User behavior sequences in search systems resemble "interest fossils", capturing genuine intent yet eroded by exposure bias, category drift, and contextual noise. Current methods predominantly follow an "identify-aggregate" paradigm, assuming sequences immutably reflect user preferences while overlooking the organic entanglement of noise and genuine interest. Moreover, they output static, context-agnostic representations, failing to adapt to dynamic intent shifts under varying Query-User-Item-Context conditions. To resolve this dual challenge, we propose the Contextual Diffusion Purifier (CDP). By treating category-filtered behaviors as "contaminated observations", CDP employs a forward noising and conditional reverse denoising process guided by cross-interaction features (Query x User x Item x Context), controllably generating pure, context-aware interest representations that dynamically evolve with scenarios. Extensive offline/online experiments demonstrate the superiority of CDP over state-of-the-art methods.

翻译：搜索系统中的用户行为序列类似于“兴趣化石”，既捕捉了真实意图，又受到曝光偏差、类别漂移和上下文噪声的侵蚀。现有方法主要遵循“识别-聚合”范式，假设序列一成不变地反映用户偏好，而忽视了噪声与真实兴趣的有机纠缠。此外，这些方法输出静态的、与上下文无关的表征，无法适应不同查询-用户-物品-上下文条件下动态的意图变化。为解决这一双重挑战，我们提出了上下文扩散净化器。通过将经过类别过滤的行为视为“受污染的观测”，CDP利用由交叉交互特征引导的前向加噪和条件反向去噪过程，可控地生成纯净的、上下文感知的兴趣表征，这些表征随场景动态演化。大量的离线/在线实验证明了CDP相对于最先进方法的优越性。