Beyond Metadata for BBC iPlayer: an autoencoder-driven approach for embeddings generation in content similarity recommendations
Download as PPTX, PDF0 likes32 views
I built a machine learning pipeline that generates metadata item embeddings for the calculation of item similarity in the context of Content-To-Content similarity recommendations for BBC iPlayer programmes.
![8982 8982
Modelling &
Embeddings
Generation
Ni [4491, 2245, 1122, 561, 280, 140, ...]
N1 < 8982 N1 < 8982
N2 < N1
Hyperparameter
Tuning](https://image.slidesharecdn.com/presentation-250327102834-ff094585/85/Beyond-Metadata-for-BBC-iPlayer-an-autoencoder-driven-approach-for-embeddings-generation-in-content-similarity-recommendations-82-320.jpg)
![8982 8982
Modelling &
Embeddings
Generation
Ni [4491, 2245, 1122, 561, 280, 140, ...]
Hyperparameter
Tuning](https://image.slidesharecdn.com/presentation-250327102834-ff094585/85/Beyond-Metadata-for-BBC-iPlayer-an-autoencoder-driven-approach-for-embeddings-generation-in-content-similarity-recommendations-83-320.jpg)
![8982 8982
Modelling &
Embeddings
Generation
Ni [4491, 2245, 1122, 561, 280, 140, ...]
Hyperparameter
Tuning](https://image.slidesharecdn.com/presentation-250327102834-ff094585/85/Beyond-Metadata-for-BBC-iPlayer-an-autoencoder-driven-approach-for-embeddings-generation-in-content-similarity-recommendations-84-320.jpg)
![8982 8982
Modelling &
Embeddings
Generation
Ni [4491, 2245, 1122, 561, 280, 140, ...]
Hyperparameter
Tuning](https://image.slidesharecdn.com/presentation-250327102834-ff094585/85/Beyond-Metadata-for-BBC-iPlayer-an-autoencoder-driven-approach-for-embeddings-generation-in-content-similarity-recommendations-85-320.jpg)
![8982 8982
Modelling &
Embeddings
Generation
Ni [4491, 2245, 1122, 561, 280, 140, ...]
Hyperparameter
Tuning](https://image.slidesharecdn.com/presentation-250327102834-ff094585/85/Beyond-Metadata-for-BBC-iPlayer-an-autoencoder-driven-approach-for-embeddings-generation-in-content-similarity-recommendations-86-320.jpg)
![8982 8982
Modelling &
Embeddings
Generation
Ni [4491, 2245, 1122, 561, 280, 140, ...]
Hyperparameter
Tuning](https://image.slidesharecdn.com/presentation-250327102834-ff094585/85/Beyond-Metadata-for-BBC-iPlayer-an-autoencoder-driven-approach-for-embeddings-generation-in-content-similarity-recommendations-87-320.jpg)
![Modelling &
Embeddings
Generation
Ni [4491, 2245, 1122, 561, 280, 140, ...]
Hyperparameter
Tuning](https://image.slidesharecdn.com/presentation-250327102834-ff094585/85/Beyond-Metadata-for-BBC-iPlayer-an-autoencoder-driven-approach-for-embeddings-generation-in-content-similarity-recommendations-88-320.jpg)
![Modelling &
Embeddings
Generation
Ni [4491, 2245, 1122, 561, 280, 140, ...]
Hyperparameter
Tuning](https://image.slidesharecdn.com/presentation-250327102834-ff094585/85/Beyond-Metadata-for-BBC-iPlayer-an-autoencoder-driven-approach-for-embeddings-generation-in-content-similarity-recommendations-89-320.jpg)
![Modelling &
Embeddings
Generation
Ni [4491, 2245, 1122, 561, 280, 140, ...]
Hyperparameter
Tuning](https://image.slidesharecdn.com/presentation-250327102834-ff094585/85/Beyond-Metadata-for-BBC-iPlayer-an-autoencoder-driven-approach-for-embeddings-generation-in-content-similarity-recommendations-90-320.jpg)
![Modelling &
Embeddings
Generation
Ni [4491, 2245, 1122, 561, 280, 140, ...]
Hyperparameter
Tuning](https://image.slidesharecdn.com/presentation-250327102834-ff094585/85/Beyond-Metadata-for-BBC-iPlayer-an-autoencoder-driven-approach-for-embeddings-generation-in-content-similarity-recommendations-91-320.jpg)
![Modelling &
Embeddings
Generation
Ni [4491, 2245, 1122, 561, 280, 140, ...]
