探索图神经网络在Zalando变革推荐系统中的潜力

Recommender systems are vital for personalizing user experiences across various platforms. At Zalando, these systems play a crucial role in tailoring content to individual users, thereby enhancing engagement and satisfaction. This is particularly important for Zalando Homepage, which serves as the customers' first impression of the company. Our current recommendation system employed on the Home page excels by leveraging user-content interactions and optimizing for predicted click through rate (CTR). The research introduced in this post focuses primarily on the approach and design of integrating GNN into the existing recommender system. We aim to validate the feasibility and effectiveness of this integration before transitioning to a fully production-ready implementation.

推荐系统对于个性化用户体验至关重要。在Zalando，这些系统在根据个人用户定制内容方面起着关键作用，从而增强了参与度和满意度。这对于Zalando主页尤为重要，因为它是客户对公司的第一印象。我们目前在主页上使用的推荐系统通过利用用户-内容交互并优化预测点击率（CTR）表现出色。本帖介绍的研究主要集中在将GNN集成到现有推荐系统中的方法和设计上。我们旨在验证这种集成的可行性和有效性，然后再过渡到完全生产就绪的实现。

The Problem Statement

问题陈述

Given a preselected pool of content that potentially can be shown to a user on Zalando Homepage, we need to predict CTR for each piece of content so that later in the system the content with the highest expected value (which predicted CTR is part of) can be shown to the user.

在给定一个预选的可能展示给用户的Zalando主页内容池的情况下，我们需要预测每个内容的点击率（CTR），以便系统稍后可以向用户展示预期价值最高的内容（预测CTR是其中的一部分）。

Our production model relies on traditional tabular data, capturing user-content interactions such as views and clicks, and contrasts with the high potential of graph neural networks. GNNs have emerged as a powerful tool for modeling relational data, offering a way to represent and learn from complex interaction patterns more effectively. GNNs operate by representing data as graphs, and recommender system can be naturally modeled as a bipartite graph with two node types: users and items, and its links connect users and items and indicate user-item interaction (e.g., click, view, order, etc.).

我们的生产模型依赖于传统的表格数据，捕捉用户-内容交互，如浏览和点击，这与图神经网络的高潜力形成对比。GNN已成为建模关系数据...