Research map / Federated learning and privacy / Privacy attacks and secure learning
General Commerce Intelligence: Glocally Federated NLP-Based Engine for Privacy-Preserving and Sustainable Personalized Services of Multi-Merchants
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds a Transformer-based NLP commerce recommender (GCI, with TransformRec) trained on merchant and product-name receipt data via glocalized federated learning that shares only internal encoder/decoder layers, and reports deployment on a food-delivery platform.
Problem
Predicting a customer's next purchase across multiple merchants without customer IDs or product IDs, while preserving privacy and letting MSMEs collaborate without sharing raw data or full models.
Approach
Treat purchased product and merchant names as text tokens for a small Transformer recommender (TransformRec), and train it with glocalized FedAvg that shares only the internal encoder/decoder layers while keeping vocabulary-dependent embedding and output layers local.
Key results
- Glocal FL achieved HR@10 of 0.478 and 0.522 on BBQ1/BBQ2 and 0.147 and 0.140 on Ulsan Pedal 1/2, beating local training (0.443, 0.455, 0.110, 0.117) and data-shared training (0.425, 0.404, 0.097, 0.099).
- TransformRec with tokenization reached HR@10 of 7.329% versus 2.182% without tokenization on Ulsan Pedal data.
- Live deployment on Ulsan Pedal recorded a 6.32% hit rate (1,891 of 29,933 purchases) from AI recommendations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add glocalized federated learning instead of local training | TransformRec · BBQ 1, BBQ 2, Ulsan Pedal 1, Ulsan Pedal 2 | HR@10: 0.443 -> 0.478; 0.455 -> 0.522; 0.110 -> 0.147; 0.117 -> 0.140 | Helped |
| add glocalized federated learning instead of local training | TransformRec · BBQ 1, BBQ 2, Ulsan Pedal 1, Ulsan Pedal 2 | Precision@10: 0.044 -> 0.048; 0.046 -> 0.052; 0.011 -> 0.015; 0.012 -> 0.014 | Helped |
| add tokenization (ablation) | TransformRec · Ulsan Pedal | HR@10: 2.182% -> 7.329% | Helped |
| replace PP2Vec with TransformRec | commerce recommendation · Ulsan Pedal | HR@10: 4.1% -> 6.0% | Helped |
| replace PP2Vec with TransformRec | commerce recommendation · BBQ Order App | HR@10: 7.8% -> 28.0% | Helped |
| replace PP2Vec with TransformRec | commerce recommendation · BBQ Mall | HR@10: 8.1% -> 14.9% | Helped |
| train on shared data instead of local data | TransformRec · BBQ 1, BBQ 2, Ulsan Pedal 1, Ulsan Pedal 2 | HR@10: 0.443 -> 0.425; 0.455 -> 0.404; 0.110 -> 0.097; 0.117 -> 0.099 | Hurt |
Surprising findings
- Training on pooled shared data performed worse than local-only training, which the authors attribute to heterogeneous data across contexts.
- Generated product names such as 'Rose Sauce Fried Chicken Set' were later launched as a real product by a partner merchant.
Limitations
- Evaluation uses proprietary, non-public receipt data from two partners, so results are hard to reproduce.
- Deployment hit rate is reported only for one partner in an early phase; the claim of double-digit future hit rates is a projection.
- Model is small (1 layer, D_MODEL 128) and the comparison to prior systems is limited to PP2Vec and internal baselines.
Experimental setup
Datasets: Ulsan Pedal transaction receipt data Genesis BBQ transaction data Genesis BBQ online mall data
Benchmarks: Ulsan Pedal held-out split BBQ held-out splits
Metrics: HR@10 Precision@10 HR@3 HR@5
Baselines: PP2Vec Local training Data-shared training Random TransformRec without tokenization
Builds on: Transformer Word2Vec FedAvg Jaccard similarity
Closest papers by meaning
- eCeLLM: Generalizing Large Language Models for E-commerce from Large-scale, High-quality Instruction Data (ICML 2024 · similarity 0.82)
- EcomGPT: Instruction-Tuning Large Language Models with Chain-of-Task Tasks for E-commerce (AAAI 2024 · similarity 0.81)
- FlowerTune: A Cross-Domain Benchmark for Federated Fine-Tuning of Large Language Models (NeurIPS 2025 · similarity 0.81)
- Mix-Ecom: Towards Mixed-Type E-Commerce Dialogues with Complex Domain Rules (ICLR 2026 · similarity 0.79)
- Harmonizing Generalization and Personalization in Federated Prompt Learning (ICML 2024 · similarity 0.79)
- ShoppingBench: A Real-World Intent-Grounded Shopping Benchmark for LLM-based Agents (AAAI 2026 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.