CausalRec: A Multi-Domain, Multi-Confounder Benchmark Dataset for Debiasing and Cold-Start Recommendations

CausalRec is a large-scale synthetic benchmark dataset containing 2.5 million interactions explicitly designed to bridge the gap between causal debiasing and cold-start optimization in Recommendation Systems (RecSys).

Unlike standard datasets that merge all biases into a single observable signal, CausalRec tracks 5 distinct system and context confounders as isolated variables alongside the absolute unpolluted ground-truth preference ($R^*$).

📊 Dataset Structure

The benchmark is split into two distinct, structured interaction environments:

1. Fast-Feed (1.5M rows): Simulates high-velocity infinite scroll apps (e.g., Short Video platforms). Highly distorted by Position Bias ($C_1$) and Temporal Bias ($C_5$).
2. E-Commerce (1.0M rows): Simulates high-intent marketplace catalogs. Highly distorted by Popularity Bias ($C_2$), Selection Bias ($C_3$), and Promotion Bias ($C_4$).

Data Fields

Each row represents a single simulated exposure event containing:

Column Name	Type	Description
domain	String	Environment source (Fast_Feed or E_Commerce)
user_id	Integer	Unique identifier for the user ($N=10,000$)
item_id	Integer	Unique identifier for the item ($N=5,000$ per domain)
user_type	String	Cohort classification: Warm or Cold (10% structural User Cold-Start pool)
item_type	String	Cohort classification: Warm or Cold (10% structural Item Cold-Start pool)
true_preference_r_star	Float	Ground Truth ($R^$)*: Pristine, unpolluted dot product of latent vectors
c1_position_bias	Float	Confounder 1: Layout presentation depth (Exponential scale)
c2_popularity_bias	Float	Confounder 2: Global Power-Law/Zipfian distribution impact
c3_selection_bias	Float	Confounder 3: Algorithmic exposure filter based on user profile historical volume
c4_promotion_bias	Integer	Confounder 4: Paid advertisement injection indicator (0 = Organic, 1 = Sponsored)
c5_temporal_bias	Float	Confounder 5: Cyclical continuous periodic wave variation
y_observed	Integer	Target Label ($Y$): The final corrupted, observable click outcome (0 or 1)

🚀 Quick Start (Usage)

You can load this benchmark instantly into any Python machine learning pipeline using the Hugging Face datasets library:

from datasets import load_dataset

# Load the benchmark dataset
dataset = load_dataset("alihassan1437/CausalRec")

# Inspect train/test structure
print(dataset)