Understanding ai music discovery and recommendation systems
Download as PPTX, PDF0 likes268 views
The document provides an overview of AI music recommendation systems, discussing their necessity due to the influx of new tracks and the challenges faced in music discovery. Various recommendation techniques such as demographic, collaborative, content-based, context-based, and hybrid approaches are outlined, along with their advantages and limitations. It also highlights current opportunities for improvement in music recommendation engines and introduces Melodrive, a company specializing in AI music solutions.
Understanding ai music discovery and recommendation systems
- 1. Understanding AI music discovery and recommendation systems
- 2. I am Valerio Velardo, PhD Cofounder & CEO @ Melodrive AI + Music Researcher Hello!
- 3. What you’ll learn... ◍ Basic grasp of how music recommendation systems work ◍ How to leverage recommendation engines to add value to music applications ◍ Understand how listeners discover music according to our research findings
- 4. Overview 1. The (music) recommendation economy 2. Music recommendation techniques 3. Current limitations and untapped MR opportunities 4. Melodrive 5. Q&A
- 7. Why is AI music recommendation needed? 40,000 tracks uploaded on Spotify every day* *Ingham T. (2019), Nearly 40,000 tracks are now being added to Spotify every single day
- 8. Why is AI music recommendation needed? Dachs P. (2019), Hear what I’ve found — A survey on music discovery 83% Music fans who actively search for new music
- 9. Why is AI music recommendation needed? Dachs P. (2019), Hear what I’ve found — A survey on music discovery
- 10. Why is AI music recommendation needed? Dachs P. (2019), Hear what I’ve found — A survey on music discovery
- 12. MR features ◍ Relevance ◍ Novelty ◍ Quality music
- 13. MR challenges ◍ Songs can be consumed many times ◍ Instant feedback ◍ Context-dependent consumption ◍ User-dependent consumption
- 14. MR use cases ◍ Suggest interesting music ◍ Suggest new artists ◍ Generate playlists ◍ Recommend loops/samples ◍ ...
- 15. Human MR ◍ Audio metadata ◍ Music metadata ◍ Cultural metadata ◍ Context metadata
- 16. Current music discovery methods Dachs P. (2019), Hear what I’ve found — A survey on music discovery
- 17. MR approaches 1. Demographic filtering 2. Collaborative filtering 3. Content-based filtering 4. Context-based filtering 5. Hybrid approaches
- 18. Demographic filtering ◍ Cluster users ◍ Personal, geographical and psychographic data
- 19. Demographic filtering + ◍ Simplest approach - ◍ Recommendations are generic ◍ Effort from the user
- 20. Collaborative filtering ◍ Recommendation based on user-item relationship ◍ Implicit feedback (music consumption habits) ◍ Explicit feedback (user ratings)
- 22. Collaborative filtering + ◍ Quite effective ◍ Simple implementation - ◍ Cold start problem ◍ No content-related descriptions ◍ Popularity bias
- 23. Content-based filtering ◍ Use tags describing the content of the musical item ◍ Tags can be extracted manually/automatically ◍ Based on similarity among items
- 24. Content-based filtering + ◍ Audio and music descriptors ◍ No cold start problem - ◍ Manual: scalability ◍ Manual: time consuming ◍ Automatic: complex implementation ◍ Subjective
- 25. Context-based filtering ◍ Use contextual information to describe musical items ◍ Social tagging ◍ Web mining
- 26. Context-based filtering + ◍ Cultural information ◍ Flexible descriptions - ◍ Inconsistent tags ◍ Coverage ◍ Time consuming
- 27. Hybrid approaches ◍ Mix different approaches to obtain better results ◍ Weighted recommendations ◍ Switch between different approaches
- 28. Hybrid approaches Velardo V. (2019), Spotify’s Discover Weekly explained — Breaking from your music bubble or, maybe not?
- 29. 3. Current limitations and untapped MR opportunities
- 30. Shortcomings of current MR systems ◍ Musical bubble ◍ Trust ◍ Contextual relevance
- 31. Being stuck in your musical bubble
- 32. Getting out of your musical bubble ◍ Exploration slider ◍ Multiple-feature exploration
- 33. Trusting MR engines ◍ Why is this music relevant to me? ◍ Discovery is in the journey not the destination
- 34. The MR friend ◍ Deep user knowledge ◍ Suggest relevant music ◍ User-AI collaboration ◍ Motivate recommendations
- 35. The MR friend for classical music ◍ Create a custom musical growth path ◍ Explain how music works ◍ Teach music history
- 36. Current MR engines ignore context
- 37. Context-aware MR system MR engine Time Location Mood Device
- 38. 4. Melodrive
- 39. Who we are ◍ Team of PhDs in Music and AI ◍ Experience in industry and academia ◍ AI music consulting
- 40. What we’ve worked on ◍ Built AI generative music engine used by 1K game developers ◍ Worked on client projects in AI music
- 41. What we do ◍ AI and data strategy/training ◍ Design and implement bespoke AI music solutions ◍ Build generative music soundtracks