pareto-efficient hybridization for multi-objective recommender systems
DESCRIPTION
Presented at RecSys2012, Dublin. Any questions or comments, email me at [email protected]TRANSCRIPT
Pareto-Efficient Hybridization forMulti-Objective Recommender
Systems
Marco Tulio Ribeiro 1,2 Anisio Lacerda 1,2
Adriano Veloso 1 Nivio Ziviani 1,2
1Universidade Federal de Minas Gerais 2Zunnit TechnologiesComputer Science Department Belo Horizonte, Brazil
Belo Horizonte, Brazil
ACM Recommender Systems 2012, Dublin, IrelandSeptember 10th, 2012
1
Pareto Efficient Hybridization forMulti-Objective Recommender Systems
I Multi-Objective:
I AccuracyI NoveltyI Diversity
I Hybridization:I Different algorithms have different strengths
I Pareto Efficient:I In a moment
2
Pareto Efficient Hybridization forMulti-Objective Recommender Systems
I Multi-Objective:
I AccuracyI NoveltyI Diversity
I Hybridization:I Different algorithms have different strengths
I Pareto Efficient:I In a moment
2
Pareto Efficient Hybridization forMulti-Objective Recommender Systems
I Multi-Objective:I AccuracyI NoveltyI Diversity
I Hybridization:I Different algorithms have different strengths
I Pareto Efficient:I In a moment
2
Pareto Efficient Hybridization forMulti-Objective Recommender Systems
I Multi-Objective:I AccuracyI NoveltyI Diversity
I Hybridization:I Different algorithms have different strengths
I Pareto Efficient:I In a moment
2
Pareto Efficient Hybridization forMulti-Objective Recommender Systems
I Multi-Objective:I AccuracyI NoveltyI Diversity
I Hybridization:I Different algorithms have different strengths
I Pareto Efficient:I In a moment
2
What’s a Good Recommendation?
I “Good” is a multifaceted concept
I Are novel recommendations goodrecommendations?
I Are accurate recommendations goodrecommendations?
I Are diverse recommendations goodrecommendations?
3
What’s a Good Recommendation?
I “Good” is a multifaceted conceptI Are novel recommendations good
recommendations?
I Are accurate recommendations goodrecommendations?
I Are diverse recommendations goodrecommendations?
3
Is Novelty Good?
I “Good” is a multifaceted conceptI Are novel recommendations good
recommendations?I Are accurate recommendations good
recommendations?I Are diverse recommendations good
recommendations?
3
Is Novelty Good?
I “Good” is a multifaceted conceptI Are novel recommendations good
recommendations?I Are accurate recommendations good
recommendations?I Are diverse recommendations good
recommendations?
3
What’s a Good Recommendation?
I “Good” is a multifaceted conceptI Are novel recommendations good
recommendations?I Are accurate recommendations good
recommendations?
I Are diverse recommendations goodrecommendations?
3
Is Accuracy Good?
I “Good” is a multifaceted conceptI Are novel recommendations good
recommendations?I Are accurate recommendations good
recommendations?I Are diverse recommendations good
recommendations?
3
Is Accuracy Good?
I “Good” is a multifaceted conceptI Are novel recommendations good
recommendations?I Are accurate recommendations good
recommendations?I Are diverse recommendations good
recommendations?
3
What’s a Good Recommendation?
I “Good” is a multifaceted conceptI Are novel recommendations good
recommendations?I Are accurate recommendations good
recommendations?I Are diverse recommendations good
recommendations?
3
Is Diversity Good?
I “Good” is a multifaceted conceptI Are novel recommendations good
recommendations?I Are accurate recommendations good
recommendations?I Are diverse recommendations good
recommendations?
3
Is Diversity Good?
I “Good” is a multifaceted conceptI Are novel recommendations good
recommendations?I Are accurate recommendations good
recommendations?I Are diverse recommendations good
recommendations?
3
Our WorkI The challenge:
I Combining multiple algorithms
I Contributions:I Domain and algorithm-independent hybridI Multi-objective in terms of accuracy, novelty
and diversity.I Adjustable compromise
4
Our WorkI The challenge:
I Combining multiple algorithmsI Contributions:
I Domain and algorithm-independent hybrid
I Multi-objective in terms of accuracy, noveltyand diversity.
