感知信息的反事实特征增强社交推荐

Perceiving information for counterfactual feature⁃enhanced social recommendation

现如今,对比学习迎来了应用狂潮,计算机多个领域对于对比学习的潜力的开发达到了新的阶段。然而现实世界用户的交互行为历史往往是充满噪声的,而且每次交互世界相隔时间较长,稀疏性较大。针对此种交互历史建模往往会导致用户表征不准确、健壮性弱以及容易崩溃的问题,设计了主要研究一种可以摆脱原始的用户交互历史中的噪声和稀疏性质,通过建模反事实数据分布来对比学习准确和健壮的用户表示的社交推荐模型。构建一种通过对比学习的反事实交互特征和观测交互特征来增强用户表征的框架。

Contrastive learning is now experiencing a frenzy of applications,and several areas of computing have reached a new stage in the exploitation of its potential.However,real‑world user interaction histories are often noisy and sparse with long intervals between each interaction world.Modelling such interaction histories often leads to inaccurate user representations,weak robustness and crash‑proneness.To address this problem,the design focuses on a social recommendation model that can escape the noisy and sparse nature of the original user interaction history and learn accurate and robust user representations by modelling counterfactual data distributions contrastively.A framework for enhancing user representations by comparing learned counterfactual interaction features with observed interaction features is constructed.