Recommendation Algorithm

[sophryu99]

Content-based recommendation algorithm

Input Data format

recommend_songs([{'name': 'Butter', 'year': 2021},
                {'name': 'Permission to Dance', 'year': 2021},
                {'name': 'Dynamite', 'year': 2020}],  track_data)

[sophryu99]

Features to train with

# Get Track info API
    track_id = models.IntegerField(default=0)
    artist_id = models.IntegerField(default=0)
    year = models.IntegerField(default=0)
    popularity = models.IntegerField(default=0)
    release_date = models.IntegerField(default=0)

    # Get audio_features API
    acousticness = models.FloatField(default=0)
    danceability = models.FloatField(default=0)
    duration_ms = models.FloatField(default=0)
    energy = models.FloatField(default=0)
    instrumentalness = models.FloatField(default=0)
    liveness = models.FloatField(default=0)
    loudness = models.FloatField(default=0)
    valence = models.FloatField(default=0)
    speechiness = models.FloatField(default=0)
    tempo = models.FloatField(default=0)
    key = models.IntegerField(default=0)
    mode = models.IntegerField(default=0)

[sophryu99]

Training Metrics

from sklearn.cluster import KMeans
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline
from sklearn.decomposition import PCA

[sophryu99]

Basic idea of the recommendation algorithm

• Through eda, it is clear that similar genres tend to have data points that are located close to each other
• Take the data points of the songs user provided, and recommend songs corresponding to nearby data points.

[sophryu99]

1. Get the average vector of the audio features for each tracks

Cosine distance function for two vectors u and v

# find the distance between two pairs of collections of points
from scipy.spatial.distance import cdist

• Documentation for cdist

2. Find the n-closest data points in the dataset to this average vector 3. Take these n-points and recommend the songs corresponding to them