基于用户多特征融合的个性化推荐模型

Personalized recommendation model based on users multi-features fusion

个性化推荐是提取特定信息的有效手段之一,针对传统推荐法中缺少用户特征问题,提出一种基于多特征融合的广义矩阵分解与深度长短期记忆网络混合推荐模型(hybrid recommendation model of generalized matrix factorization and deep long short-term memory network based on multi-features fusion,HMF)。该模型利用潜在特征向量维度表征年龄、职业对项目的选择性影响,凸显用户个性化;利用长短期记忆(long short-term memory,LSTM)获取用户与项目间的时序特征,再通过多层感知机挖掘其深层次的时序非线性高阶交互关系;将广义矩阵分解(generalized matrix factorization,GMF)获得的简单非线性低阶交互关系与复杂时序非线性高阶交互关系融合,经全连接层得出最终预测评分。HMF有效地利用用户多特征和用户—项目交互信息,实现个性化动态精准推荐。为验证模型的有效性和可行性,在公开数据集MovieLens-1M上进行测试。仿真实验表明,当HMF的潜在特征向量维度为50、MLP层数为7时,HR@10和NDCG@10分别为0.9831、0.9749,相比传统单特征模型NCF最优解分别提高了27.61%、54.29%。

Personalized recommendation is one of the most effective means to extract specific information.Aiming at the problem that lacking of users feature in traditional recommendation methods,a hybrid recommendation model of generalized matrix factorization and deep long short-term memory network based on multi-features fusion(HMF)was proposed.The model used potential eigenvector factors to characterize the selective impact of age and occupation on the project,highlighting user personalization.Long short-term memory(LSTM)was used to obtain the temporal characteristics between users and projects,and then the deep temporal nonlinear higher-order interaction relationship was mined by multi-layer perceptron.The simple nonlinear low-order interaction obtained by generalized matrix factorization(GMF)was fused with the complex time-series nonlinear high-order interaction,and the final prediction score was obtained through the full connection layer.HMF effectively utilized the user’s multi-feature and user-project interaction information to realize personalized dynamic and accurate recommendation.In order to verify the validity and feasibility of the model,the test was conducted on the public dataset MovieLens-1M.The simulation experiment shows that when the potential eigenvector factors of HMF is 50 and the MLP layer is 7,HR@10 and NDCG@10 are 0.9831 and 0.9749 respectively,which are 27.61%and 54.29%higher than the optimal solution of the traditional single feature model NCF.