Local

Music

Recommendation

Douglas Turnbull

Ithaca College

Ithaca, New York

USA

Part 1

Recommender Systems

Personalized Radio

Celestial Jukebox

Recommender System (RecSys)

A computational system that

recommends

specific items to a

specific user.

Recommender System (RecSys)

A computational system that

recommends

specific [artists, songs, events] to a

specific [listener, fan].

Content-based

(CB)

​

Analyze Substance of Items

​

Collaborative

Filtering

(CF)

​

Analyze User Interactions with Items

​

​

​

Hybrid

​

Combine CB and CF Approaches

Content-Based

(CB)

​

Analyze Substance of Items

​

1. Human Annotation
2. Social Tagging
3. Text-Mining Reviews
4. Computer Vision/Audition

Collaborative

Filtering

(CF)

​

Analyze User Interactions with Items

​

• Explicit Feedback
• Implicit Feedback

​

​

“Cold Start” Problems

New Item has no user ratings

Solution: Content-based RecSys

​

​

New User has no interaction with the RS

Solutions: Social Login, UX Onboarding

CF Matrix Representation

r_i,u = how much user u likes (or dislikes) item i

​

​

item i

user u

|U|x|I|

r_u,i

Neighborhood Algorithms

​

• Collect Joe’s favorite movies
• Find users who also like the same movies
• Recommend movies that they like to Joe

​

Two Ways to Compute

• User-User Similarity
• Item-Item Similarity

​

User-User Neighborhood

​

1. Represent Joe as a column vector
2. Calculated the similarity between Joe’s vector and all users
• e.g., cosine distance
3. Find k most similar “neighbors”
4. Average rows from neighbors
5. Rank items

​

Item-Item Neighborhood

​

• Compute Item-Item Similarity Matrix
• Recommend most similar items to items in the user model

​

User Model

Problems with using Raw Data Matrix X

Size: millions of items, millions of users

Sparse: inefficient storage + computation

Noisy: lots of meaningless junk

Synonymy: “hip hop” vs. “rap”

Polysemy: “romantic”

​

Matrix Factorization “fixes” these problems

​

Part 2 �Matrix Factorization

Embedding Items and Users into Latent “Semantic” Space

Matrix Factorization

min_U,I ∑_ratings (r_u,i - x_u^T y_i)²

​

k - selected dimension of embedded space

​

​

Matrix Factorization

min_U,I ∑_ratings (r_u,i - x_u^T y_i)²

+ λ (||U||² + ||I||²)

λ - regularization parameter

​

​

Matrix Factorization

min_U,I ∑_ratings (r_u,i - x_u^T y_i)²

+ λ (||U||² + ||I||²)

​

Alternating Least Squares (ALS)

• Fix I ￫ solve for U
• Fix U ￫ solve for I

​

​

Implicit Data

min_U,I ∑_u,i c_u,i(r_u,i - x_u^T y_i)²

+ λ (||U||² + ||I||²)

​

​

​

​

r_i,u = 1 when user u has interacted with item i

0 otherwise

c_i,u = confidence in rating r_i,u

C_i,u= 50 when r_i,u = 1

C_i,u= 1 when r_i,u = 0

Explicit vs. Implicit Data Loss Function

Implicit Feedback

min_U,I ∑_u,i c_u,i(r_u,i - x_u^T y_i)

+ λ (||U||² + ||I||²)

​

​

​

​

Explicit Ratings

min_U,I ∑_ratings (r_u,i - x_u^T y_i)

+ λ (||U||² + ||I||²)

Part 3�Long-Tail

Music Recommendation

Most Popular →

Users - Playlists

Items - Tracks→

Least Popular →

Short

Head

Mid

Body

Long

Tail

min_U,I ∑_u,i c_u,i(r_u,i - x_u^T y_i)²

Popularity Bias

Small minority of Short-Head have big impact on loss function.

min_U,I ∑_u,i c_u,i(r_u,i - x_u^T y_i)

Item-Balanced Loss Function:

Scale c_u,i so each track has same weight.

​

c_u,i ← c_u,i / ||c_*,i||

​

​

​

Real Plot from Gabe

​

Million Playlist Dataset

​

1M Playlists

2.2M Tracks

​

Retain tracks with on at least 10 playlists

→ 370K tracks

​

One-Third Partitions:

Short-Head → 2K tracks

Mid-Body → 18K tracks

Long-Tail → 350K tracks

​

​

Training Playlists

Test Playlist

Item-Balanced Loss Function

Item-Balanced Loss Function

Item-Balanced Loss Function

Long-tail Recommendation Takeaways

Item-Balanced Loss Function results in

• big improvement for long-tail tracks
• overall improvement

​

Neighborhood-based approaches can outperform Matrix Factorization for long-tail data

• see recent publications

​

​

​

​

​

Part 3

Local

Music Recommendation

Kevin Kinsella - Grassroots 2013 – Mark Anbinder

Why MegsRadio Failed?

• too many music streaming giants
• YouTube, Spotify, Apple Music, Google Play

​

• personalized Internet radio is dead
• Long live celestial jukeboxes

​

• our inability to maintain professional-grade service
• Needed to be more responsive to users
• Students graduate too quickly

​

• playlist algorithm is not tuned correctly
• too many WTF music recommendations

​

• obtaining music is hard

​

Northeastern Music Informatics Special Interest Group

Ithaca College

Februrary 2020

JimiLab @ Ithaca College:

Just Inspired Music Innovation

​

Douglas Turnbull

Ithaca College

dougturnbull.org

​

​

(Truncated) Singular Value Decomposition

k is a hyperparameter

X_k = best k-rank approximation of X with respect to |X - X_k|₂ (Frobenius norm)

​

https://en.wikipedia.org/wiki/Latent_semantic_analysis

u

​

users

i

items

|I| x |U|

X

Kevin Kinsella - Grassroots 2013 – Mark Anbinder

◉

◉

◉

◉

◉

Popularity Bias - Algorithmic

1. Popular Items have more Ratings
2. More Ratings ￫ More Signal ￫ More Confidence
3. Structural Bias in Model

Add Matrix Visualization HereasdfAAAA

Local Music Data

• 8 Cities in the USA
• ~36 Artists Per City
• 99.999% Sparse

​

Local Music Experiment

6 RecSys Algorithms

• Baselines: Random, Popularity
• Neighborhood: ItemItem, UserUser
• Matrix Factorization: ALS, BPS

​

3 Evaluation Metrics

• NCDG, Precision@R, Precision@1

​

(Surprising) Results #1

Item-Item Neighborhood works better than Matrix Factorization

• Track & Artist Recommendation
• Perhaps due to extreme sparsity

​

​

​

Per-Track Loss Function

Since popular tracks have more ratings, the model designed to overfit them, while underfitting obscure tracks.

​

Solution: Alter the loss function so that each track has some influence.

​

​

​

Popularity Bias - Commercial

​

​