T-ECD: A Large-Scale Cross-Domain E-Commerce Dataset for Industrial Recommender Systems

Abstract

We introduce T-ECD, an open large-scale dataset for recommender systems research, containing over 135 billion interactions from 44 million users and 30 million items. The dataset is derived from anonymized real-world data within a unified banking ecosystem and spans five interconnected e-commerce domains: marketplace, retail, payments, promotional offers, and reviews. The core value of T-ECD lies in its rich integration of implicit feedback (e.g., clicks, purchases), explicit feedback (e.g., ratings, review embeddings), and contextual signals (e.g., transaction receipts, item embeddings). Crucially, the dataset maintains cross-domain consistency through aligned user and brand identifiers, enabling research on transfer learning, cross-domain recommendation, and multi-task modeling under realistic industrial conditions. To demonstrate the utility of T-ECD, we conduct benchmark experiments using a range of recommendation models, including matrix factorization and sequential models, evaluated in both single-domain and cross-domain settings. The results establish initial performance baselines and highlight the challenges of modeling in a rich, multi-domain environment. By making T-ECD publicly available, we aim to narrow the gap between academic research and production systems and to foster innovation in large-scale, multi-domain recommendation. To accommodate diverse research needs, T-ECD is released in both large-scale and reduced-size versions.