ABOUT THE COURSE...

1) Can we use the information provided in publicly available listening data to transform this information into the space defined by these alternative preference models?

>> Some of the attributes provided in the data were analogous to the dimensions of the alternative preference models (ex. "energy," "valence," "acoustic"...). However, others were more difficult to relate to the models. Given more time, it would likely be possible to write functions to more directly translate songs in the listening data to the alternative model spaces using information from the literature. However, due to time constraints, the attributes given from the data are used as is.

2) Given a representation of a user's musical preferences, can we classify/predict songs which they preferred (vs. others)?

>> I was able to classify user preferred vs. not preferred songs (as labeled by their relative playcounts) using various methods (see decision tree, naive Bayes, SVM, and NN sections), although classification accuracy overall was fairly low. SVMs achieved the highest classification accuracy of 61%.

3) Using this representation, can we further estimate the degree of preference (how much they like it) for a given item (song/artist/genre)?

>> Due to the relatively poor binary classification accuracies for preferred vs. not preferred songs, I did not attempt any regression models. However, future work could explore the use of linear regression or mixed modeling to model playcounts using preference distance (as well as other data, like popularity).

4) Which of the two genre-free multi-dimensional representations of preference yields the most accurate predictions/estimations? The three-dimensional VAL-ARO-DEP model, or the five-dimensional MUSIC model?

>> Although I wasn't able to transform the given data into these models exactly (due to time constraints), I did experiment with estimating preference using three vs. six attributes from the dataset. The six-attribute metric appeared to perform better overall (see SVM section).

5) What is the distribution of breadth of preferences in the dataset? Do the majority of recorded users have a wide or narrow stylistic range of musical preferences, or somewhere in between?

>> In the dataset overall there seems to be a wide distribution of musical styles, with more popular, "umbrella" styles appearing more frequently (ex. rock, pop, metal, dance...) (see "clustering," Figure 27). The average user in this dataset has just one song per unique artist in their data, which suggests a breadth of artists (see "data exploration," Figures 8 & 9). Within-user variances of the Spotify attributes suggests that there is also a fairly broad distribution of styles represented in each user's data (see dataset creation code).

6) Within either of the proposed models, is it more accurate to represent the user as a convex hull of their data points in the space, or as a single point (the centroid of the convex hull)? See Figure 2 below.

>> Due to time constraints, it was not feasible to attempt the more complex convex hull modelling. Centroid modelling was used throughout this project. The convex hull model is left as an implementation for future work.

7) How many data points (songs/artists/genres...) are needed to best represent a user?

>> This really varied from user to user. Because I labeled preferred vs. not preferred songs based on within-user playcount, the range of playcounts that a user had was more important to consider than just how many data points the user had. A higher range of playcounts provided more confidence that higher-playcount songs were in fact more preferred than lower-playcount songs.

8) What level of data granularity (songs vs. artists vs. genres...) best represents a user?

>> I used songs as the level of data granularity throughout this project. Since users in this dataset tended to not have multiple listed songs per artist, it didn't make sense to use this to capture user preferences. Genres could be an interesting level to explore, although the genre information provided by Spotify would need to be simplified in some systematic way order to do so. Then playcounts could be aggregated by genre within a user.

9) Which features/attributes are most important in answering the outlined questions? The mappings to the dimensions of these alternative models? Or are genre labels/other generic metadata actually sufficient?

>> Artist and track popularity metrics were found to improve preference classifications. Track duration, key, and mode were not found to be useful features for the model. The music style features "acoustic," "danceable," "energy," "instrumentalness," "speechiness," and "valence" were also found to improve classification. Some results also suggested that emphasizing "acoustic," "energy," and "valence" features provided additional classification improvement (see SVM section).​

10) Given the extreme availability of music nowadays ("old" or "new") via streaming platforms, is "year of release" a relevant/useful attribute to consider in addition to the other descriptors when representing musical preferences?

>> I tested the usefulness of the "year of release" attribute by building a model with and without this feature and comparing the classification accuracies. Year of release did not increase the classification accuracy of any of the models, suggesting that this attribute does not play a significant role in user's preferences within this data (see decision trees and naive Bayes sections).