Assignment #

Make Recommendations for Users #

• Take the current diversity of users
• generate a lot of tracks for them
  • if we want stuff "further" from preference, we use a track from them as seed, then use one of the results as seed, and so on
  • if we want closer, we just use the 'root tracks'
• compared to their current diversity, generate one list that's higher and one that's lower

Present Recommendations Back to Users #

We can use the spotify api to make a playlist for each user

• present them with a more diverse and a less diverse playlist

Analyze Feedback from Users #

• get it back in a google form or smth like that
• check their satisfaction of high and low diversity lists against user metrics (sophistication, preference)

How to use the API (or another module) #

1. Place/fill an .env file in the implicit directory.
2. Place the credentials.py file in the modules directory.

To import them in any file, use the following lines of code (API as example):

import sys sys.path.append("./modules/") from API import API

Where the file path in the append() method should point to the modules directory.

TODOs #

Work #

1. Generate lots of songs to build lists from - V
2. Build diversity score based lists - K
3. Present playlists back to users - A
   DEADLINE DEC1
4. Analyize insights - G
   DEADLINE DEC6

Presentation #

all hands on deck

Report #

• Intro - A
• Related work - A
• Method - V, K
• Results - G
• Discussion - K, V

Qs for Martijn #

• what metrics to use when asking users for feedback
  • comparative questions or objective questions?
    • we should use objective questions
  • USE EXISTING QUESTIONNAIRES from [@ekstrandUserPerceptionDifferences2014] and [@liangPersonalizedRecommendationsMusic2019] papers
    • Use 2-3 that had high factor loading
  • diversity questions from [@heInteractiveRecommenderSystems2016] paper discussed recently
• diversifying on stuff other than genre
  • diversifying on song characteristics / features is a better idea, just be careful not to go too far
    • plot a contour plot of the feature values to see how diverse we are
    • check users given feature preference variance, and stay close to those
    • try it on ourselves first
  • using artists might be better than genre