Datasets:

TOPAPEC
/

cross-domain-recsys-interactions

pretty_name: Cross-Domain RecSys Interactions (15 sources, ~883M)
license: other
language:
  - en
  - zh
  - 'no'
tags:
  - recommendation
  - sequential-recommendation
  - cross-domain
  - collaborative-filtering
  - implicit-feedback
size_categories:
  - 100M<n<1B
task_categories:
  - other
configs:
  - config_name: default
    data_files: data/interactions_v*.parquet

Cross-Domain RecSys Interactions — 15 sources, integer-indexed

One unified, integer-indexed interaction corpus glued from 15 recommendation sources for foundation / sequential recommenders. Built in 3 additive versions that share one schema and one global index space, so reading all of them together = the full corpus, collision-free. Interactions only — no model, no embeddings. Item text is provided separately as title+description metadata, joinable 1:1 by item_idx.

Full corpus: 882,999,680 interactions · ~52.6M users · ~12.1M items.

Versions (read all three for the full corpus)

version	file	interactions	sources	feedback
v1 — base (web/review)	`data/interactions_v1.parquet`	706,928,307	amazon, goodreads, movielens, yelp, mind, steam	rating / read / click / recommend
v2 — sequential	`data/interactions_v2.parquet`	60,093,215	hm, amazon_m2, foursquare(Gowalla), trivago, lfm(Last.fm-1k), adressa, foodcom	purchase / click / checkin / play / rating
v3 — chinese short-video	`data/interactions_v3.parquet`	115,978,158	pixelrec(Pixel8M), microlens(MicroLens-1M)	comment / click

Versions are append-only & index-continuous: v2 indices start at max(v1)+1, v3 at max(v1∪v2)+1, so user_idx/item_idx ranges are disjoint across versions and a plain union never collides. Each version was built independently (its own k-core), and earlier versions are never rewritten.

Per-source breakdown (interactions in train)

amazon 441,756,006 · goodreads 226,413,879 · pixelrec 106,893,985 · movielens 31,904,072 · hm 28,384,429 · amazon_m2 16,543,664 · microlens 9,084,173 · yelp 5,718,931 · foursquare 4,369,335 · trivago 4,198,784 · lfm 3,218,125 · adressa 2,783,832 · mind 1,104,003 · foodcom 595,046 · steam 31,416.

Files (uniform `_v{1,2,3}` naming; all parquet, homogeneous per group)

data/interactions_v{1,2,3}.parquet   # the corpus (8 cols, identical schema) — glob all 3
maps/user_map_v{1,2,3}.parquet       # user_uid  <-> user_idx
maps/item_map_v{1,2,3}.parquet       # item_uid <-> item_idx <-> (source, native_item)  [meta bridge]
meta/meta_v{1,2,3}.parquet           # item_idx, source, native_item, title, description
valid_items/valid_items_{source}.parquet   # native_item, item_uid, n_inter  (per source, the meta entrypoints)
stats/*.json                         # per-phase build stats

v1=base, v2=sequential, v3=chinese (suffix is consistent across interactions, maps and meta).

Field meanings

data/interactions_v*.parquet (one row = one user→item interaction)

field	type	meaning
`user_idx`	int32	dense user id, global across versions (disjoint ranges), ordered by descending frequency within its version
`item_idx`	int32	dense item id, global across versions; the key to join `meta`/`item_map`
`timestamp`	int64	event time, unix seconds UTC. Order-only source amazon_m2 uses a synthetic monotonic per-session step
`rating`	float32	rating normalized to [0,1] per source (implicit positives = 1.0)
`rating_raw`	float32	original source value (e.g. 5.0 stars, Food.com 0–5; 1.0 for implicit)
`feedback_type`	int8	0 explicit_rating · 1 implicit_click · 2 implicit_recommend · 3 implicit_read · 4 implicit_purchase · 5 implicit_checkin · 6 implicit_play · 7 implicit_comment
`domain`	int16	sub-domain: amazon category 0–32; other sources 100–115 (one per source/sub)
`source`	int8	0 amazon · 1 goodreads · 2 yelp · 3 movielens · 4 mind · 5 steam · 6 amazon_m2 · 7 hm · 8 adressa · 9 trivago · 10 foursquare · 11 lfm · 12 foodcom · 13 microlens · 14 pixelrec · 15 douban

Rows are sorted within each user by (timestamp, item_idx) — a stable tie-break so sequence slicing is reproducible even at daily timestamp granularity. No train/test split is applied (do the temporal cutoff at train time).

maps/user_map_v*.parquet: user_uid (str, namespaced "{source}:{native_user_id}") ↔ user_idx. maps/item_map_v*.parquet: item_uid ↔ item_idx ↔ (source, native_item). native_item is the raw item key in the source (parent_asin / book_id / business_id / movieId / article_id / …) — the join key for external metadata. meta/meta_v*.parquet: item_idx, source, native_item, title, description (title+description only). Languages: v1/v2 mostly English (amazon_m2 multilingual, adressa Norwegian); v3 = Chinese. valid_items/valid_items_{source}.parquet: native_item, item_uid, n_inter — the per-source item lists that survived the k-core; entry points to filter+index any external metadata.

Metadata coverage

meta/meta_v1.parquet — 10,331,150 items (100% of v1), ~3.3 GB.
meta/meta_v2.parquet — 920,925 items (text-bearing v2 sources: amazon_m2, hm, foodcom, lfm), ~104 MB. Textless v2 sources have no item title/description upstream and are intentionally absent: adressa (url only), trivago (amenity properties only), foursquare/Gowalla (geo only).
meta/meta_v3.parquet — 471,446 items (100% of v3; Chinese), ~39 MB.

Usage

# Full corpus (all 3 versions) with 🤗 datasets — schemas are identical so they load as one table
from datasets import load_dataset
ds = load_dataset("TOPAPEC/cross-domain-recsys-interactions", split="train", streaming=True)
print(next(iter(ds)))

# Local DuckDB: glob all versions + attach titles by item_idx (download large files first)
import duckdb
from huggingface_hub import snapshot_download
d = snapshot_download("TOPAPEC/cross-domain-recsys-interactions", repo_type="dataset",
                      allow_patterns=["data/*", "meta/*", "maps/*"])
con = duckdb.connect()
con.sql(f"""
  SELECT m.source, m.title, count(*) n
  FROM   read_parquet('{d}/data/interactions_v*.parquet') t
  JOIN   read_parquet('{d}/meta/meta_v*.parquet')         m USING (item_idx)
  GROUP  BY 1,2 ORDER BY n DESC LIMIT 10
""").show()

Metadata-join contract (for attaching your own metadata)

Per source: meta_raw → filter(native_item IN valid_items_{source}) → join(item_map_v* on (source, native_item)) → keyed by item_idx. Guarantees (asserted in the build tests): every item_idx in the corpus has exactly one item_map row; every (source, native_item) resolves to one item_uid; no uid collisions across sources; version index ranges are disjoint.

Provenance & licenses

Derived ETL (interactions + title/description only) from: Amazon Reviews 2023 (McAuley-Lab), Goodreads (UCSD), MovieLens 32M (GroupLens), Yelp Open Dataset, MIND (Microsoft), Steam (McAuley), Amazon-M2 (KDD Cup 2023), H&M (Kaggle), Adressa (NTNU), Trivago (RecSys'19), Gowalla (SNAP), Last.fm-1k, Food.com (Kaggle), MicroLens & PixelRec (Westlake). You must comply with each source's original license/terms. Only anonymized interaction tuples + item title/description are redistributed.