Dataset Card for Lastfm-VADS

This dataset contains Valence, Arousal, Dominance and Sentiment Ratio values for over 800K tracks, with their respective artist, album and ranked tags.

Dataset Details

This dataset was curated to be used in a Bachelor Thesis, consisting in the integration of sentiment features into music recommendation along with inherent track features, such as artists, albums, interaction timestamps and (automatically assigned) ratings.

Track data was gathered with last.fm's API, and sentiment features were extracted from tags, with a sentiment analyzer that used generalized Wikipedia definitions.

Dataset Description

• Curated by: Javier Wang
• Language(s) (NLP): English

Dataset Sources

• Repository: https://huggingface.co./datasets/Acervans/Lastfm-VADS
• Demo: https://github.com/Acervans/lastfm_RS

Uses

Features may be integrated into any ML model, specifically RecSys models, to evaluate the model performance considering the included types (track-inherent or sentiment). These files are structured to be used directly within the RecBole framework, specifically for context-aware models if all features need be integrated.

Dataset Structure

This dataset has three atomic files, all of which use tabs as field separators:

• lastfm_recbole.inter - Contains user-track interactions, organized in:
  • user_id: User ID.
  • track_id: Track ID.
  • rating: Rating assigned to the track based on the type of interaction of the user, keeping the highest if more than one.
  • timestamp: Timestamp of the interaction.
• lasftm_recbole.item - Contains item features, organized in:
  • track_id: Track ID.
  • tags: Tags as sequential token, repeated by user assignment count, and separated by a space.
  • artist_id: Artist ID.
  • album_id: Album ID.
  • v: Valence score for the track.
  • a: Arousal score for the track.
  • d: Dominance score for the track.
  • stsc: Sentiment Ratio for the track.
• lastfm_recbole.user - Contains users' IDs. This file was kept to have the same IDs assigned during processing.

File lastfm_data.tar.gz contains the raw files scraped with last.fm's API, distributed in several JSON and DAT files. Please check the Readme.txt inside for the structure of these files.

Source Data

The source data used for this dataset is Last.fm's API. The text source used to analyze sentiment (tag definitions) was obtained from Wikipedia's API.

Data Collection and Processing

This dataset was obtained through a simple web scraper. It retrieves the top 50 chart tags using Last.fm's API, representing the most listened tags at the time. For each tag, we obtained the top unique artists associated with them and the top-30 unique listeners. Then, we used the API to gather data from the top listeners, acquiring their top 20 tracks, recent tracks, and loved tracks, each with the corresponding timestamp, and artist and album information. Additionally, we collected the top 10 artists and albums for each listener. Finally, we fetched the top 10 tags assigned by users to each unique track, artist, and album, which were associated definitions according to the summaries retrieved from Wikipedia's API.

For processing, ratings in the .inter file were assigned based on listen count, type of interaction (love, recent, top) and ranking. All tags in the .item file were preprocessed to reduce ambiguities, removing spaces and dashes and converting to lowercase. These tags were then repeated based on user assignment count, to increase the weight given for each track. Finally, unique IDs were assigned for each track, user and album comply with Data Protection Principles.