Research On Recommendation Algorithm Based On Neural Network Fusion User Interaction History

With the rapid development of the Internet,Internet information has grown exponentially,and information overload has become a problem we have to face.The network is full of all kinds of information,including some low-quality or even useless information,users will always receive this information inadvertently while browsing the required information.The recommendation algorithm came into being,which helps people obtain useful information more conveniently by customizing the information and pushing it.With the development of artificial intelligence,various deep learning algorithms are widely used.How to integrate deep learning into recommendation algorithms has become one of research focuses of scholars.The main work of this thesis will focus on two research directions: implicit feedback recommendation and explicit feedback recommendation.(1)As to the noise in the feature selection of traditional collaborative filtering recommendation algorithm,the method of feature extraction is analyzed,and Autoencoders is extended to the matrix decomposition framework with good scalability.We propose a implicit feedback recommendation algorithm which incorporates Autoencoders into item's modeling.In the feature extraction,the inherent information of the item is used as the input of the Auto-encoders,and the low-dimensional nonlinear features are obtained by multilayer dimension reduction.When constructing the recommendation model,the low-dimensional features of the items are integrated into the potential feature vectors of the items in the matrix decomposition model,and the user's historical interactive items are then used to reconstruct the potential feature vectors of the user,which alleviates the cold start problem.Experiments were conducted using two public datasets,MovieLens and Pinterest.Compared with the traditional implicit feedback recommendation algorithms,the predicted hit rate of our model is significantly improved.The experimental results show that Auto-encoders can effectively extract the nonlinear characteristics of the item and provide accurate recommendations to users.(2)Aiming at the problem of information imbalance in the traditional explicit feedback recommendation algorithm,the weight calculation method is analyzed in this dissertation.By applying the neural network to the calculation of historical interaction weights,an explicit feedback recommendation model based on neural attention mechanism is proposed.When distinguishing the importance of different historical interactions,the target item and the historical interactions are integrated into a vector as the input of the attention mechanism network,and the weights of each historical interaction are learned.In the final prediction phase,the neural network is used instead of the traditional inner product method to better capture the high-level interaction between the user and the item.Extensive experiments on two real-world datasets show significant improvements of our proposed method over the traditional explicit feedback recommendation methods.