Pareto-Efficient Hybridization for Multi-Objective Recommender Systems

1,770 views

Published on

Presented at RecSys2012, Dublin. Any questions or comments, email me at marcotcr@gmail.com

Published in: Technology

Pareto-Efficient Hybridization for Multi-Objective Recommender Systems

  1. 1. Pareto-Efficient Hybridization for Multi-Objective Recommender Systems Marco Tulio Ribeiro 1,2 Anisio Lacerda 1,2 Adriano Veloso 1 Nivio Ziviani 1,21 2 Universidade Federal de Minas Gerais Zunnit Technologies Computer Science Department Belo Horizonte, Brazil Belo Horizonte, Brazil ACM Recommender Systems 2012, Dublin, Ireland September 10th, 2012 1
  2. 2. Pareto Efficient Hybridization forMulti-Objective Recommender Systems 2
  3. 3. Pareto Efficient Hybridization forMulti-Objective Recommender Systems Multi-Objective: 2
  4. 4. Pareto Efficient Hybridization forMulti-Objective Recommender Systems Multi-Objective: Accuracy Novelty Diversity 2
  5. 5. Pareto Efficient Hybridization forMulti-Objective Recommender Systems Multi-Objective: Accuracy Novelty Diversity Hybridization: Different algorithms have different strengths 2
  6. 6. Pareto Efficient Hybridization forMulti-Objective Recommender Systems Multi-Objective: Accuracy Novelty Diversity Hybridization: Different algorithms have different strengths Pareto Efficient: In a moment 2
  7. 7. What’s a Good Recommendation? “Good” is a multifaceted concept 3
  8. 8. What’s a Good Recommendation? “Good” is a multifaceted concept Are novel recommendations good recommendations? 3
  9. 9. Is Novelty Good? 3
  10. 10. Is Novelty Good? 3
  11. 11. What’s a Good Recommendation? “Good” is a multifaceted concept Are novel recommendations good recommendations? Are accurate recommendations good recommendations? 3
  12. 12. Is Accuracy Good? 3
  13. 13. Is Accuracy Good? 3
  14. 14. What’s a Good Recommendation? “Good” is a multifaceted concept Are novel recommendations good recommendations? Are accurate recommendations good recommendations? Are diverse recommendations good recommendations? 3
  15. 15. Is Diversity Good? 3
  16. 16. Is Diversity Good? 3
  17. 17. Our Work The challenge: Combining multiple algorithms 4
  18. 18. Our Work The challenge: Combining multiple algorithms Contributions: Domain and algorithm-independent hybrid 4
  19. 19. Our Work The challenge: Combining multiple algorithms Contributions: Domain and algorithm-independent hybrid Multi-objective in terms of accuracy, novelty and diversity. 4
  20. 20. Our Work The challenge: Combining multiple algorithms Contributions: Domain and algorithm-independent hybrid Multi-objective in terms of accuracy, novelty and diversity. Adjustable compromise 4
  21. 21. Weighted Aggregation Combine the algorithms using standard weighted aggregation 5
  22. 22. Weighted Aggregation Combine the algorithms using standard weighted aggregation Problem: finding the vector of weights W 5
  23. 23. Weighted Aggregation Combine the algorithms using standard weighted aggregation Problem: finding the vector of weights W Example: W = [SVD: 2.3, TopPop: −5, ItemKNN : 1] 5
  24. 24. Weighted Aggregation Combine the algorithms using standard weighted aggregation Problem: finding the vector of weights W Example: W = [SVD: 2.3, TopPop: −5, ItemKNN : 1] Easy to add or remove algorithms 5
  25. 25. Evolutionary Algorithms A population is created with a group of random individuals 6
  26. 26. Evolutionary Algorithms A population is created with a group of random individuals For each generation: The individuals of the population are evaluated (cross validation) The best individuals are combined, mutated or kept 6
  27. 27. Evolutionary Algorithms A population is created with a group of random individuals For each generation: The individuals of the population are evaluated (cross validation) The best individuals are combined, mutated or kept Good for search spaces where little is known 6
  28. 28. Evolutionary Algorithms A population is created with a group of random individuals For each generation: The individuals of the population are evaluated (cross validation) The best individuals are combined, mutated or kept Good for search spaces where little is known Domain and algorithm-independent 6
  29. 29. SPEA2 Strength Pareto Evolutionary Algorithm [Zitzler, Laumanns and Thiele] 7
  30. 30. SPEA2 Strength Pareto Evolutionary Algorithm [Zitzler, Laumanns and Thiele] Multi-Objective Evolutionary Algorithm 7
  31. 31. SPEA2 7
  32. 32. SPEA2 Strength Pareto Evolutionary Algorithm [Zitzler, Laumanns and Thiele] Multi-Objective Evolutionary Algorithm Uses the Pareto Dominance concept 7
  33. 33. Pareto Dominance 7
  34. 34. Pareto Dominance 7
  35. 35. Pareto Dominance 7
