基于分离嵌入交叉网络的推荐模型  

Recommendation model based on separated embedding interaction networks

作　　者：封书蕾 蒋中云[2] FENG Shulei;JIANG Zhongyun(School of Information,Shanghai Ocean University,Shanghai 201306,P.R.China;School of Information Technology,Shanghai Jian Qiao University,Shanghai 201306,P.R.China)

机构地区：[1]上海海洋大学信息学院,上海201306 [2]上海建桥学院信息技术学院,上海201306

出　　处：《深圳大学学报（理工版）》2023年第4期513-520,共8页Journal of Shenzhen University(Science and Engineering)

基　　金：上海高校应用型本科试点建设资助项目(Z32004-17-84);上海市教育委员会一流本科专业建设专项资助项目(JYLB202002)。

摘　　要：针对现有深度学习推荐模型中的特征交叉方法存在无法充分利用嵌入向量信息与预测的精度不够的问题,提出一种基于分离嵌入交叉网络(separated embedding interaction networks,SEIN)的深度学习推荐模型.该模型先采用嵌入神经网络层将稀疏的特征向量转化为稠密的嵌入向量,再对不同维度的特征矩阵分离进行特征交叉,并且通过分离嵌入交叉网络层数显式控制特征交叉的阶数,最终将得到的各隐藏层矩阵求和池化,并通过预测层得到最后的输出.在Criteo、AutoML与Movielens公开数据集上,以推荐结果的曲线下面积、对数损失、准确率与召回率作为评估指标进行了点击率预测与top-k推荐实验.结果表明,与点击率预测基线模型DeepFM、Deep&Cross与xDeepFM对比,SEIN模型在Criteo数据集上的曲线下面积分别提升了2.38%、2.31%与2.35%,对数损失分别下降了1.81%、1.99%与1.85%;在AutoML数据集上曲线下面积分别提高了1.17%、2.60%与0.57%,对数损失分别下降了0.66%、2.53%与0.35%.与基于图神经网络的推荐模型HeteGraph、IR-Rec、GSIRec、KGNCF-RNN和ITRA相比,基于SEIN的推荐模型的准确率(k=5)分别提升1.27%、0.47%、0.48%、0.56%与2.59%.基于SEIN的推荐模型能够有效解决无法充分利用嵌入向量信息的问题,提高推荐准确度.Aiming at the problems that the existing feature interaction methods in deep learning recommendation models cannot fully utilize the embedding vector information and thus have the insufficient accuracy,we propose a deep learning recommendation model based on separated embedding interaction networks(SEIN).This model first uses the embedding neural network layer to convert the sparse feature vectors into dense embedding vectors,then separates the feature matrices of different dimensions for feature interaction,and explicitly controls the order of feature interaction through the number of SEIN layers.Finally,the obtained hidden layer matrices are pooled by summation,and the final output is obtained through the prediction layer.In public datasets of Criteo,AutoML and Movielens,click-through rate prediction and top-k recommendation experiments are carried out by using the area under the curve(AUC),log-loss,accuracy and recall rate as evaluation indicators.The experimental results demonstrate that compared with the baseline models for click-through rate prediction,namely DeepFM,Deep&Cross and xDeepFM,the SEIN model improves the AUC by 2.38%,2.31%and 2.35%on the Criteo dataset,and reduces the log-loss by 1.81%,1.99%and 1.85%,respectively.On the AutoML dataset,the SEIN model increases the AUC by 1.17%,2.60%and 0.57%,and reduces the logloss by 0.66%,2.53%and 0.35%,respectively.Compared with the recommendation models based on graph neural networks,specifically HeteGraph,IR-Rec,GSIRec,KGNCF-RNN,and ITRA,the recommendation model based on SEIN improves the accuracy(k=5)by 1.27%,0.47%,0.48%,0.56%and 2.59%,respectively.Therefore,the SEIN model can effectively solve the problem of not fully utilizing the embedding vector information and thus improve the accuracy of recommendation.

关 键 词：人工智能 推荐技术 深度学习 特征交叉 向量稠密化 数据挖掘 点击率预测 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] TP183[自动化与计算机技术—控制科学与工程] TP391