be a group of users or items, instead of a single user or
item. E.g. what is the best season to Poland by this
group of travelers (e.g. Tour Group, etc), where the
suggested contexts should meet the requirement of the
group of users, instead of a single user.
• In this paper, we focus on the general form:
{User, Item} Contexts; a pair of <user, item> as input
Input Output App
{User, Item} Contexts The best season to Poland for Tom
{User} Contexts The best travel season for Tom
{Item} Contexts The best travel season to Poland
12. Context Recommender
3. Algorithmic Paradigms
What are the possible algorithms to recommend contexts?
After analysis, we propose two series of frameworks:
1). Direct Context Prediction
We infer suggested contexts by <user, item, preferences>;
In other words, contexts are predicted based on users’
previous preference histories associated with <item,
contexts>
2). Indirect Context Recommendation
We reuse the context-aware recommendation algorithms:
In CARS,
Therefore, we vary the choices of contexts, and finally
recommend the contexts which can contribute to the best
ratings user will give to the item.
13. Context Recommender
3. Algorithmic Paradigms
Which algorithms can be applied in each category?
1). Direct Context Prediction
Classification algorithms are the popular ones which are
applied to this category, where they have been adopted in
context predictions in the pervasive computing area.
2). Indirect Context Recommendation
We reuse the context-aware recommendation algorithms;
Therefore, all available CARS algorithms can be applied to.
However, the drawback of this category is the high
computational cost if there are too many contextual
conditions in the data. And it also relies on how the CARS
algorithms perform.
In this paper, we focus on Direct Context Prediction using
multi-label classification algorithms.
14. Context Recommender
4. Why Multi-label Classifications (MLC)?
1).Binary Classification (is this an apple?)
2).Multi-class Classification (is it an apple|orange|pear?)
3).Multi-label Classification (<round, apple, fruit, Mac>)
In other words, MLC allows the system to
select more than 1 labels from the set.
Classification is used for context predictions;
and MLC just fits the requirement of context
recommendation task.
15. Context Recommender
4. Why Multi-label Classifications (MLC)?
Two series of MLC algorithms:
1). Transformation Algorithms
They can use traditional classification algorithms (e.g.
decision trees, SVM, etc), and they transform the MLC task
to multiple binary or multi-class classification tasks. So they
do not need to develop new algorithms.
E.g., binary relevance (BR), label powerset (LP), classifier
chains (CC), k-labelsets (RAkEL)
2). Adaptation Algorithms
Develop new classification algorithms to adapt to the MLC
task. E.g., Binary relevance KNN (BRKNN), Multi-label KNN
(MLKNN)
16. Context Recommender
5. Experiments and Evaluations (Algorithms)
Toolkit: Mulan (MLC toolkit) and Weka Java-based library
MLC algorithms: BR, LP, CC, RAkEL, MLKNN, BRKNN;
Classification methods used in BR, LP, CC and RAkEL:
KNNclassifier (KNN), decision trees (J48), naive bayes
(NB),Bayesian nets (BN) and support vector machine (SMO)
Baseline algorithms:
1).the three KNN classifiers: RatingBased (RB),
BestContextVectorBased (BCVB) and BestContextBased
(BCB) proposed by L. Baltrunas, et al
2).Non-personalized methods: such as most popular algs
Data # of users # of items # of ratings # of labels Rating scale
AdomMovie 69 176 1,010 8 1-13
LDOS 113 1186 2,094 17 1-5
TripAdvisor 2731 2269 14,175 5 1-5
17. Context Recommender
5. Experiments and Evaluations (Metrics)
Inputs: User, Item, Binary Preference (Good or Bad)
Outputs: A list of predicted contexts
Time =
Weekend
Time =
Weekday
Companion
= Kids
Companion
= Parents
Companion
= Girlfriend
Real 0 1 0 0 1
Prediction 1 0 0 0 1
Y is the set of TRUE labels in the ground truth, and Z is
the set of predicted TRUE labels. m = # of examples.
Another metric is hamming loss which measures the
average percentage of incorrectly predicted labels.
19. Context Recommender
5. Experimental Results (findings)
1).Personalization is required, because personalized
algorithms work much better than the non-personalized
ones. i.e., simply recommending most popular ones (e.g.
most popular season people visiting Poland) is not enough;
2).LP algorithm using SMO as classifier is the best MLC
algorithms among all data sets and all methods examined.
They beat all the other algorithms. KNN-based approaches
worked bad, because context-aware data are usually sparse.
3).SMO is the best classifier used for MLC algorithms, but it
increases computation costs if data is large and there are
many contextual labels. The alternative choice is Bayesian
Nets which worked good and not time-consuming
4). About running performance: LP using SMO is the best
choice, but both LP and SMO increase computation costs if
data is large and there are many contextual labels.
Solutions:
a). Reduce the number of labels by pre-selections;
b). Choose LP using Bayesian Nets
20. Conclusions and Future Work
• We formally introduce and discuss the application and
research problem of context recommendations (CR). We
believe that context recommenders will provide many more
novel applications and new recommendation opportunities
for both practical use and the research community.
• We propose the formal framework of CR applications and
discuss the algorithmic paradigms for CR.
• We examine the algorithms using multi-label classifications,
and infer some significant findings and patterns as
introduced previously.
• Future work: Examine more other algorithms, and develop
new evaluation metrics for this domain.
21. References
• Context Recommendations
[1].Baltrunas, Linas, Marius Kaminskas, Francesco Ricci, Lior
Rokach, Bracha Shapira, and Karl-Heinz Luke. "Best usage
context prediction for music tracks." In Proceedings of the 2nd
Workshop on Context Aware Recommender Systems. 2010.
[2].Yong Zheng, Bamshad Mobasher, Robin Burke. "Context
Recommendation Using Multi-label Classification". In Proceedings
of the 13th IEEE/WIC/ACM International Conference on Web
Intelligence, 2014
• Multi-label Classifications
[1].Tsoumakas, Grigorios, and Ioannis Katakis. "Multi-label
classification: An overview." International Journal of Data
Warehousing and Mining (IJDWM) 3, no. 3 (2007): 1-13.
[2].Tsoumakas, Grigorios, Ioannis Katakis, and Ioannis Vlahavas.
"Mining multi-label data." In Data mining and knowledge
discovery handbook, pp. 667-685. Springer US, 2010.
22. Thanks!
Yong Zheng, Bamshad Mobasher, Robin Burke
Center for Web Intelligence, DePaul University, Chicago
IEEE/WIC/ACM Conference on Web Intelligence
Aug 14, Warsaw, Poland