Music Recommendation System

Team Name: APT

​

강한이, 이민희, 이서일

​

Contents

1. Why Music?
2. Project Overview
1. Problem Definition
2. Dataset
3. Evaluation Metric
3. Methodology
• Model-based Approach
• Additional Approach

4. Project Plans
1. Schedule
2. Team Roles
5. Q&A

2

Why Music?

Music recommendation system is on high demand for music streaming services.

3

Why Music?

Music consumption does not necessarily guarantee a correlation with user preference.

4

Why Music?

Personalization is not always the best approach in music recommendations.

​

5

Project Goal

KKBox's Music Recommendation Challenge¹

• The 11th ACM International Conference on Web Search and Data Mining (WSDM 2018)

6

[1] [Deleted User], Howard, A., Chiu, A., McDonald, M., msla, Kan, W., & Yianchen. (2017). WSDM - KKBox's Music Recommendation Challenge. Kaggle. https://kaggle.com/competitions/kkbox-music-recommendation-challenge

Dataset

Use the dataset provided by the challenge.

• User-song play data
• User info. & Song info.

The train/test split is provided:

• Training data: 7,377,419 entries
• Test data: 2,556,791 entries

7

Dataset

Play information for each song per user is provided.

• Users: 34,404, Songs: 2,296,834
• Whether a user has chosen a specific song is represented by 0 or 1 # target

8

* msno: user_id

Dataset

Additional information about users and songs is provided.

• Users: region, age, gender, registered_via, etc.
• Songs: length, genre, artist_name, composer_name, etc.

​

9

Evaluation Metric & Our Goal

• AUROC
• Metric used in the challenge
• Predicts the probability that an item will be marked as "1," indicating it has been chosen by the user.
• AUROC metric is calculated based on these.
• The goal is to exceed the given benchmark.
• 0.61337

10

Methodology:

Model Based

Recommendation System

11

LightGBM

• Tree-based Gradient Boosting Machine
• Builds a sequence of decision trees, where each tree corrects the errors of the previous ones.
• Assigns prediction scores based on the value of the leaf node where each example lands.

12

LightGBM

Node Functionality:

• Non-Leaf Nodes: Serve as decision points that guide the example through the tree based on feature values, routing it to a specific leaf node.
• Leaf Nodes: Contain the final score that contributes to the overall prediction.

13

LightGBM

e.g.

14

LightGBM

e.g.

15

Feature 1 <= 0.8

Feature 2 <= 1.0

0.5

1

-0.3

Tree 1

Leaf node values represent adjustments to the log-odds of the prediction.

LightGBM

e.g.

16

Feature 1 <= 0.8

Feature 2 <= 1.0

-0.3

0.5

1

Leaf node values represent adjustments to the log-odds of the prediction.

Tree 1

LightGBM

e.g.

17

Feature 1 <= 0.8

Feature 2 <= 1.0

-0.3

0.5

1

Leaf node values represent adjustments to the log-odds of the prediction.

Tree 1

LightGBM

e.g.

18

Feature 1 <= 0.8

Feature 2 <= 1.0

0.5

1

-0.3

Leaf node values represent adjustments to the log-odds of the prediction.

Tree 1

LightGBM

19

In each boosting round, LightGBM grows a tree by recursively splitting nodes

based on feature thresholds to maximize information gain,

assigning leaf values as adjustments to predictions.

Feature 1 <= 0.8

Feature 2 <= 1.0

0.5

1

-0.3

Tree 1

LightGBM

20

Tree 1

Tree k

…

Trains multiple trees iteratively, with each tree learning to correct the residual errors

from the ensemble of all previously trained trees

to minimize the overall loss.

Why LightGBM?

• Excellent for tabular data
• Can process categorical features directly without requiring one-hot encoding or manual transformation.
• Supports feature selection
• Prioritizes important features during training by evaluating feature importance based on split gain or frequency
• Plain, simple, and out-of-the-box model
• Able to processing large-scale data with limited resources

21

Features to Use

​

• This dataset is much richer than what we have learned during classes
• It gives information about user and item, in addition to user-item matrix.

