Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Pulpix - Video Recommendation at Scale

1,074 views

Published on

Talk given by Thomas Belhalfaoui & Lucas Charrier, Pulpix, during the RecsysFR meetup on February 1st 2017.

Published in: Internet
  • Be the first to comment

  • Be the first to like this

Pulpix - Video Recommendation at Scale

  1. 1. Video Recommendation at Scale January 2017 RecSysFR Meetup
  2. 2. Lucas Charrier Software Engineer
  3. 3. Thomas Belhalfaoui Data Scientist
  4. 4. About Pulpix
  5. 5. About Pulpix AI startup based in New-York and Paris 100+ sites using our technology in the world $850,000 seed round raised in the US Accelerated in Silicon Valley by
  6. 6. About Pulpix AI startup based in New-York and Paris 100+ sites using our technology in the world $850,000 seed round raised in the US Accelerated in Silicon Valley by
  7. 7. About Pulpix 100+ sites using our technology in the world Accelerated in Silicon Valley by $850,000 seed round raised in the US AI startup based in New-York and Paris
  8. 8. About Pulpix 100+ sites using our technology in the world $850,000 seed round raised in the US AI startup based in New-York and Paris Accelerated in Silicon Valley by
  9. 9. What do we do at Pulpix?
  10. 10. Video recommendation
  11. 11. How?
  12. 12. People binge-watch videos on social platforms
  13. 13. Why?
  14. 14. UX AI There is one simple reason: they invest in UX and AI.
  15. 15. Media Website vs. Social Platform Media Website Media Website
  16. 16. So, what do we do at Pulpix?
  17. 17. 1 7
  18. 18. 1 8
  19. 19. 1 9
  20. 20. 2 0
  21. 21. 2 1
  22. 22. 2 2
  23. 23. 2 3
  24. 24. 2 4
  25. 25. 2 5
  26. 26. 26
  27. 27. 27
  28. 28. 28
  29. 29. 29
  30. 30. 30
  31. 31. 31
  32. 32. 32
  33. 33. 33
  34. 34. 34
  35. 35. 35
  36. 36. 36
  37. 37. Video to Video
  38. 38. Video to Video
  39. 39. Article to Video
  40. 40. in Video
  41. 41. Recommendation at Scale
  42. 42. Recommendation at Scale Key figures 10 million videos Less than 100 ms response time 10 million events a day More than one billion training events
  43. 43. Recommendation at Scale Key figures 10 million videos Less than 100 ms response time 10 million events a day More than one billion events
  44. 44. Less than 100 ms response time More than one billion training events 10 million events a day 10 million videos Recommendation at Scale Key figures
  45. 45. Less than 100 ms response time 10 million events a day 10 million videos More than one billion training events Recommendation at Scale Key figures
  46. 46. From R&D to production Iterate fast
  47. 47. From R&D to production Iterate fast Idea Prototype Retrospective evaluation
  48. 48. From R&D to production Iterate fast Idea Prototype Retrospective evaluation A/B test 24h to 10 days
  49. 49. Content-Based Engine First approach
  50. 50. Content-based Engine First approach Speech-to-text Metadata Keywords extraction Weighting
  51. 51. Content-based Engine First approach Speech-to-text Content ScoreMetadata Keywords extraction Weighting
  52. 52. Content-based Engine First approach Speech-to-text Content ScoreMetadata Keywords extraction Weighting Recency boost
  53. 53. Collaborative Filtering
  54. 54. Collaborative Filtering Matrix factorization Videos
  55. 55. Collaborative Filtering Matrix factorization Videos
  56. 56. Collaborative Filtering Matrix factorization Implicit rating
  57. 57. Example L’Equipe.fr
  58. 58. Example L’Equipe.fr Football Basket Tennis 54% 12% 26% User preferences
  59. 59. Collaborative Filtering How to put it into practice? • User-based recommendations - Known users only - Not contextual
  60. 60. Collaborative Filtering How to put it into practice? • User-based recommendations - Known users only - Not contextual • Video-based recommendations - For all users - Fully contextual
  61. 61. Hybrid Engine
  62. 62. Hybrid engine Key dimensions
  63. 63. Context Hybrid engine Key dimensions
  64. 64. PersonalizationContext Hybrid engine Key dimensions
  65. 65. Personalization RecencyContext Hybrid engine Key dimensions
  66. 66. Hybrid Engine Linear score combination Content Engine Collaborative Engine Context Recency Global score Scores Linear Model
  67. 67. Hybrid Engine Nonlinear embedding combination Content Engine Collaborative Engine Context Recency Scores Candidate videos
  68. 68. Hybrid Engine Nonlinear embedding combination Content Engine Collaborative Engine Context Recency Global score Scores Candidate videos Nonlinear Model Features
  69. 69. Scalable Data Science stack
  70. 70. Flexibility queuing Scalability sharding Fault tolerance replication High throughput replication Recommendation at Scale Requirements
  71. 71. Flexibility queuing Scalability sharding Fault tolerance replication High throughput replication Recommendation at Scale Requirements
  72. 72. Flexibility queuing Scalability sharding Fault tolerance replication High throughput replication Recommendation at Scale Requirements
  73. 73. Flexibility queuing Scalability sharding Fault tolerance replication High throughput replication Recommendation at Scale Requirements
  74. 74. Recommendation at Scale Our Data Science stack
  75. 75. Recommendation at Scale Our Data Science stack tracking
  76. 76. Recommendation at Scale Our Data Science stack tracking ML data processing
  77. 77. Recommendation at Scale Our Data Science stack tracking reco ML data processing
  78. 78. What’s next?
  79. 79. • Reinforcement Learning What’s next? Our current R&D
  80. 80. • Reinforcement Learning • Deep Learning: ○ Recommendation ○ Video recognition What’s next? Our current R&D
  81. 81. @pulpix denis.vilar@pulpix.com PARIS NEW-YORK sabry.otmani@pulpix.com 124 rue d’Aboukir 75002 Paris, France 584 Broadway New York 10012 NY, USA Pulpix Pulpix Inc. hello@pulpix.com +33 (0)6 66 15 02 42 +1 (415) 996 4453 www.pulpix.com Pulpix is recruiting! jobs@pulpix.com

×