I Adjustable compromise
4
Our WorkI The challenge:
I Combining multiple algorithmsI Contributions:
I Domain and algorithm-independent hybridI Multi-objective in terms of accuracy, novelty
and diversity.
I Adjustable compromise
4
Our WorkI The challenge:
I Combining multiple algorithmsI Contributions:
I Domain and algorithm-independent hybridI Multi-objective in terms of accuracy, novelty
and diversity.I Adjustable compromise
4
Weighted Aggregation
I Combine the algorithms using standardweighted aggregation
I Problem: finding the vector of weights W
I Example:W = [SVD: 2.3,TopPop: −5, ItemKNN : 1]
I Easy to add or remove algorithms
5
Weighted Aggregation
I Combine the algorithms using standardweighted aggregation
I Problem: finding the vector of weights W
I Example:W = [SVD: 2.3,TopPop: −5, ItemKNN : 1]
I Easy to add or remove algorithms
5
Weighted Aggregation
I Combine the algorithms using standardweighted aggregation
I Problem: finding the vector of weights W
I Example:W = [SVD: 2.3,TopPop: −5, ItemKNN : 1]
I Easy to add or remove algorithms
5
Weighted Aggregation
I Combine the algorithms using standardweighted aggregation
I Problem: finding the vector of weights W
I Example:W = [SVD: 2.3,TopPop: −5, ItemKNN : 1]
I Easy to add or remove algorithms
5
Evolutionary Algorithms
I A population is created with a group ofrandom individuals
I For each generation:I The individuals of the population are
evaluated (cross validation)I The best individuals are combined, mutated
or kept
I Good for search spaces where little isknown
I Domain and algorithm-independent
6
Evolutionary Algorithms
I A population is created with a group ofrandom individuals
I For each generation:I The individuals of the population are
evaluated (cross validation)I The best individuals are combined, mutated
or kept
I Good for search spaces where little isknown
I Domain and algorithm-independent
6
Evolutionary Algorithms
I A population is created with a group ofrandom individuals
I For each generation:I The individuals of the population are
evaluated (cross validation)I The best individuals are combined, mutated
or kept
I Good for search spaces where little isknown
I Domain and algorithm-independent
6
Evolutionary Algorithms
I A population is created with a group ofrandom individuals
I For each generation:I The individuals of the population are
evaluated (cross validation)I The best individuals are combined, mutated
or kept
I Good for search spaces where little isknown
I Domain and algorithm-independent
6
SPEA2
I Strength Pareto Evolutionary Algorithm[Zitzler, Laumanns and Thiele]
I Multi-Objective Evolutionary AlgorithmI Uses the Pareto Dominance conceptI Returns a Pareto FrontierI O(M 2logM), but performed offline
7
SPEA2
I Strength Pareto Evolutionary Algorithm[Zitzler, Laumanns and Thiele]
I Multi-Objective Evolutionary Algorithm
I Uses the Pareto Dominance conceptI Returns a Pareto FrontierI O(M 2logM), but performed offline
7
SPEA2
I Strength Pareto Evolutionary Algorithm[Zitzler, Laumanns and Thiele]
I Multi-Objective Evolutionary AlgorithmI Uses the Pareto Dominance conceptI Returns a Pareto FrontierI O(M 2logM), but performed offline
7
SPEA2
I Strength Pareto Evolutionary Algorithm[Zitzler, Laumanns and Thiele]
I Multi-Objective Evolutionary AlgorithmI Uses the Pareto Dominance concept
I Returns a Pareto FrontierI O(M 2logM), but performed offline
7
Pareto Dominance
I Strength Pareto Evolutionary Algorithm[Zitzler, Laumanns and Thiele]
I Multi-Objective Evolutionary AlgorithmI Uses the Pareto Dominance conceptI Returns a Pareto FrontierI O(M 2logM), but performed offline
7
Pareto Dominance