Hyperparameter
Tuning](https://image.slidesharecdn.com/presentation-250327102834-ff094585/85/Beyond-Metadata-for-BBC-iPlayer-an-autoencoder-driven-approach-for-embeddings-generation-in-content-similarity-recommendations-92-320.jpg)
![Modelling &
Embeddings
Generation
Ni [4491, 2245, 1122, 561, 280, 140, ...]
Hyperparameter
Tuning](https://image.slidesharecdn.com/presentation-250327102834-ff094585/85/Beyond-Metadata-for-BBC-iPlayer-an-autoencoder-driven-approach-for-embeddings-generation-in-content-similarity-recommendations-93-320.jpg)
![Modelling &
Embeddings
Generation
Ni [4491, 2245, 1122, 561, 280, 140, ...]
Hyperparameter
Tuning](https://image.slidesharecdn.com/presentation-250327102834-ff094585/85/Beyond-Metadata-for-BBC-iPlayer-an-autoencoder-driven-approach-for-embeddings-generation-in-content-similarity-recommendations-94-320.jpg)
![Modelling &
Embeddings
Generation
Ni [4491, 2245, 1122, 561, 280, 140, ...]
Hyperparameter
Tuning](https://image.slidesharecdn.com/presentation-250327102834-ff094585/85/Beyond-Metadata-for-BBC-iPlayer-an-autoencoder-driven-approach-for-embeddings-generation-in-content-similarity-recommendations-95-320.jpg)
![Modelling &
Embeddings
Generation
Ni [4491, 2245, 1122, 561, 280, 140, ...]
Hyperparameter
Tuning](https://image.slidesharecdn.com/presentation-250327102834-ff094585/85/Beyond-Metadata-for-BBC-iPlayer-an-autoencoder-driven-approach-for-embeddings-generation-in-content-similarity-recommendations-96-320.jpg)
![Modelling &
Embeddings
Generation
Ni [4491, 2245, 1122, 561, 280, 140, ...]
Hyperparameter
Tuning](https://image.slidesharecdn.com/presentation-250327102834-ff094585/85/Beyond-Metadata-for-BBC-iPlayer-an-autoencoder-driven-approach-for-embeddings-generation-in-content-similarity-recommendations-97-320.jpg)
![Modelling &
Embeddings
Generation
Ni [4491, 2245, 1122, 561, 280, 140, ...]
Hyperparameter
Tuning](https://image.slidesharecdn.com/presentation-250327102834-ff094585/85/Beyond-Metadata-for-BBC-iPlayer-an-autoencoder-driven-approach-for-embeddings-generation-in-content-similarity-recommendations-98-320.jpg)
![[ろれ雑ろ]Session 2 2. Future of Team Collaboration](https://cdn.slidesharecdn.com/ss_thumbnails/session2-2-161124020920-thumbnail.jpg?width=560&fit=bounds)
More Related Content
Similar to Beyond Metadata for BBC iPlayer: an autoencoder-driven approach for embeddings generation in content similarity recommendations (20)
More from Simone Spaccarotella (8)
Recently uploaded (20)
![PLAN_OF_WORK_PPT_BY_ROHIT_BHAIRAM_--2212020201003[1] new.pptx](https://cdn.slidesharecdn.com/ss_thumbnails/planofworkpptbyrohitbhairam-22120202010031new-250329072214-99e4fccd-thumbnail.jpg?width=560&fit=bounds)
Beyond Metadata for BBC iPlayer: an autoencoder-driven approach for embeddings generation in content similarity recommendations
- 1. Beyond Metadata for BBC iPlayer AN AUTOENCODER-DRIVEN APPROACH FOR EMBEDDING GENERATION IN CONTENT SIMILARITY RECOMMENDATIONS 18TH NOVEMBER 2024 SIMONE SPACCAROTELLA
- 2. Beyond Metadata for BBC iPlayer AN AUTOENCODER-DRIVEN APPROACH FOR EMBEDDING GENERATION IN CONTENT SIMILARITY RECOMMENDATIONS
- 3. Introduction WHO I AM AND WHAT MY TECHNICAL BACKGROUND IS
- 4. Introduction
- 6. Project ideas Use of Contextual Data in personalized recommendations Play pattern and interaction pattern analysis Sequential recommender system Use of Passport tags to improve recommendations Item representation with Autoencoders Multi-product recommendations
- 7. Project ideas Use of Contextual Data in personalized recommendations Play pattern and interaction pattern analysis Sequential recommender system Use of Passport tags to improve recommendations Item representation with Autoencoders Multi-product recommendations
- 8. Project Overview A HIGH-LEVEL SUMMARY OF THE MAIN ASPECTS OF THE PROJECT
