Residual Network Deep Learning Recommendation Model Based on Multi-head Attention Mechanism
张圆梦 1李少波 2周鹏 3杨明宝1
扫码查看
点击上方二维码区域，可以放大扫码查看
作者信息
1. 贵州大学计算机与科学技术学院 贵阳 550025
2. 贵州大学贵州省公共大数据国家重点实验室 贵阳 550025
3. 贵州大学机械工程学院 贵阳 550025
折叠
摘要
深度学习由于其强大的特征表达能力,在推荐研究领域的应用逐渐广泛.DIN(Deep Interest Network)是一种基于注意力机制和用户兴趣进行推荐的深度学习模型,针对其存在的特征训练完备性较低、推荐精度有待提高的问题,提出一种基于DIN改进的融合多头注意力模块与残差网络的深度学习推荐模型:MHAR-DIN(Multi-Head Attention Residual Deep Interest Network).利用多头注意力模块基于用户历史行为进行注意力的打分,充分考虑用户的兴趣偏好,并引入残差网络结构将特征越过训练直接接入全连接器,解决过深网络难以训练的问题.在公开数据集MovieLens上与经典深度学习推荐模型的对比实验表明,所提MHAR-DIN模型具有一定有效性和可行性.
Abstract
Due to its strong feature expression ability,deep learning is gradually widely used in the field of recommendation research.DIN(Deep Interest Network)is a deep learning model for recommendation based on attention mechanism and user inter-est.Aiming at the problems of low completeness of feature training and improvement of recommendation accuracy,MHAR-DIN(Multi-Head Attention Residual Deep Interest Network)is proposed,which is an improved DIN based deep learning recommenda-tion model integrating multi-head attention module and residual network.The multi-head attention module is used to score the atten-tion based on the user's historical behavior,and the user's interest preference is fully considered.The residual network structure is introduced to directly connect the features across the training to the full connector,so as to solve the problem that it is difficult to train in too deep network.The comparative experiment between the proposed model and the classical deep learning recommendation model on the public data set movielens shows that the proposed MHAR-DIN model is effective and feasible.
维普
万方数据