Research And Application On Recommendation Technology Based On Deep Learning

With the development of the Internet,users are increasingly relying on the recommender system to provide reference for the selection of their items.The recommendation models are mainly divided into collaborative filtering,content-based recommendation models,and hybrid recommendation models.These models have their own limitations in the face of data sparsity issues and cold start problems,and affect the quality of recommendations in terms of different assessment indicators.Applying deep learning technology to the recommendation field is a challenging and significant research topic.This thesis combines the deep neural network model with the collaborative filtering based shallow model,and proposes a hybrid recommendation model that can utilize both user's behavior information and item's textual content information,the main work of the thesis includes:(1)For the characteristics of the user's implicit feedback data,this thesis introduces the confidence into the user's preference learning,and improves the traditional alternating least squares optimization process,then this thesis proposes a weighted latent factor model for implicit feedback and its learning algorithm to learn effective user's latent factor vector and the item's latent factor vector.(2)Based on the characteristics of item's textual content information,this thesis proposes an attention mechanism based gated recurrent unit network model.First,use the gated recurrent unit network to accurately semantically code the words in the text of the item,and then use the attention mechanism based pooling method to automatically learn the weights of the words in the text to distinguish the importance of different words.In this way,the model effectively models and deeply mines temporal information of textual content of the item.(3)By introducing the user attention mechanism,this thesis improves the word pooling method for modeling the importance of words in the textual content of the item,and proposes a new user's attention mechanism based gated recurrent unit network model.The model enables the user to access the entire word sequence of the item's textual content more directly,thereby extracts the textual content information of the item that the user pays attention to,and further improves the accuracy of the item's textual content information mining.(4)This thesis proposes a new hybrid recommendation model,through the attention mechanism based gated recurrent neural network model,learns the high-level feature representation from the item's textual content information to predict the latent factor vector of the item learned by the weighted latent factor model,then uses the latent factor model to make recommendations.(5)This thesis applies the hybrid recommendation model to the new item recommender task,and evaluates the performance of each component and the overall recommendation of the hybrid recommendation model through the comparative experiments.The experimental results prove that the recommendation performance of the proposed hybrid recommendation model is significantly outperform than the existing recommendation model.