I Strength Pareto Evolutionary Algorithm[Zitzler, Laumanns and Thiele]
I Multi-Objective Evolutionary AlgorithmI Uses the Pareto Dominance conceptI Returns a Pareto FrontierI O(M 2logM), but performed offline
7
Pareto Dominance
I Strength Pareto Evolutionary Algorithm[Zitzler, Laumanns and Thiele]
I Multi-Objective Evolutionary AlgorithmI Uses the Pareto Dominance conceptI Returns a Pareto FrontierI O(M 2logM), but performed offline
7
Pareto Dominance
I Strength Pareto Evolutionary Algorithm[Zitzler, Laumanns and Thiele]
I Multi-Objective Evolutionary AlgorithmI Uses the Pareto Dominance conceptI Returns a Pareto FrontierI O(M 2logM), but performed offline
7
SPEA2
I Strength Pareto Evolutionary Algorithm[Zitzler, Laumanns and Thiele]
I Multi-Objective Evolutionary AlgorithmI Uses the Pareto Dominance conceptI Returns a Pareto Frontier
I O(M 2logM), but performed offline
7
Pareto Frontier
I Strength Pareto Evolutionary Algorithm[Zitzler, Laumanns and Thiele]
I Multi-Objective Evolutionary AlgorithmI Uses the Pareto Dominance conceptI Returns a Pareto FrontierI O(M 2logM), but performed offline
7
SPEA2
I Strength Pareto Evolutionary Algorithm[Zitzler, Laumanns and Thiele]
I Multi-Objective Evolutionary AlgorithmI Uses the Pareto Dominance conceptI Returns a Pareto FrontierI O(M 2logM), but performed offline
7
Adjusting the System Priority
I The recommender system may desire toadjust the compromise
I We do not return a single solution, but thePareto Frontier
I Given the priority of each objective, wechoose one individual from the frontier
8
Adjusting the System Priority
I The recommender system may desire toadjust the compromise
I We do not return a single solution, but thePareto Frontier
I Given the priority of each objective, wechoose one individual from the frontier
8
Adjusting the System Priority
I The recommender system may desire toadjust the compromise
I We do not return a single solution, but thePareto Frontier
I Given the priority of each objective, wechoose one individual from the frontier
8
Adjusting the System Priority
I The recommender system may desire toadjust the compromise
I We do not return a single solution, but thePareto Frontier
I Given the priority of each objective, wechoose one individual from the frontier
8
Evaluation Methodology
I Task: Top-N Item Recommendation
I Evaluation methodology similar to[Cremonesi, Koren and Turrin, RecSys 2010]
I With novelty and diversity from[Vargas and Castells, RecSys 2011]
9
Evaluation Methodology
I Task: Top-N Item RecommendationI Evaluation methodology similar to
[Cremonesi, Koren and Turrin, RecSys 2010]
I With novelty and diversity from[Vargas and Castells, RecSys 2011]
9
Evaluation Methodology
I Task: Top-N Item RecommendationI Evaluation methodology similar to
[Cremonesi, Koren and Turrin, RecSys 2010]
I With novelty and diversity from[Vargas and Castells, RecSys 2011]
9
Datasets
Movielens Last.fm
Recommends movies music
Users 6,040 992
Content 3,883 movies 176,948 artists
Ratings/Feedback 1,000,209 19,150,868
Feedback explicit implicit
Table: Summary of Datasets
10
Recommendation Algorithms
I PureSVD (50 and 150 factors)[Cremonesi, Koren and Turrin, RecSys 2010]
I KNNs: Item and User-basedI Most PopularI WRMF
[Hu et al, ICDM 2008, Pan et al ICDM 2008]
I Content-based:I Item Attribute KNN (movielens only)I User Attribute KNN
11
Recommendation Algorithms
I PureSVD (50 and 150 factors)[Cremonesi, Koren and Turrin, RecSys 2010]
I KNNs: Item and User-based
I Most PopularI WRMF
[Hu et al, ICDM 2008, Pan et al ICDM 2008]
I Content-based:I Item Attribute KNN (movielens only)I User Attribute KNN
11
Recommendation Algorithms
