Music and Machines by Mira Savchuk

Machine Learning

mirasavchuk — 2025-05-12 16:59:36 UTC

A technology type that teaches computers to learn from data and make decisions without exactly being told so. Some examples are:

Email spam filters
Siri/Alexa speech recognition
Recommendation systems

Types of Filtering Methods

mirasavchuk — 2025-05-12 17:14:20 UTC

Content-based filtering recommends things by comparing them to what a user has liked in the past. It looks at different features like genres or keywords.

Collaborative filtering recommends things based on what similar users liked. However, it doesn’t look at its details—just patterns.

mirasavchuk — 2025-05-12 17:18:37 UTC

Why 23 songs?

mirasavchuk — 2025-05-13 16:37:16 UTC

We rated 23 songs to create data for a machine learning model. These ratings help the system learn our music preferences. Then it can recommend new songs each might like using methods like content-based or collaborative filtering.

TOP ATTRIBUTES

mirasavchuk — 2025-05-13 16:38:47 UTC

Danceability
Energy
Popularity
I chose those three because for me it's important for my song to be energized so people can dance to it. Also, popular, because most popular songs are very good.

Levels of Complexity

mirasavchuk — 2025-05-13 16:48:02 UTC

-If the model is too simple it may end up in not learning enough from the data, make poor predictions and miss patterns.

-Too complex models, have too many unnecessary details, works great on training data but fails on new data.

-Models that are just right test the model with new data to make sure it works well in different situations

One-Attribute Model

mirasavchuk — 2025-05-13 17:15:52 UTC

This graph gave a rough idea of the song rating. Points don't really line up and while its not bad the graph isn't quite accurate since it didn’t consider other important for me features. Test and train error was high and not close at all, making this model a bad fit.

Two-Attribute Model

mirasavchuk — 2025-05-13 17:18:07 UTC

This graph was much more accurate. Test and train error was low and close together. It included more features and was a better fit overall.

Class Rating Matrix

mirasavchuk — 2025-05-13 17:29:31 UTC

My ratings didn't perfectly match with anybody's. Coming at the highest similarity of 83% my ratings aligned with Syd, Landon, Alexis, Amaya, Jada, Lydia, Morgan and Dani

Predictions

mirasavchuk — 2025-05-13 17:37:48 UTC

Predictions #2

mirasavchuk — 2025-05-13 17:39:23 UTC

What do your results look like compared to your actual rating?
Some were quite close, very few were off
Were there any surprises? Why do you think that happened?
Yes, some predictions were much lower than my actual ratings, but it could be because those songs are personal favorites
What does this add to your understanding of collaborative filtering?
It can’t predict personal taste well and works great when many people rate the same things
How might you change your model to be more accurate?
Combine together with content based filtering and add more data

Data Ethics of Machine Learning

mirasavchuk — 2025-05-13 19:43:37 UTC

Machine learning is important for people to understand because everything around us uses it everyday. Starting with recommendation systems to medical tools. However, depending on the bias the outcome can change. For example, collaborative filtering may prefer popular items and ignore other preferences and content-based filtering can limit diversity by only recommending similar items

mirasavchuk — 2025-05-13 19:54:15 UTC

Data Ethics of Machine Learning

mirasavchuk — 2025-05-14 16:59:12 UTC

What public needs to know is that machine learning happens from data, meaning if data is biased, the result will be too. Collaborative filtering often shows popular things and follows what most people like, so it can miss new or different items. Content-based filtering shows similar things based on what you liked before, but it can miss variety if the item info is not complete or fair.