matcav/datacookers-processed

Preprocessed Data for the ADA Project 2025

By DataCookers

General Datasets

• channels_clean.parquet: A cleaned version of the df_channels_en split from the Youniverse Dataset. This dataset corrects parsing errors where missing values in the Category column caused subsequent columns to shift one position to the left.
• grammy_raw.parquet: The foundational dataset containing Grammy winners and nominees (1965–2024), sourced from Kaggle.
• grammy_metadata.parquet: A merged dataset combining the grammy_raw data with the yt_metadata_en split from Youniverse. It is further enriched with web-scraped song lyrics.
• grammy_channels.parquet: A filtered subset of df_channels_en containing only those YouTube channels that have published a Grammy-nominated or winning song.
• grammy_metadata_extended.parquet: An expansion of the metadata containing all videos belonging to the channels identified in grammy_channels (not just the Grammy-winning videos).
• grammy_timeseries.parquet: Temporal data (time series) specifically associated with the channels found in the grammy_channels dataset.
• music_metadata.parquet: A subset of the original yt_metadata_en Youniverse split, filtered to include only entries classified under the Music Category.
• music_video_ids.parquet: A lightweight dataset containing the specific video IDs filtered from the yt_metadata_en split, strictly for videos listed in the Music category.
• music_comments.parquet: A collection of user comments obtained by filtering youtube_comments.tsv.gz. It retains only comments posted on videos present in the music_video_ids dataset.

Collaborative Filtering Data

The following files represent the output of Matrix Factorization performed on the merged_comments.tsv.gz split.

Methodology: 1. We constructed a sparse interaction matrix of shape (users, items) with binary entries (0 or 1), where a 1 indicates the user commented on a specific video. 2. We performed Matrix Factorization using the Alternating Least Squares (ALS) algorithm via the implicit library.

Resulting Files:

• CF_item_factors.parquet: The latent space vectors representing every video_id (Item Factors).
• CF_user_factors.parquet: The latent space vectors representing every author (User Factors).
• CF_als_model.pkl: The serialized weights of the trained ALS model.
• CF_user_id_map.pkl: A dictionary mapping the original author (string) to the integer ID used during ALS model training.
• CF_item_id_map.pkl: A dictionary mapping the original video_id (string) to the integer ID used during ALS model training.