【摘要】 针对传统推荐模型难以实现对同一个主题的文章连续扩展的问题，提出一种基于门控混合专家网络的实时相关推荐方法。从低维稠密向量交互、语义特征相似性和不同特征字段之间的依赖程度等多个维度捕获特征作为专家网络；通过多门控制的混合专家策略和分层注意力机制，综合考虑这些专家网络；利用最终学习到的深层特征，预测推荐评分和项目点击概率，获得用户对项目的满意度。实验结果表明，与其它基线模型对比，AUC指标最多可提高0.35%,Logloss指标最多可降低0.76%,消融实验也验证了各个部分的有效性，说明了该模型的可行性与准确性。更多还原

【Abstract】 Aiming at the problem that the traditional recommendation model is difficult to realize the continuous expansion of articles on the same topic, a real-time relevant recommendation method based on gated hybrid expert network was proposed. Features were captured as an expert network from multiple dimensions such as low-dimensional dense vector interaction, semantic feature similarity and dependency degree between different feature fields. These expert networks were comprehensively considered through the hybrid expert strategy of multi-gate control and the hierarchical attention mechanism. The final learned deep features were used to predict the recommendation score and the item click probability, so as to obtain the user’s satisfaction with the item. Experimental results show that compared with other baseline models, the AUC index is improved by up to 0.35%, and the Logloss index can be reduced by up to 0.76%. Ablation experiments also verify the effectiveness of each part, and illustrate the feasibility and accuracy of the proposed model.更多还原