- 9. What I built A MACHINE LEARNING PIPELINE CONTENT-TO-CONTENT (C2C) SIMILARITY RECOMMENDER VIDEO-ON-DEMAND (VOD) "MORE LIKE THIS" SECTION ON BBC IPLAYER
- 10. More like this
- 12. Potentials Improve the quality of the recommendations Provide a common embeddings generator for content metadata Reduce costs and duplications The what
- 13. Contributions The how Ingested a feature-rich dataset that better describe the content Applied novel techniques to improve the descriptive power of the transformed content metadata
- 14. Motivations Improve on the identified issues Learn and apply novel ML techniques Align with the business objectives and needs The why
- 15. Issues with the current solution FEATURE POOR LACK OF PREPROCESSING SUBOPTIMAL SIMILARITY SCORING HARDCODED PARAMETERS
- 16. Techniques
- 18. Measure of Success KEY PERFORMANCE INDICATORS AND SCOPE OF THE PROJECT
- 19. Key Performance Indicators INTEGRATION WITH PASSPORT TAGS COMPARABILITY OF THE RECOMMENDATIONS WITH THE CURRENT SOLUTION
- 21. Cost Reduction
- 22. Potential cost reduction Linked to the Passport integration and adoption of the solution C N /
- 23. User Engagement
- 24. Diversity A MEASURE OF VARIETY WITHIN SIMILARITY RECOMMENDATIONS
- 25. Diversity A MEASURE OF VARIETY WITHIN SIMILARITY RECOMMENDATIONS Metadata Diversity
- 26. User Engagement
- 27. Out of Scope A/B test experimentation
- 28. In Scope Select & Proprocess the data Train and validate a model Generate and visualize the recommendations Build an MVP Share the results with stakeholders
- 29. The Main Intuition THE IDEA BEHIND THE SOLUTION
- 30. The Main Intuition Similarity Distance in space
- 31. The Data AN OVERVIEW OF THE IPLAYER CATALOGUE AND THE DATA STRUCTURE
- 32. Main Idea
- 33. Main Idea
- 34. The Data The iPlayer programmes structure
- 35. The Data The iPlayer programmes structure
- 36. The Data The iPlayer programmes structure episode
- 37. The Data The iPlayer catalogue
- 38. Catalogue size
- 39. The Data
- 40. Data drift
- 41. The Data The selected Passport tags
- 42. The Data The selected Passport tags Passport Tags Genre Format Contributor Motivation Editorial Tone Narrative Theme Relevant To About
- 43. Passport tags: example values
- 44. Source of bias
- 45. The Data The BBC Ontologies
- 46. The Data The BBC Things entities
- 47. The Data The BBC Things entities
- 48. The Data The BBC Things entities
- 49. The (Linked) Data Graph Data Structure
- 51. Implementation METHODS, JUSTIFICATION, AND TOOLS USED
- 52. Tools
- 53. The Pipeline Data Loading Data Preprocessing Modelling & Embeddings Generation C2C Similarity Recommendation Model Evaluation
- 55. Data Loading
- 60. Source of errors TAG A TAG B TAG C 1 0 1
- 62. Source of errors TAG A TAG B TAG C 1 0 1 0 0 0
- 72. Modelling & Embeddings Generation Data Split Ratio: 80%, 10%, 10% Batch Size: 300 Epochs: 100 Optimizer: Adam Loss Function: Binary Cross Entropy Early Stopping Patience: 10 Dropout Rate: 0.2 Hyperparameters
- 73. Modelling & Embeddings Generation ( , )= 1 =1 ln( 多)+(1 )ln(1 )多 =0 多 ( ln ( )+(1 ) ln(1 )) let and 0 1ln (1 ) ln(1 ) Binary Cross Entropy
- 74. Modelling & Embeddings Generation ( , )= 1 =1 ln( 多)+(1 )ln(1 )多 =1 多 ( ln ( )+(1 ) ln(1 )) let and ln ( )0 ln( ) Binary Cross Entropy
- 76. Modelling & Embeddings Generation Learning Rate Dropout Batch Normalization Hyperparameter Tuning
- 82. 8982 8982 Modelling & Embeddings Generation Ni [4491, 2245, 1122, 561, 280, 140, ...] N1 < 8982 N1 < 8982 N2 < N1 Hyperparameter Tuning
- 83. 8982 8982 Modelling & Embeddings Generation Ni [4491, 2245, 1122, 561, 280, 140, ...] Hyperparameter Tuning
- 84. 8982 8982 Modelling & Embeddings Generation Ni [4491, 2245, 1122, 561, 280, 140, ...] Hyperparameter Tuning
- 85. 8982 8982 Modelling & Embeddings Generation Ni [4491, 2245, 1122, 561, 280, 140, ...] Hyperparameter Tuning
