COMP90049 - Introduction to Machine Learning - Project 4 - Solved

Music Genre Prediction from Audio, Metadata and Text Features!

Task:                       Build a music genre classifier


Overview
The goal of this project is to build and critically analyse some supervised Machine Learning algorithms, to automatically identify the genre of a song on the basis of its audio, metadata and textual features. That is, given a list of songs, your job is to implement one or more Machine Learning model(s), train them using the training dataset, and evaluate using the test validation and test dataset.

This project aims to reinforce the largely theoretical Machine Learning concepts around models, data, and evaluation covered in the lectures, by applying them to an open-ended problem. You will also have an opportunity to practice your general problem-solving skills, written communication skills, and creativity.

This project has two stages. At Stage I, the focus will be the implementation, experimentation, and the report, where you will demonstrate the knowledge that you have gained, in a manner that is accessible to a reasonably informed reader. At Stage II, you will review 2 reports written by your peers in Stage I and submit a review of about 200-400 words.


Dataset
Each song (instance) is represented through a large set of features (described in detail in the README), and listed in the features.csv files. Each song is labelled with a single genre tag, which is provided in the labels.csv files.

The data files are available via the LMS. You will be provided with a set of training, validation and test instances. The files are provided in csv format, which stands for comma-separated values.

•    train features.csv: Contains features of 7678 training instances.

•    train labels.csv: Contains a single genre label for each training instance

•    valid features.csv: Contains features of 450 validation instances.

•    valid labels.csv: Contains a single genre label for each validation instance.

•    test features.csv: Contains features of 428 test instances.

Each song in the data set is indexed with a unique trackID. We provide three different types of features. Details are provided in the README file, as well as the references listed under Terms of Use:

•    Metadata features: For each song, we provide its title, loudness, tempo, key, mode, duration, and time_signature .

•    Text features: For each song, we provide a list of tags representing the words that appeared in the lyrics of the song and are human annotated (such as ‘dance’, ‘love’, or ‘never).

•    Audio features: We provide 148 pre-computed audio features that were pre-extracted from the 30 or 60 second snippets of each track, and capture timbre, chroma, and ‘Mel Frequency Cepstral Coefficients’ (MFCC) aspects of the audio. Each feature is continuous and the values are not interpretable.

Each song is labelled with its genre, i.e., with a single label from one of 8 possible genre labels:

‘Soul and Reggae’, ‘Pop’, ’Punk’, ‘ Jazz and Blues’, ‘Dance and Electronica’, ‘Folk’, ‘Classic Pop and Rock’, ‘Metal’

Task
You will develop Machine Learning models which predict the music genre based on a diverse set of features, capturing audio features of the track, as well as song metadata such as its title, key, duration, and textual tags representing the words in the song lyrics. You will implement and analyze different Machine Learning models in their performance; and explore the utility of the different types of features for music genre prediction.

We will use a hold-out strategy to evaluate the trained model using a validation, and a test set:

1.    The training phase: This will involve training your classifier(s) and parameter tuning where required.

You will use the train  features and train lyrics files.

2.    The validation phase: This is where you observe the performance of the classifier(s). The validation data is labelled: you should run the classifier that you built in the training phase on this data to calculate one or more evaluation metrics to discuss in your report. This phase will help you to find the best model that can be used for the testing phase.

3.    The testing phase: The test data is unlabeled; you should use your preferred model to produce a prediction for each test instance, and submit your predictions to Kaggle website; we will use this output to confirm the observations of your approach.

N.B: Various Machine Learning techniques have been (or will be) discussed in this subject (Naive Bayes, Decision Trees, 0-R, etc.); many more exist. You are strongly encouraged to make use of existing Machine Learning libraries (such as sklearn) in your attempts at this project.

[1] This should be at a conceptual level; a detailed description of the code is not appropriate for the report.