Music Recommendation Algorithm

Artwork created by Dalle 3 AI

Project Overview

This project focuses on the development of an audio recommendation algorithm using unsupervised learning model called K-means. The dataset comprises songs spanning from 1950 to 2019, featuring various lyrical features and metadata. The goal is to create a machine learning pipeline that clusters songs based on their lyrical content and other features, allowing for personalized song recommendations.

The project is structured into the following phases:

1. Exploratory Data Analysis (EDA):

• Analyzed data to gain insights into music genres, themes, and artist distribution.
• Explored correlations between different features.
• Identified trends in song release dates, genres, and themes over time.
• Visualized distributions of song lengths, genres, and themes.

Insights from EDA:

• Found that out of 5424 artists, only 3308 have more than one song in the training dataset.

• Discovered that the top 10 artists in the dataset are not primarily hip-hop, reggae, or pop artists.

• Bar plot of song counts by genre, shows that the POP genre is the most popular one among every decade represented in our dataset

  

• Bar plot of song counts by theme, shows that the violence and obsenity is the dominant theme of the songs since 2010s

  

2. Data Cleaning:

• Prepared the TEST and TRAIN data for K-Mean Classification Model.
• Dropped irrelevant and strongly correlated columns indentified by use of HEATMAPs.
• Manipulate the dataframes to extract identifiers and features to meet the requirements for classification model.

3. Data Modeling:

• Utilized K-means clustering algorithm with 5 clusters, identified by combination of Elbow and Silhouette Plot.
• Use the developed model on the test dataset.
• Applied Principal Component Analysis (PCA) to reduce dimensionality for visualization.
• Assigned cluster labels to each song and identifier.
• Developed an interactive recommendation tool based on cluster classification.

4. Reporting:

• Summarized insights gained from EDA.
• Described the process of determining optimal clusters.
• Provided human-interpretable descriptions of the clusters.
• Presented a recommendation tool for users based on their song preferences.

Recomentation Tool:

At the end of the phase 3 and on the phase 4, I introduce the simple python script that use findings of our K-means model to recomend songs to users who will select a prefered song from the list of our songs in our TRAIN dataset. This tool use a simple python commands like if and elif.

Future Steps:

• Enhance Model Performance: Experiment with different clustering algorithms and hyperparameters to improve the accuracy and robustness of the recommendation algorithm.

• Lyrical Analysis: Optionally, conduct a sentiment analysis on the song lyrics using techniques like Word2Vec to extract deeper insights into the emotional content of the songs.

• Improving Recommendation Tool: Enhance the recommendation tool by exploring alternative ways to provide the end user with the song recomentdations based on their favorite song (using the clusters created by our K-Means model)

• Deployment on AWS and Telegram: Explore hosting the recommendation system on Amazon Web Services (AWS) for scalability and reliability. Additionally develop a Telegram bot interface for users to interact with the recommendation system seamlessly.