Researchers from T-Technologies presented the T-ECD dataset at the main conference of ACM KDD 2026 in South Korea. The dataset contains over 135 billion anonymized interactions from five domains of a banking ecosystem — marketplace, retail, payments, special offers, and reviews — and is published on Hugging Face under the Apache 2.0 license. The report showed that cross-domain data improves recommendations by 31.5% for promotional campaigns but worsens them by 25.5% for food products, raising questions about the universality of multi-domain models.

image
image

What happened

The technical report on the T-ECD dataset (T-Tech E-commerce Cross-Domain Dataset) was included in the main program of the 32nd ACM SIGKDD 2026 conference. The dataset was collected based on data from the T-Bank banking ecosystem and includes 44 million users, 30 million items, and five interaction domains. 92.6% of users interact with at least two domains. After synthetic anonymization, the absolute value of Recall@200 decreases by 20–35%, but the relative ranking of models is preserved. The benchmarks cover models from ALS to HSTU and show the effect of cross-domain data: an improvement of 31.5% for promotional campaigns, 13.3% for non-food items, and a drop of 25.5% for food products. The dataset is available in a full version (2.81 TB) and a small version (~1 billion interactions, ~3.5 million users) on Hugging Face, and the benchmark code is in the tecd-baselines repository.

Context

Open industrial-scale datasets for recommender systems are rare. Existing datasets like Amazon Review Data or Taobao are often focused on a single domain or do not provide a fully reproducible pipeline with code. T-ECD ranks among the top open industrial datasets, comparable in size to Amazon Review and Taobao. Publication in the main program of ACM KDD — the leading conference on data mining and knowledge discovery — confirms the scientific significance of the work. The Apache 2.0 license does not impose restrictions on commercial use, which distinguishes T-ECD from datasets with educational licenses.

Why this matters for the industry

T-ECD creates an industrial benchmark for cross-domain personalization. The cross-domain structure (92.6% of users in 2+ domains) opens up tasks for transfer learning and multi-domain personalization, which are relevant for MultiApp platforms and ecosystem products. The result of negative transfer of up to −25.5% for FMCG is critical for industrial applications: it shows that cross-domain data does not always improve recommendations, contradicting the simplified narrative that "more data is better." This poses the task of adaptive cross-domain fusion for research teams, taking into account domain-specific characteristics. For RecSys startups, open access to the dataset reduces R&D costs and provides a baseline against which their own models can be validated.

Why this matters for users

The small version of the dataset (~1 billion interactions, ~3.5 million users) works without a GPU — it can be downloaded from Hugging Face and used on a CPU for hypothesis testing and training recommender models. The tecd-baselines repository with benchmark code (ALS, NCF, SASRec, HSTU) allows reproducing the results from the report and comparing their own models with an industrial standard. For students and research groups, this provides instant access to realistic industrial data without the need to deploy GPU infrastructure.

What is still unknown / limitations

The benchmark set covers four models from different generations (ALS, NCF, SASRec, HSTU), but modern architectures based on graph neural networks and transformers, such as LightGNN, SimGCL, and B4Rec, are missing. This limits the scientific rigor of claims about the reproducibility of results. The absolute drop in Recall@200 by 20–35% after anonymization indicates a loss of information that may be systematic — the methodology of synthetic anonymization and its impact on specific user groups require further study.

Sources

Author

Look at AI, editorial team