计算机科学与探索 ›› 2023, Vol. 17 ›› Issue (6): 1417-1426.DOI: 10.3778/j.issn.1673-9418.2109017

• 人工智能·模式识别 • 上一篇    下一篇

双感知门控交互的多任务推荐模型

林建,吴云,陈育康   

  1. 贵州大学 计算机科学与技术学院,贵阳 550025
  • 出版日期:2023-06-01 发布日期:2023-06-01

Multi-task Recommendation Model of Dual Perception Gated Interaction

LIN Jian, WU Yun, CHEN Yukang   

  1. School of Computer Science and Technology, Guizhou University, Guiyang 550025, China
  • Online:2023-06-01 Published:2023-06-01

摘要: 针对多任务推荐中存在负迁移的问题,提出一种双感知门控交互的多任务推荐模型(DPGI-MTRM)。首先,在多任务共享网络和专有网络中,创新性地设计了双感知特征提取模块(称为双感知专家层),其作用是对输入特征得到元素级和向量级的双感知特征表示。其次,在门控网络的基础上提出了任务交互层,对经过门控网络输出的特征进行交互计算提取任务之间的高级语义相关性,同时采用残差方式加上原始输入门控的特征向量减少任务交互可能带来的噪音干扰。最后,通过堆叠双感知专家层、门控交互层,连接特定任务的神经网络输出层得到双感知门控交互的多任务推荐模型。此外,在模型训练时使用了梯度归一化的多目标优化方法,使该模型更好地收敛。在Census-income、Synthetic Data和Ali-CCP数据集上进行实验,采用AUC和MSE指标进行评估,实验结果表明,提出的模型表现优于其他基准模型,达到较为先进的性能。

关键词: 多任务, 双感知专家层, 门控交互层, 推荐模型

Abstract: Aiming at the problem of negative migration in multi-task recommendation, the multi-task recommen-dation model of dual perception gated interaction (DPGI-MTRM) is proposed. Firstly, in the multi-task sharing network and the proprietary network, the dual-sensing feature extraction module (called the dual-sensing expert layer) is innovatively designed. Its function is to obtain the element-level and vector-level dual-sensing feature representation for the input features. Secondly, a task interaction layer is proposed on the basis of the gated network, which interactively calculates the features output by the gated network to extract high-level semantic relevance between tasks, and at the same time uses the residual method plus the original input gated feature vector to reduce possible noise interference caused by task interaction. Finally, by stacking a dual perception expert layer and a gated interaction layer, and then connecting the neural network output layer of a specific task, a multi-task recommendation model of dual perception gated interaction is obtained. In addition, the multi-objective optimization method of gradient normalization is used during model training, so that the model can better converge.  Experiments are conducted on the Census-income, Synthetic Data and Ali-CCP datasets, and the AUC (area under curve) and MSE (mean square error) indicators are used for evaluation. Experimental results show that the proposed model performs better than other benchmark models and achieves more advanced performance.

Key words: multi-task, dual perception expert layer, gated interaction layer, recommendation model