I PureSVD (50 and 150 factors)[Cremonesi, Koren and Turrin, RecSys 2010]
I KNNs: Item and User-basedI Most Popular
I WRMF[Hu et al, ICDM 2008, Pan et al ICDM 2008]
I Content-based:I Item Attribute KNN (movielens only)I User Attribute KNN
11
Recommendation Algorithms
I PureSVD (50 and 150 factors)[Cremonesi, Koren and Turrin, RecSys 2010]
I KNNs: Item and User-basedI Most PopularI WRMF
[Hu et al, ICDM 2008, Pan et al ICDM 2008]
I Content-based:I Item Attribute KNN (movielens only)I User Attribute KNN
11
Recommendation Algorithms
I PureSVD (50 and 150 factors)[Cremonesi, Koren and Turrin, RecSys 2010]
I KNNs: Item and User-basedI Most PopularI WRMF
[Hu et al, ICDM 2008, Pan et al ICDM 2008]
I Content-based:I Item Attribute KNN (movielens only)I User Attribute KNN
11
Hybrid Baselines
I Borda Count
I STREAM (stacking-based approach)[Bao, Bergman and Thompson, RecSys 2009]
I Weighted aggregation with equal weights
12
Hybrid Baselines
I Borda CountI STREAM (stacking-based approach)
[Bao, Bergman and Thompson, RecSys 2009]
I Weighted aggregation with equal weights
12
Hybrid Baselines
I Borda CountI STREAM (stacking-based approach)
[Bao, Bergman and Thompson, RecSys 2009]
I Weighted aggregation with equal weights
12
Some of Our Solutions - Movielens
I PO-acc:
I PO-acc2:
I PO-nov:
I PO-div:
13
14
15
16
Some of Our Solutions - Last.fm
I PO-acc:
I PO-nov:
I PO-div:
17
18
19
20
Conclusions
I A multi-objective hybridization technique forcombining recommendation algorithms
I “Tune” the system to different priority needsI Highly reproducible experiments:
I Public datasetsI Open-source implementations
(MyMediaLite, DEAP)
I Competitive with the best algorithmsaccording to each objective
21
Conclusions
I A multi-objective hybridization technique forcombining recommendation algorithms
I “Tune” the system to different priority needs
I Highly reproducible experiments:I Public datasetsI Open-source implementations
(MyMediaLite, DEAP)
I Competitive with the best algorithmsaccording to each objective
21
Conclusions
I A multi-objective hybridization technique forcombining recommendation algorithms
I “Tune” the system to different priority needsI Highly reproducible experiments:
I Public datasetsI Open-source implementations
(MyMediaLite, DEAP)
I Competitive with the best algorithmsaccording to each objective
21
Conclusions
I A multi-objective hybridization technique forcombining recommendation algorithms
I “Tune” the system to different priority needsI Highly reproducible experiments:
I Public datasetsI Open-source implementations
(MyMediaLite, DEAP)
I Competitive with the best algorithmsaccording to each objective
21
Future Work
I Test these assumptions using onlineAB-testing, in real world E-commercewebsites
I Try maximizing other objectives:I profit, stock diversity, etc
I Figuring out how often the weights need tobe re-adjusted
22
Future Work
I Test these assumptions using onlineAB-testing, in real world E-commercewebsites
I Try maximizing other objectives:I profit, stock diversity, etc
I Figuring out how often the weights need tobe re-adjusted
22
Future Work
I Test these assumptions using onlineAB-testing, in real world E-commercewebsites
I Try maximizing other objectives:I profit, stock diversity, etc
I Figuring out how often the weights need tobe re-adjusted
22
Pareto-Efficient Hybridization forMulti-Objective Recommender
Systems
Marco Tulio Ribeiro 1,2 Anisio Lacerda 1,2
Adriano Veloso 1 Nivio Ziviani 1,2
1Universidade Federal de Minas Gerais 2Zunnit TechnologiesComputer Science Department Belo Horizonte, Brazil
Belo Horizonte, Brazil
ACM Recommender Systems 2012, Dublin, IrelandSeptember 10th, 2012
23