22

Features to Use

1. user, song metadata
• information given by the dataset -> likes other did at the competition
2. user-song Collaborative Filtering score
3. song audio embedding (not used)

23

1. User, Song Metadata

• Songs: length, genre, artist_name, composer_name, lyricist, language
• Users: region, age, gender, registered_via, registration_init_time, expiration_date

24

• User, Song Metadata

​

25

• User, Song Metadata

26

2. Collaborative Filtering Score

​

From User-Song Matrix

1. Compute the reduced representations (ft. SVD)
2. Compute the Similarity between Users or Songs
3. Construct Peer groups: Top 100 neighborhood
4. Generate Top-100 CF score for each User-Song pair

* For Similarity: Cosine similarity or NN-based similarity

27

3. Song Audio Embedding (not used)

Get audio embedding by a pretrained model

How to obtain audio

• Download a 30-second audio clip from YouTube

Pre-trained Model: MusicFM-MSD²

• Output: Embeddings with dimensions (4, 1024)

28

[2] M. Won, Y.-N. Hung, and D. Le, “A foundation model for music informatics,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), 2024.

3. Song Audio Embedding (not used)

• Crawling audio data within the project period is challenging.
• While using music embeddings as features could enhance performance, we decided not to include them in this project.

29

Project Plans

30

Project Timeline

11/13 ~ 11/19: Data Preprocessing & Studying Reference

11/20 ~ 11/26: Training the Model

11/27 ~ 12/3: Evaluating & Modifying the Model

12/4 ~ 12/11: A Preliminary Period

12/11: Final Presentation

31

Each Member’s roles

All Members: Data Processing, Making the Model, Studying Reference

강한이: Explore LightGBM applications

이민희: Tune LightGBM Hyperparameters

이서일: Set the Dataset

32

Q & A

Additional Approach:

Neural Network Based

Recommendation System

34

Motivation

Unlike other types of contents, music consumption does not necessarily guarantee a correlation with preference.

Content-based item (song) embeddings might give more reliable result.

35

Proposed Method

36

Proposed Method

37

Pretrained

Model

Item 1 embeddings

Item 3 embeddings

User 1 embeddings

Item 4 embeddings

Average

Proposed Method

38

I_a

I_b

I_c

I_d

U₁

U₁

​

Calculate the probability of being chosen by computing the cosine similarity between user and item embeddings.

​

Proposed Method

• Obtain item embeddings using a pre-trained model.
• Calculate user embeddings by averaging the embeddings of items the user has chosen.
• For test items, calculate the probability of being chosen by computing the cosine similarity between user and item embeddings.

39

Details

How to Obtain Audio

• Download a 30-second audio clip from each song, specifically from the 30- to 60-second mark.
• Use the ‘yt-dlp’ Python library.
• Query YouTube with ‘{artist} {title}’ and select the most suitable video.

40

Details

Pre-trained Model: MusicFM-MSD²

• Input: 30 seconds, 24kHz audio
• Output: Embeddings with dimensions (4, 1024)
• Take the mean across the time axis

41

[2] M. Won, Y.-N. Hung, and D. Le, “A foundation model for music informatics,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), 2024.

Is it Possible?

• Time required to crawl audio from YouTube
• ​

42

GBM 모델 설명 (내용 덤프)

43

GBM

• Tree-based, Gradient Boost Machine (GBM)
• 그래서 이게 뭔데? (설명할 정도의 공부 필요)
• classifier, ranker 등의 종류
• feature를 이용해서 분류해준다
• 그리고 feature는 뭐가 되든 간에 상관 없다
• user, item, user-item
• 결국 다 이 방식을 쓴다면 feature를 어떻게 정하느냐가 중요
• 우리는 이걸 잘하겠다. ㅋㅋ

44

LightGBM + FeatureEngineering

• Tree-based Gradient Boosting Machine
• Features: User info. & Song info. & User-Song info.
• Target: User-Song Play Probability (Binary Classification)

​

• Collaborative Filtering used in Feature Engineering

45