I really enjoyed this project, reviewing how recommendation systems work throughout music streaming services (Spotify, Apple Music, etc). My early notes:
- Collaborative Filtering in a nutshell: If user A & B have similar preferences, then songs liked by A but not yet considered by B will be recommended to B.
- Performing Collaborative Filtering is based on Matrix Factorization.
- Content based filtering is used often in such services.
- The basic concept in content based filter: compare sets of audio signals which represent items in a meaningful way.
- Spotify does this by comparing audio signals. E.g. Spotify radio.
- Deep learning method also used (machine learning).
- Deep learning method “mimics the architecture of mammalian brains”. It can learn features at multiple levels from low-level data without resorting to manually crafted features.
- Deep learning “strives to provide a model for human cognition”.
- Semantic issues in music recommendation systems…
- Various properties cannot be obtained from audio signals.
- Characteristics which influence user preference are not equal to the corresponding audio signal.
- Item & user vector shortcoming or the Harry Potter effect.
- Popular items in a recommender can lead to feedback loops. Rich get richer. If Queen gets purchased by more users. Likely to be listened by more users. Likely to be recommended to more users. Harry Potter effect lessens the diversification of items.