Hello, dear friend, you can consult us at any time if you have any questions, add WeChat: daixieit

Capstone Project Introduction to Machine Learning

Classification flavor

Here is what we want you to do:

*Download the file musicData.csv from Brightspace

*Write code that creates a classification model (of your choice, can be anything we covered in class –  including SVM, trees, forests, boosting methods, neural networks, etc.; for good classification performance, make sure to include a suitable dimensionality reduction and clustering step before attempting the classification) that predicts the genre of a song in the test set from the rest of the data. *In the code, before anything else happens, we want you to initialize the seed of the random number  generator with your NYU N-number. For instance, if your N-number (on the back of your NYU ID) is       N18994097, we would expect the first lines of your code to be:

import random

random.seed(18994097)

Note: Don’t use “18994097”, unless that is your N-number. Use yours instead. This is so that the correct answers (which depend on which movie ratings go into the test set) are specific to *you*.

*Make sure to do the following train/test split: For *each* genre, use 500 randomly picked songs for the test set and the other 4500 songs from that genre for the training set. So the complete test set will be     5000x1 randomly picked genres (one per song, 500 from each genre). Use all the other data in the           training set and make sure there is no leakage .

*Use this test set to determine the AUC of your model. Try to get the AUC as high as possible without having any leakage between training and test set.

*Write a brief report (1-2 pages) as to how you built your model, how you made your design choices     (why you did what you did) and addressing how you handled the challenges below. Make sure to state  your final AUC at the bottom of the report and please include a plot of the ROC curve in your report. As dimensionality reduction will be critical, please also include a visualization of the genres as clusters in    the lower dimensional space (lower than the dimensionality of the original data) you uncovered and     include a comment as to what you think about this space/clustering . Also make sure to comment on     what you think is the most important factor that underlies your classification success.

*For extra credit, include some interesting non-trivial observation or data visualization in your report.    *Upload both the report (pdf only) and your code (some kind of python or notebook format only) to the Brightspace portal by the due date.

Some key challenges you can expect when taking on this project:

*There is randomly missing data, e.g. some of the durations of some of the songs are missing, as well as some of the auditory feature values. There are not many missing values, but you have to handle them   somehow, either by imputation or by removing the missing data in some reasonable way.                         *The acoustic features are unlikely to be normally distributed.

*Some of the data is provided in string format, e.g. the key. This will need to be transformed into numerical data to be useful.

*Some of the data is provided in categorical format, e.g. mode. This will need to be dummy coded. *The category labels of the genres will need to be transformed into numerical labels.

*As this is a multi-class classification (not just binary), you need to be careful – only if the predicted       classification of a genre in the test set matches the actual genre of a song is the classification correct. In other words, there are many more ways to be wrong than to be right.

*Make sure not to normalize categorical values (like mode) for the purposes of doing dimensionality reduction.

*There might be information in the linguistic properties of artist and song, but you don’t have to include that you in your model.