基于序贯三支神经网络的个性化推荐

虽然协同过滤可以实现用户的个性化推荐,但是大多数协同过滤及其改进模型未考虑用户和项目等特征,因而不能发掘样本间的非线性关系.与协同过滤相比,深度学习能挖掘丰富的用户兴趣模式,但网络拓扑结构是基于二支决策的方式,忽略了推荐样本的难易程度.为了增强模型的非线性表达,同时区分推荐样本的难易,受序贯三支决策的启发,提出序贯三支决策神经网络个性化推荐模型(personalized recommendation model based on sequential three-way decision with single feedforward neural network, STWD-SFNN-PR).首先,为了将高维稀疏特征向量映射为低维稠密的特征向量, STWD-SFNN-PR采用嵌入进行特征处理.其次,在增量式的网络结构中学习推荐样本,使用Adam优化网络参数,并返回难以推荐的样本.再次,利用序贯三支决策增加延迟决策的策略,并在不同的粒度层采用序贯的阈值,从而动态地实现难以推荐样本的划分.最后,为了验证模型的可行性和有效性,选择多种电影推荐数据集进行研究,并选择经典的神经网络推荐、经典的...

Personalized recommendation based on sequential three-way decisions with single feedforward neural network

Although collaborative filtering(CF) can realize the personalized recommendation of users, most of the CF and their improved models do not consider the characteristics of users and items, hence they cannot explore the nonlinear relationship among samples. Compared with the CF, although deep learning can mine rich patterns of user interests, the network topology is based on two-way decision-making, ignoring the difficulty of recommended samples.To enhance the nonlinear expression of the model and distinguish the difficulty of the recommended samples, a personalized recommendation model based on sequential three-way decision with single feedforward neural network(STWD-SFNN-PR) is proposed. Firstly, to map high-dimensional sparse feature vectors to low-dimensional dense feature vectors, the STWD-SFNN-PR adopts embedding technology for feature preprocessing. Secondly, the characteristics of users and items are learned in the incremental network structure, Adam is employed to optimize network parameters, and samples that are difficult to recommend are returned. Thirdly, sequential three-way decision-making is utilized to increase the strategy of delayed decision-making, and the sequential threshold parameters are adopted at different granularity levels, thus realizing the classification of the recommended samples dynamically. Finally, to verify the feasibility and effectiveness of the model, a variety of movie recommendation datasets are selected for research, and the classic neural network recommendation model, the classic deep learning recommendation model, and the latest CF recommendation models with three-way are compared. Experimental results show that the STWD-SFNN-PR has better recommendation quality.