RelBench dbinfer datasets

This repository hosts the dbinfer family of relational datasets in the RelBench 3.0 manifest format, one subdirectory per dataset. The datasets originate from the 4DBInfer benchmark (data version 20240304), built directly from the original archives that dbinfer_bench itself downloads (https://data.dgl.ai/mtbench/20240304-<name>.tar). Labels are the source's own, served as-is (every task has kind: external).

Each subdirectory is a self-describing RelBench dataset (manifest.yaml + plain db/*.parquet + tasks/<task>/); open its schema.svg for a zoomable entity-relationship diagram.

Datasets

dataset	domain	tables	tasks
dbinfer-amazon	E-commerce (reviews)	3	churn, purchase, rating
dbinfer-avs	Retail (Acquire Valued Shoppers)	8	repeater
dbinfer-diginetica	E-commerce (sessions)	12	ctr, purchase
dbinfer-outbrain-small	Content recommendation	9	ctr
dbinfer-retailrocket	E-commerce (behaviour)	7	cvr
dbinfer-seznam	Digital advertising	4	charge, prepay
dbinfer-stackexchange	Online community (Q&A)	9	churn, upvote

Table counts include the single-column key tables materialized for the foreign-key targets that 4DBInfer declares without a payload of their own (Item, Visitor, Customer, Chain, Brand, Category, Company, Session, User, Orders, Token); see each dataset card.

dbinfer-outbrain-small has almost no referential integrity, in the source. 4DBInfer subsampled its tables independently, so ~99.9% of its foreign keys -- and all but 58 of 69,543 distinct train entities of its ctr task -- point at rows that were not kept. The previous revision hid this behind all-null keys. There is no full-size outbrain archive upstream. Its card has the numbers.

Several tasks derive their label from a column that is also in the database (retailrocket/cvr <- View.added_to_cart, seznam/charge <- Dobito.sluzba, seznam/prepay <- Probehnuto.sluzba, outbrain-small/ctr <- Click.clicked, amazon/rating <- Review.rating). Each such task declares remove_columns; the column is kept in db/ so the database stays faithful to 4DBInfer, and Dataset.get_db drops it, so relbench.load_task(...) hands you a graph without it. If you read the parquet directly, drop it yourself.

Loading

import relbench
ds = relbench.load_dataset("relbench/dbinfer/dbinfer-diginetica")
task = relbench.load_task("relbench/dbinfer/dbinfer-diginetica", "ctr")
db = ds.get_db()
train = task.get_table("train")

See the RelBench CONTRIBUTING guide for the manifest layout.

Provenance and revision history

Generated by provenance/dbinfer.py, which pins the sha256 of each source archive; verified by provenance/check_dbinfer.py. Port decisions -- which columns are kept, how implicit key domains are materialized, how keys are reindexed -- are documented in that file's module docstring, and each dataset card lists the payload columns it drops.

This collection was rebuilt from the original 4DBInfer archives. The previous revision was derived from pre-built db.zip artifacts produced by the upstream dbinfer-relbench-adapter export pipeline, which corrupted the data in ways that were not visible from the schema:

• Every declared foreign key was 99.9-100% null. The adapter validated foreign keys against len(parent_table) where the parent was a {column: array} dict -- i.e. against the parent's column count -- so all larger key values were set to NaN. None of the seven databases could be joined.
• Real primary keys were overwritten with np.arange(n) before that validation, so the key correspondence was already gone.
• Foreign keys to non-materialized key domains were dropped, amputating the schema (retailrocket/View was left with no foreign keys at all).
• Task entity ids were remapped through a task-local mapping unrelated to the database's, so label rows did not reference the database.
• Task time columns were destroyed (1970-01-01 00:00:00.000000022) by casting the source's integer-valued columns with astype('datetime64[ns]').
• Classification targets were silently relabelled by sorted-string order, and task_type was inferred from target cardinality -- turning amazon/rating, a 4DBInfer regression/RMSE task, into multiclass, and both retrieval tasks into multiclass.
• Undeclared payload columns were published, including Posts.Score (from which stackexchange/upvote's label is derived), History.repeater (avs/repeater's label itself), and full-history aggregates like Users.Reputation/UpVotes/Views.
• val_timestamp / test_timestamp fell after the end of the data, so get_db(upto_test_timestamp=True) trimmed nothing and gave no temporal protection at all (dbinfer-retailrocket claimed 2015-09-21 against a last event of 2015-09-18; dbinfer-diginetica claimed 2016-11-12; dbinfer-seznam 2015-10-04 against 2015-10-01). They also did not bracket the source's own splits -- seznam's val labels start at 2015-04-01 and its test labels at 2015-07-01.

The current revision fixes all of the above: foreign keys resolve, primary keys are dense, implicit key domains are materialized as tables, task entity columns index their entity table, time columns are real timestamps, targets keep their source values, task_type follows the source, and each task declares remove_columns for any database column its label is derived from. Results computed against the previous revision are not comparable.

Citation

These datasets are from the 4DBInfer benchmark. If you use them, please cite:

@inproceedings{wang2024fourdbinfer,
  title     = {{4DBInfer}: A {4D} Benchmarking Toolbox for Graph-Centric Predictive Modeling on Relational Databases},
  author    = {Wang, Minjie and Gan, Quan and Wipf, David and Cai, Zhenkun and Li, Ning and Tang, Jianheng and Zhang, Yanlin and Zhang, Zizhao and Mao, Zunyao and Song, Yakun and Wang, Yanbo and Li, Jiahang and Zhang, Han and Yang, Guang and Qin, Xiao and Lei, Chuan and Zhang, Muhan and Zhang, Weinan and Faloutsos, Christos and Zhang, Zheng},
  booktitle = {Advances in Neural Information Processing Systems 37 (NeurIPS 2024) Datasets and Benchmarks Track},
  year      = {2024}
}