Scholarlink

<p>Understanding how customers behave on digital platforms and predicting their interactions in advance are valuable assets for fostering customer loyalty and stimulating engagement within e-commerce applications. In this context, clickstream (customer interaction patterns through their clicks) prediction can be useful in identifying future behaviors and supporting strategic institutional decisions. Thus, we propose ClickstreamGPT, an innovative Generative Pre-trained Transformer (GPT) approach for predicting clickstream behavior on e-commerce platforms. For this study, our GPT-based model was trained using clickstream data from 130,000 customers within an e-banking application, comprising 100&#xa0;million clickstream records across approximately 700 unique application screens and their execution flows. Experimental results indicate that our proposed algorithm is capable of predicting up to 37 distinct click patterns for each customer with an average accuracy of 78%. This predictive capability can be crucial for institutions in anticipating customer churn, optimizing marketing campaigns, and enhancing content recommendation within e-commerce platforms.</p>

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

ClickstreamGPT: a transformer-based generative AI approach to clickstream prediction for E-commerce platforms

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月4日

查看全文
  • Felipe Mancini,
  • Victor Augusto Sampaio Precoma,
  • Leticia Alves De Oliveira,
  • Fabio Teixeira

摘要

Understanding how customers behave on digital platforms and predicting their interactions in advance are valuable assets for fostering customer loyalty and stimulating engagement within e-commerce applications. In this context, clickstream (customer interaction patterns through their clicks) prediction can be useful in identifying future behaviors and supporting strategic institutional decisions. Thus, we propose ClickstreamGPT, an innovative Generative Pre-trained Transformer (GPT) approach for predicting clickstream behavior on e-commerce platforms. For this study, our GPT-based model was trained using clickstream data from 130,000 customers within an e-banking application, comprising 100 million clickstream records across approximately 700 unique application screens and their execution flows. Experimental results indicate that our proposed algorithm is capable of predicting up to 37 distinct click patterns for each customer with an average accuracy of 78%. This predictive capability can be crucial for institutions in anticipating customer churn, optimizing marketing campaigns, and enhancing content recommendation within e-commerce platforms.