MMG-Pop Benchmark Dataset

This dataset supports the MMG-Pop (Multimodal Graph-based Popularity Prediction) Benchmark.

Directory Overview

metadata/

Contains the full cascade files for each dataset. These files are the main source for labels, post content, and complete conversation structure.

thread_metadata_updated_added.parquet or reddit_*_metadata.parquet

One row per discussion tree or thread. Includes cascade-level metadata and final prediction targets, such as:

• width
• depth
• virality
• size (given as num_posts)
• unique users
• like score (given as score in Reddit datasets or like_count in Bluesky)

thread_posts_with_all_labels2.parquet or reddit_*_posts.parquet

One row per post or comment in the full cascade. Contains raw post text and post-level fields used to build node features and reconstruct the conversation tree.

user_degrees.csv

Bluesky-only auxiliary file mapping users to follower counts. Used to derive root-user follower features.

---

snapshots/

Contains early-window versions of the cascade graph. These are the forecasting inputs: what the model observes after only the first few minutes of cascade growth.

edges_{window}min.parquet

Reply edges observed within a specific early window. Defines the partial graph structure at that time.

meta_{window}min.parquet

Cascade-level snapshot records for the same window. Used to identify valid trees at that early stage and align them with the full metadata and post files.

---

splits/

Optional saved train, validation, and test splits for reproducibility.

train_ids_*.npy, val_ids_*.npy, test_ids_*.npy

Lists of cascade IDs assigned to each split for a given observation window.

---

future_horizons/

Optional files for future-horizon prediction beyond the default final-state prediction setting.

gt_4h.parquet, gt_8h.parquet, gt_16h.parquet, gt_24h.parquet

Ground-truth cascade outcomes measured at intermediate future times.

valid_mask.parquet

Marks which cascade/horizon pairs are valid and should be used for training or evaluation.

---

File Structure

bluesky/
  metadata/
    thread_metadata_updated_added.parquet
    thread_posts_with_all_labels2.parquet
    user_degrees.csv
  snapshots/
    edges_2min.parquet
    meta_2min.parquet
    edges_10min.parquet
    meta_10min.parquet
    splits/
      train_ids_2min.npy
      val_ids_2min.npy
      test_ids_2min.npy
      train_ids_10min.npy
      val_ids_10min.npy
      test_ids_10min.npy
    future_horizons/
      gt_4h.parquet
      gt_8h.parquet
      gt_16h.parquet
      gt_24h.parquet
      valid_mask.parquet

reddit/
  gaming/
    metadata/
      reddit_gaming_metadata.parquet
      reddit_gaming_posts.parquet
    snapshots/
      edges_20min.parquet
      meta_20min.parquet
      edges_50min.parquet
      meta_50min.parquet
      splits/
        train_ids_20min.npy
        val_ids_20min.npy
        test_ids_20min.npy
        train_ids_50min.npy
        val_ids_50min.npy
        test_ids_50min.npy
      future_horizons/
        gt_4h.parquet
        gt_8h.parquet
        gt_16h.parquet
        gt_24h.parquet
        valid_mask.parquet

  futurology/
    metadata/
      reddit_futurology_metadata.parquet
      reddit_futurology_posts.parquet
    snapshots/
      edges_90min.parquet
      meta_90min.parquet
      splits/
        train_ids_90min.npy
        val_ids_90min.npy
        test_ids_90min.npy
      future_horizons/
        gt_4h.parquet
        gt_8h.parquet
        gt_16h.parquet
        gt_24h.parquet
        valid_mask.parquet

  ama/
    metadata/
      reddit_ama_metadata.parquet
      reddit_ama_posts.parquet
    snapshots/
      edges_30min.parquet
      meta_30min.parquet
      splits/
        train_ids_15min.npy
        val_ids_15min.npy
        test_ids_15min.npy
        train_ids_30min.npy
        val_ids_30min.npy
        test_ids_30min.npy
      future_horizons/
        gt_4h.parquet
        gt_8h.parquet
        gt_16h.parquet
        gt_24h.parquet
        valid_mask.parquet