  36. 36. Pareto Dominance 7
  37. 37. SPEA2 Strength Pareto Evolutionary Algorithm [Zitzler, Laumanns and Thiele] Multi-Objective Evolutionary Algorithm Uses the Pareto Dominance concept Returns a Pareto Frontier 7
  38. 38. Pareto Frontier 7
  39. 39. SPEA2 Strength Pareto Evolutionary Algorithm [Zitzler, Laumanns and Thiele] Multi-Objective Evolutionary Algorithm Uses the Pareto Dominance concept Returns a Pareto Frontier O(M 2logM ), but performed offline 7
  40. 40. Adjusting the System Priority The recommender system may desire to adjust the compromise 8
  41. 41. Adjusting the System Priority The recommender system may desire to adjust the compromise We do not return a single solution, but the Pareto Frontier 8
  42. 42. Adjusting the System Priority The recommender system may desire to adjust the compromise We do not return a single solution, but the Pareto Frontier Given the priority of each objective, we choose one individual from the frontier 8
  43. 43. Adjusting the System Priority 8
  44. 44. Evaluation Methodology Task: Top-N Item Recommendation 9
  45. 45. Evaluation Methodology Task: Top-N Item Recommendation Evaluation methodology similar to [Cremonesi, Koren and Turrin, RecSys 2010] 9
  46. 46. Evaluation Methodology Task: Top-N Item Recommendation Evaluation methodology similar to [Cremonesi, Koren and Turrin, RecSys 2010] With novelty and diversity from [Vargas and Castells, RecSys 2011] 9
  47. 47. Datasets Movielens Last.fmRecommends movies musicUsers 6,040 992Content 3,883 movies 176,948 artistsRatings/Feedback 1,000,209 19,150,868Feedback explicit implicit Table: Summary of Datasets 10
  48. 48. Recommendation Algorithms PureSVD (50 and 150 factors) [Cremonesi, Koren and Turrin, RecSys 2010] 11
  49. 49. Recommendation Algorithms PureSVD (50 and 150 factors) [Cremonesi, Koren and Turrin, RecSys 2010] KNNs: Item and User-based 11
  50. 50. Recommendation Algorithms PureSVD (50 and 150 factors) [Cremonesi, Koren and Turrin, RecSys 2010] KNNs: Item and User-based Most Popular 11
  51. 51. Recommendation Algorithms PureSVD (50 and 150 factors) [Cremonesi, Koren and Turrin, RecSys 2010] KNNs: Item and User-based Most Popular WRMF [Hu et al, ICDM 2008, Pan et al ICDM 2008] 11
  52. 52. Recommendation Algorithms PureSVD (50 and 150 factors) [Cremonesi, Koren and Turrin, RecSys 2010] KNNs: Item and User-based Most Popular WRMF [Hu et al, ICDM 2008, Pan et al ICDM 2008] Content-based: Item Attribute KNN (movielens only) User Attribute KNN 11
  53. 53. Hybrid Baselines Borda Count 12
  54. 54. Hybrid Baselines Borda Count STREAM (stacking-based approach) [Bao, Bergman and Thompson, RecSys 2009] 12
  55. 55. Hybrid Baselines Borda Count STREAM (stacking-based approach) [Bao, Bergman and Thompson, RecSys 2009] Weighted aggregation with equal weights 12
  56. 56. Some of Our Solutions - Movielens PO-acc: PO-acc2: PO-nov: PO-div: 13
  57. 57. 14
  58. 58. 15
  59. 59. 16
  60. 60. Some of Our Solutions - Last.fm PO-acc: PO-nov: PO-div: 17
  61. 61. 18
  62. 62. 19
  63. 63. 20
  64. 64. Conclusions A multi-objective hybridization technique for combining recommendation algorithms 21
  65. 65. Conclusions A multi-objective hybridization technique for combining recommendation algorithms “Tune” the system to different priority needs 21
  66. 66. Conclusions A multi-objective hybridization technique for combining recommendation algorithms “Tune” the system to different priority needs Highly reproducible experiments: Public datasets Open-source implementations (MyMediaLite, DEAP) 21
  67. 67. Conclusions A multi-objective hybridization technique for combining recommendation algorithms “Tune” the system to different priority needs Highly reproducible experiments: Public datasets Open-source implementations (MyMediaLite, DEAP) Competitive with the best algorithms according to each objective 21
  68. 68. Future Work Test these assumptions using online AB-testing, in real world E-commerce websites 22
  69. 69. Future Work Test these assumptions using online AB-testing, in real world E-commerce websites Try maximizing other objectives: profit, stock diversity, etc 22
  70. 70. Future Work Test these assumptions using online AB-testing, in real world E-commerce websites Try maximizing other objectives: profit, stock diversity, etc Figuring out how often the weights need to be re-adjusted 22
  71. 71. Pareto-Efficient Hybridization for Multi-Objective Recommender Systems Marco Tulio Ribeiro 1,2 Anisio Lacerda 1,2 Adriano Veloso 1 Nivio Ziviani 1,21 2 Universidade Federal de Minas Gerais Zunnit Technologies Computer Science Department Belo Horizonte, Brazil Belo Horizonte, Brazil ACM Recommender Systems 2012, Dublin, Ireland September 10th, 2012 23

×