作者:黄英来，冀宇超，刘镇波
Authors:HUANG Yinglai,JI Yuchao,LIU Zhenbo摘要:针对农产品电商平台,产品季节性强、地域性强、用户行为多变,导致推荐效果不理想的问题 ,提出了 一种融合表征的农产品推荐算法 。首先,用长短期记忆网络和注意力网络相结合组成深度兴趣网络 , 以此来捕获物品的潜在特征 ;其次 ,构建用户- 商品二部图;再次 ,利用图神经网络提取图数据的连接信息对每个节点的影响 , 并更新节点的嵌入式表示 ,以获取用户的潜在特征;最后 ,将两种潜在特征通过多层感知机得到待推荐农产品的购买概率 ,进一步提取和利用了用户行为序列中的用户深度兴趣 ,并将其融合深度兴趣网络进行推荐 。实验结果表明:融合表征的农产品推荐算法相较于原有模型 AUC 指标提高 9% 以上 ,准确率和召回率提高约 6% 以上 ;相较于不考虑节点嵌入式表示的情况,AUC 和准确率、召回率也均有提高。
Abstract:This paper proposes a kind of recommendation algorithm for agricultural commodities with fusion representation, in response to the issue of unexpected results on agricultural product e-commerce platforms due to the strong seasonality and regionality of products, as well as the variable user behaviors. Firstly, it integrates Long Short-Term Memory Networks and Attention Network to make up Deep Interest Network. This step aims to catch the potential feature of the item. Secondly, it builds up user-product bipartite graph. Then, it uses Graph Neural Network to abstract the impacts that connection information of graph data has on each node. And it also updates the embedded presentation of the node to catch the potential feature of user. Last, the two potential features are fed into a Multilayer Perceptron to get the order rate of the to-be-recommended agricultural commodities. This step combines the user ′ s deep interests derived from their behavior sequence with deep interest network to generate personalized recommendations. The results of experiment have shown that, compared with the previous model, the AUC target of recommendation algorithm for agricultural commodities with fusion representation has increased over 9% . Compared with the situation without taking the embedded presentation of the node into consideration, the AUC, Accuracy and Recall have all increased.

PDF全文下载地址:

可免费Download/下载PDF全文