- 86. 8982 8982 Modelling & Embeddings Generation Ni [4491, 2245, 1122, 561, 280, 140, ...] Hyperparameter Tuning
- 87. 8982 8982 Modelling & Embeddings Generation Ni [4491, 2245, 1122, 561, 280, 140, ...] Hyperparameter Tuning
- 88. Modelling & Embeddings Generation Ni [4491, 2245, 1122, 561, 280, 140, ...] Hyperparameter Tuning
- 89. Modelling & Embeddings Generation Ni [4491, 2245, 1122, 561, 280, 140, ...] Hyperparameter Tuning
- 90. Modelling & Embeddings Generation Ni [4491, 2245, 1122, 561, 280, 140, ...] Hyperparameter Tuning
- 91. Modelling & Embeddings Generation Ni [4491, 2245, 1122, 561, 280, 140, ...] Hyperparameter Tuning
- 92. Modelling & Embeddings Generation Ni [4491, 2245, 1122, 561, 280, 140, ...] Hyperparameter Tuning
- 93. Modelling & Embeddings Generation Ni [4491, 2245, 1122, 561, 280, 140, ...] Hyperparameter Tuning
- 94. Modelling & Embeddings Generation Ni [4491, 2245, 1122, 561, 280, 140, ...] Hyperparameter Tuning
- 95. Modelling & Embeddings Generation Ni [4491, 2245, 1122, 561, 280, 140, ...] Hyperparameter Tuning
- 96. Modelling & Embeddings Generation Ni [4491, 2245, 1122, 561, 280, 140, ...] Hyperparameter Tuning
- 97. Modelling & Embeddings Generation Ni [4491, 2245, 1122, 561, 280, 140, ...] Hyperparameter Tuning
- 98. Modelling & Embeddings Generation Ni [4491, 2245, 1122, 561, 280, 140, ...] Hyperparameter Tuning
- 100. Modelling & Embeddings Generation N. Hidden Layers: 1 Embedding Size: 561 Batch Normalization: False Dropout: True Learning Rate: 0.01 Hyperparameter Tuning
- 114. Timmy Time Timmy is a little lamb with lots to learn. Join him as he heads off on adventures. Confidence: 94.737% Patchwork Pals The Patchwork Pals live on a patchwork blanket and pull together to solve problems. Confidence: 96.017% Fireman Sam Fun with the friendly fireman and the villagers of Pontypandy. Confidence: 94.199% Arthur Animation following the adventures of the world's most famous aardvark. Confidence: 94.061% Postman Pat: Special Delivery Service Children's animation with Postman Pat, the new Head of the Special Delivery Service. Confidence: 93.940% Tish Tash Following the adventures of a young bear called Tish. Confidence: 93.734% Octonauts Animated deep sea adventures with Captain Barnacles and his band of explorers. Confidence: 93.241% Hey Duggee Duggee runs the Squirrel Club, where children can earn badges for learning new skills. Confidence: 93.237% Bob the Builder Bob and the gang have so much fun because working together they get the job done. Confidence: 93.012% Tee and Mo Explore the amazing world around us with baby monkey Tee and his Mum. Confidence: 92.761% Raa Raa the Noisy Lion Hang out with Raa Raa and his animal friends as they solve all sorts of mysteries. Confidence: 92.487% Mr Bear's Christmas A heartwarming, festive tale of friendship and love, celebrating the magic of Christmas. Confidence: 92.303%
- 121. Conclusions A CRITICAL EVALUATION OF THE PROJECT
- 122. Strengths Inference time & space complexity Improved similarity semantic General solution
- 123. Weaknesses Lack of model interpretability Pipeline re-training for unseen tags Hyperparameter tuning time
- 124. Outcomes General solution for multimodal BBC content Comparable recommendations Positive feedback from stakeholders
- 125. Questions Answered What do recommendations look like using these embeddings? Can we assess the embeddings subjectively? Can we assess the embeddings using offline scoring? How do we know that the embeddings make any sense? How might we use these embeddings for recs?
- 126. Recommendations Build a fully automated pipeline on Sagemaker Integrate with real-time Passport Tags Evaluate different similarity-score metrics A/B test the solution
- 127. Future Improvements Use of Graph Neural Network (Graph Autoencoder) Plug the Diversity metric in the training phase
- 128. THANKS
- 129. Q & A ANY QUESTIONS?