融合上下文和视觉信息的多模态电影推荐模型  

Multimodal Movie Recommendation Model Integrating Context and Visual Information

作　　者：朱昆 刘姜[1] 倪枫[1] 朱佳怡 ZHU Kun;LIU Jiang;NI Feng;ZHU Jiayi(Business School,University of Shanghai f or Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学管理学院,上海200093

出　　处：《软件工程》2024年第6期68-73,共6页Software Engineering

基　　金：国家自然科学基金资助项目(12371508);国家自然科学基金资助项目(11701370);上海市“系统科学”高峰学科建设项目。

摘　　要：针对传统的上下文电影推荐模型只采用文本数据,从单模态数据获取的信息有限,无法充分解决数据稀疏性带来的问题,提出了一种融合文本和图像数据的多模态电影推荐模型(VLPMF)。首先,VLPMF集成了长短期记忆网络(LSTM)和概率矩阵分解(PMF)。其次,将VGG16提取的图像特征以概率的角度结合到PMF中并构建融合层,将文本特征和图像特征融合后得出预测评分。最后,在Movielens-1M、Movielens-10M和亚马孙AIV数据集上进行对比实验,结果表明,VLPMF模型的均方根误差比对比实验中最优模型的均方根误差分别降低了1.26百分点、1.51百分点和4.30百分点。Traditional context-based movie recommendation models only adopt text data,resulting in limited information obtained from a single modality and inability to fully address the issues caused by data sparsity.To address these problems,this paper proposes a multimodal movie recommendation model(VLPMF)that integrates text and image data.Firstly,VLPMF integrates Long Short-Term Memory(LSTM)and Probabilistic Matrix Factorization(PMF).Secondly,image features extracted by VGG16 are integrated into PMF from a probabilistic perspective to construct a fusion layer,where the fused text and image features are combined to predict ratings.Finally,comparative experiments conducted on the Movielens-1M,Movielens-10M,and Amazon AIV datasets demonstrate that the Root Mean Square Error(RMSE)of the VLPMF model is reduced by 1.26 percentage points,1.51 percentage points,and 4.30 percentage points respectively compared to the best-performing models in the experiments.

关 键 词：推荐系统 图像内容 深度卷积神经网络 概率矩阵分解模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]