General Commerce Intelligence: Glocally Federated NLP-Based Engine for Privacy-Preserving and Sustainable Personalized Services of Multi-Merchants

Kyoung Jun Lee, Baek Jeong, Suhyeon Kim, Dam Kim, Dongju ParkAAAI 2024 · DOI

Introduction

Small and Medium-sized Enterprises (MSMEs) often find themselves in a "data dilemma." To compete with major retail platforms, they require sophisticated AI-driven services like personalized recommendations and demand forecasting. However, building these models typically requires massive datasets that individual small businesses do not possess. While they could turn to Big Tech providers, doing so often necessitates surrendering proprietary customer data, leading to a loss of business autonomy and privacy concerns.

Furthermore, traditional recommendation systems are built on specific Product IDs. This approach works within a single, unified store but fails in multi-merchant environments where every shop uses its own naming conventions and taxonomies. To address these challenges, this paper introduces General Commerce Intelligence (GCI), a system designed to be privacy-preserving, scalable across diverse merchants, and accessible to smaller players.

The GCI Framework

The GCI engine moves away from rigid ID-based systems by treating commerce data—specifically product and merchant names—as natural language. This allows the system to leverage the semantic power of modern NLP architectures.

Overview of the General Commerce Intelligence (GCI) architecture, utilizing federated learning to share AI models across merchants while keeping transaction data local.

The framework relies on two primary technical innovations:

1. TransformRec: NLP-Based Recommendation

Instead of mapping a user's history to a sequence of IDs, GCI uses TransformRec, a Transformer-based architecture that analyzes the semantic relationships within purchase histories. By tokenizing product names, the model can generalize across different merchants. For instance, by breaking "Seafood Pizza" into "Seafood" and "Pizza," the model learns that a customer interested in this item might also appreciate "Seafood Pasta" from an entirely different vendor.

Comparison of recommendation logic without and with tokenization, demonstrating how tokenizing product names allows the model to find semantic links across different merchants.

2. Glocalized Federated Learning

To enable cross-merchant learning without compromising privacy, GCI employs a "glocalized" (global + local) Federated Learning (FL) approach. In this setup, only the internal layers of the Transformer—which capture general linguistic and behavioral patterns—are synchronized globally across the network. The embedding and output layers remain local to each merchant. This strategy prevents "parameter explosion" and ensures the model remains finely tuned to the specific vocabulary and inventory of each individual store.

Experimental Results and Real-World Impact

The researchers validated GCI using real-world transaction data from major South Korean partners, including Ulsan Pedal (a food delivery app with ~200,000 entries) and Genesis BBQ (a major franchise with ~350,000 entries).

Key findings include:

  • The Power of Tokenization: TransformRec with tokenization achieved a Hit-Rate (HR@10) of 7.329%, a significant improvement over the 2.182% achieved without it.
  • Federated Superiority: Surprisingly, the Glocal FL approach outperformed centralized "Data-Shared" models. This suggests that the federated approach helps the model avoid overfitting on the noise inherent in massive, aggregate datasets.
  • Product Brainstorming: Because the model is NLP-based, it can suggest novel product names. Genesis BBQ successfully launched a "Rose Sauce Fried Chicken Set" after the AI "hallucinated" it as a potential high-demand item.

In a live deployment on the Ulsan Pedal platform, the system achieved a 6.32% hit rate. For comparison, industry estimates place Amazon’s recommendation accuracy at approximately 5%, highlighting GCI's effectiveness in a real-world, multi-merchant setting.

Conclusion

General Commerce Intelligence demonstrates that treating commerce as a language allows for powerful, cross-merchant AI services that respect user privacy. By combining NLP with glocalized federated learning, GCI empowers small businesses to provide high-quality personalized experiences without sacrificing their data or autonomy. Future work will focus on integrating Generative AI for conversational commerce and addressing the "cold start" problem for new merchants entering the network.

Summary updated

BibTeX
@inproceedings{lee2024general,
  author = {Kyoung Jun Lee and Baek Jeong and Suhyeon Kim and Dam Kim and Dongju Park},
  title = {{General Commerce Intelligence: Glocally Federated NLP-Based Engine for Privacy-Preserving and Sustainable Personalized Services of Multi-Merchants}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i21.30309},
}
Figure