Scholarlink

The rapid rise in the e-commerce activities over the recent years led to increasing congestion in transportation systems hindering the timely delivery of regular packages. Smarter planning on last mile delivery is needed, thus our study proposes an NLP solution to classify receiver addresses as home, corporate business address, and private business address so that delivery time can be arranged accordingly considering the working days and hours. However, because address data is considered personal information and obtained from disparate e-commerce platforms that are unable to share addresses with each other, a Federated Learning (FL) approach has become essential. This paper proposes a FedAvg based FL solution utilizing pretrained BERT and RoBERTa models for address classification. The experimental results demonstrate the performance of the BERT algorithm in comparison with other classification algorithms in FedAvg in terms of accuracy as \(94.4\%\) and F1-score as \(90.1\%\) .

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Address Classification in E-Commerce Logistics Using Federated Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Barış Bayram,
  • Alaeddin Türkmen,
  • Ayşe Dilara Türkmen,
  • Merve Karademir

摘要

The rapid rise in the e-commerce activities over the recent years led to increasing congestion in transportation systems hindering the timely delivery of regular packages. Smarter planning on last mile delivery is needed, thus our study proposes an NLP solution to classify receiver addresses as home, corporate business address, and private business address so that delivery time can be arranged accordingly considering the working days and hours. However, because address data is considered personal information and obtained from disparate e-commerce platforms that are unable to share addresses with each other, a Federated Learning (FL) approach has become essential. This paper proposes a FedAvg based FL solution utilizing pretrained BERT and RoBERTa models for address classification. The experimental results demonstrate the performance of the BERT algorithm in comparison with other classification algorithms in FedAvg in terms of accuracy as \(94.4\%\) and F1-score as \(90.1\%\) .