Combining Collaborative & Content Filtering in a Recommendation System for a Web-based DAW
summary
EarSketch implements a hybrid recommender that combines collaborative co-usage statistics from 20,000 scripts with content similarity computed via STFT and MFCCs, producing a per-sample score that weights co-usage and acoustic distance (scaled by sqrt(n)) and exposes four labeled recommendation modes. The web DAW adds live recommendation folders that refresh on keystroke and avoids user profiles. Offline and live evaluations (Mechanical Turk and deployed users) show strong uptake of the discover different kinds of sounds mode and increased diversity in chosen samples.
AI-written summary · about
chapters (13)
- 0:00Introduction to EarSketch
- 1:08Overview of recommendation system
- 3:18Problem with sample usage patterns
- 5:50Recommendation system design
- 8:24Content filtering techniques
- 10:55Recommendation score categories
- 13:03Updated sound browser integration
- 14:08User study with Mechanical Turk
- 15:32EarSketch user integration study
- 17:40User intent and future work
- 19:11Comparison with commercial tools
- 20:52Impact on musical output
- 24:25Q&A and closing
related talks
04mThe Sound Sketchpad: Expressively Combining Large and Diverse Audio Collections
43mBuilding your first recommender
21mRecsys Keynote: Improving Recommendation Systems & Search in the Age of LLMs
41mRecommender Systems: Crafting Effective Personalised Experiences
20mLive Coding With Crowdsourced Sounds And A Virtual Agent Companion (second session)
20mLive Coding With Crowdsourced Sounds And A Virtual Agent Companion
28mMaking Recommendations without Data
43mRecommendation System beyond traditional Collaborative filtering