Notes

Part 2.1

○ Users’ tastes and musical needs are highly dependent on a multitude of factors
○ They argue to satisfy user needs they need to consider: intrinsic, extrinsic, and context of listeners
§ Intrinsic - Personality and emotional state
§ Extrinsic - Weather, surroundings, place of interest
○ Particularities
§ Duration
□ Songs last 3-5 min typically
§ Magnitude
□ Millions of music pieces, streaming services have giant catalogues
§ Sequential Consumption
□ Music is listened to in rapid succession vs movies which is a 1 time thing
§ Recommending what have been previously recommended
□ Users may appreciate it
§ Consumption Behavior
□ Music is typically passively listened to, but can affect preference elicitation. Using skipping as a tool to infer that the user doesn’t like a song is bad because it’s just background noise to them and not active listening
§ Listening Intent and Purpose
□ 3 Fundamental Intents of Listening to Music
® Self Awareness
◊ Reflective
} Who they are
} Who they would like to be
® Social Relatedness
◊ Identity
} Feel Close to Friends
} Express Identity and Value to Others
® Arousal/Mood Regulation
○ Emotions Music
§ MER = Music Emotion Recognition
§ MER’s can be used to categorize music by emotion terms, the task of actually doing that is difficult for 3 reasons
□ MER commonly neglects the distinctions between Intended Emotion (by the artist) and Perceived Emotion (which is recognized by user while listening)
□ You don’t know if the user is trying to enhance or modulate their mood
□ Emotional changes occur within the same music piece, but tags are extracted for the piece as a whole.

Part 2.2 - The “Cold Start” Problem

○ New users don’t have sufficient data for recommendations
○ No one rates - The given number of ratings is much lower than the number of possible ratings - Sparsity (High Sparsity means low coverage/not enough info)
§ Yahoo Data Musicset
□ Sparsity is 99.96%
§ Netflix Dataset
□ Sparsity of only 98.82%

Part 2.2 - Tackling the “Cold Start” Problem
○ Content Based Recommendation (CB)
§ Doesn’t require ratings
§ When a new item is added to catalogue, they extract features from it and make recommendations based off that
○ Mentions Pandora’s Music Genome Project
§ Pandora uses up to 450 descriptors per song
□ “abstract lyrics, unusual harmonies, aggressive vocalist, etc…”
§ Helps tackle the “New Item” Problem
○ Hybridization
§ CB Component
□ Computes acoustic features like
® Rhythm
® Timbre
® Pitch
§ The CB then assists the CF (Collab. Filtering)
○ Active Learning
§ Not personalized because it doesn’t learn the context and all other humanly situations

Part 2.3 - APC (Automatic Playlist Continuation)
○ APG (Automatic Playlist Generation)
○ APC task is to infer intended purpose of playlist
○ Playlists account for 31% of music listening
○ Albums account for 22%
○ Single tracks account for 46%

○ Making a playlist for a certain mood (Train music, Party Music, etc) helps the APC and APG help link and make connections for recommendations for others with similar tastes
○ A mix of familiar songs + unknown songs are an important requirement - serendipity
○ How do we bring tech to understand sociological and psychological factors