KNN Genre Prediction Model

Idea

The idea behind this model is to take in a bunch of numerical values based around values in a dataset and see if it can predict the genre of a song if given those values

Creation

As can be seen from some of the EDA above, the following parameters were chosen to predict the genre:

• valence
• tempo
• energy
• instrumentalness
• speechiness
• acoustics
• liveliness
• loudness
• key
• danceability

From there, we used the KNN section of the sklearn library. To put it simply, KNN (k-nearest neighbors) is a classification algorithm that assigns a label to a new data point based on the majority class of its k nearest neighbors in a labeled training dataset.

Training and Fine-tuning

The EDA cleaned up our data, so now all that was needed was to prep the data to be fed to the model. To do this we took the aforementioned statistics on songs, along with what genre the song was (as numbers that represented what genre the song was since the model only accepts numerical values in this scenario), and randomized the order of the songs in such a way that every time the model was run it randomized the order so we could make sure the model would work on more data and not just memorize a few values, along with making it so we could get a range of how well the model does on average.
Afterward, we split up the data into 4 different sets, two holding input, one for training and the other testing, and the same for holding the 'targets.' Base testing yeilded around a 44-45% success rate, which is pretty good for a start.
When setting up the model, you can set various hyperparameters that affect how the model goes about sorting data. The following is the hyperparameters we used in the final version of our model:

• n_neighbors (the number of 'neighbors' in similar data values it checks for trends): 134 (so it checks the 134 data points in the training set with the most similar values to make it's decision on which genre a song is)
• algorithm (which of 3 possible KNN methods it uses to check the data): ball_tree (which orgainizes the data into 'balls' that it puts the data points into based on what balls it's neighbors in terms of data values have gone into; each 'ball' represents a genre in this case)
• weight (how all of the 'neighbors' are weighed in the internal decision process of the algorithm): distance (meaning 'neighbors' that have more similar data values have more of an impact on which 'ball' a data point will be sorted into)
• p (the power parameter of the Minkowski distance metric, which decides if the internal decision process should pay more attention to small changes in data or large changes): 1 (meaning it will pay more attention to small changes; otherwise known as the Manhattan distance type)
• leaf size (how many data points can be checked per 'leaf node' in the decision process; a higher number can reduce how long it takes for the algorithm to run, but requires more memory and may reduce accuracy as a tradeoff): 1 (meaning it may take longer to run than it possibly could, but takes up less memory and is likely more accurate as a result)

Model Results

The finalized model averages about an upper-46-to-mid-47% success rate, with this particular run managing about 47.65%. Considering a Naive Model, which is the baseline of whether or not a classification model has is at the very least ok, as it is simply a 1/n chance of success where n is the number of possible classifications (in this case genres), has a 1/21 (or about a 4.76%) chance of being correct, the model averages out to be about 10x better on average.