使用LLMs从DoorDash餐厅订单推断杂货偏好
Consumers enjoy DoorDash deliveries from a variety of merchants, ranging from restaurants to pet stores. To provide top-quality customer service, it is critical that we can recommend useful items, even if it is a consumer’s first time shopping within a given vertical. This is commonly referred to as the cold start problem. Here we discuss one of those intersections where we tackled how to help consumers new to grocery and convenience delivery. Our efforts identified relevant items using consumers’ DoorDash restaurant histories to build a set of explicit recommendations.
消费者享受来自各种商家的DoorDash配送，从餐厅到宠物商店。为了提供优质的客户服务，能够推荐有用的商品至关重要，即使这是消费者第一次在特定垂直领域购物。这通常被称为冷启动问题。在这里，我们讨论了一个交集，探讨如何帮助新接触杂货和便利配送的消费者。我们的努力利用消费者的DoorDash餐厅历史识别相关商品，以建立一组明确的推荐。
Restaurant order history provides a rich source of implicit consumer preferences, from culinary tastes to lifestyle and dietary habits. We wanted to determine whether we could leverage this data to understand their potential grocery needs. Large language models, or LLMs, became a powerful tool for interpreting semantic nuances alongside trained world knowledge to infer underlying preferences.
餐厅订单历史提供了一个丰富的隐性消费者偏好的来源，从烹饪口味到生活方式和饮食习惯。我们想确定是否可以利用这些数据来理解他们潜在的杂货需求。大型语言模型，或称LLMs，成为了解释语义细微差别的强大工具，结合训练的世界知识来推断潜在偏好。
Our solution employs LLMs to translate a customer's restaurant order history into personalized grocery and convenience recommendations. For example, through statistical analysis and LLM inference, our system analyzed my restaurant order history to surface highly relevant grocery recommendations, shown in Figure 1, such as hot pot soup base, potstickers, and burritos — all items I personally love and frequently purchase. This blog post details how we developed a scalable, evaluation-driven pipeline to tackle the cold start problem and deliver relevant recommendations from the outset.
我们的解决方案利用LLMs将客户的餐厅订单历史转换为个性化的杂货和便利推荐。例如，通过统计分析和LLM推断，我们的系统分析了我的餐厅订单历史，以呈现高度相关的杂货推荐，如火锅底料、锅贴和墨西哥卷饼——这些都是我个人喜爱并经常购买的商品。本文详细介绍了我们如何开发一个可扩展的、以评估为驱动的管道，以解决冷启动问题并从一开始就提供相关的推荐。
Figur...