MovieLens 32M Data Adapted for Graph Database Import

Adapted from the MovieLens 32M dataset, this collection includes user ratings, movie metadata, and tags. The data has been restructured into CSV files for efficient import into graph database systems like TigerGraph. The adaptation was created by Zachary Zhao and hosted on Harvard Dataverse, with a repository last updated in March 2026.

Use Cases

• Benchmark graph database import and query performance using the `genre_vertices.csv` and `has_genre.csv` edge files.
• Train graph neural network models for movie recommendation using the `clean_ratings_df.csv` file containing `userId`, `movieId`, and `rating` features.
• Analyze user tagging behavior and movie-tag associations via the `tags_with_id.csv` vertex file and the `TAGGED` and `TAGGED_MOVIE` edge files.
• Study multi-genre movie classification by leveraging the split genre data derived from the original `movies.csv` file's pipe-delimited `genres` column.

Strengths

• Derived from the established MovieLens 32M dataset, which contains tens of millions of ratings.
• Data is explicitly structured for graph databases, with separate vertex and edge CSV files like `has_genre.csv` and `TAGGED_MOVIE`.
• Pre-processing steps include handling null values, generating unique IDs for tags, and converting UNIX timestamps to a human-readable format.

Limitations

• The exact row counts for the derived files (e.g., `clean_ratings_df.csv`) are not provided, making scale assessment difficult.
• The adaptation may introduce biases or errors from the data transformation process, such as the splitting of multi-genre entries.
• The temporal coverage of the underlying MovieLens data is not specified, which could affect the relevance of trends.

Provenance

Source

Adapted from the MovieLens 32M dataset (grouplens.org).

Collection Method

Original data was transformed using Spark for graph database compatibility, including splitting genres, generating IDs, and converting timestamps.

Freshness

Repository last updated March 2026; however, the temporal coverage of the source MovieLens 32M data is null.