Perfomance Optimization Of Recommendation Algorithm Based On Knowledge Map

Nowadays,people increasingly realize the importance of big data.The emergence of big data provides more choices for people,but it will become more difficult for people to get interested information from a large amount of data.When faced with large-scale data,due to the large amount and variety of data,people can not quickly filter out the content they are interested in,so it will reduce the efficiency of information filtering.In order to solve this kind of problem,many researchers begin to build a recommendation model to extract features effectively,so as to complete the task of data filtering.The amount of data is growing,and the amount of information to be carried by the recommendation system is also increasing,so the problems of data sparsity,cold start and interpretability are becoming more and more prominent.At present,many recommendation algorithm models combine the user’s rating information of an item with all kinds of built models,and constantly update the parameters in the model to get the recommendation results.Its defect is that it is unable to analyze the interpretability of the prediction results.In addition,the algorithm can not effectively solve the cold start problem.Therefore,aiming at the problems of interpretability and cold start,this paper introduces the concept of knowledge map,and combines knowledge map with deep learning model,so as to improve the accuracy of the algorithm.The main research work is as follows:（1）Aiming at the interpretability of recommendation algorithm,a multi-path bidirectional recurrent neural network recommendation algorithm based on knowledge map（KG-BiLSTM）is proposed.First,preprocess the acquired data set.Then all kinds of data are constructed in the form of graph,and head entity,relationship entity and tail entity are added.According to the constructed knowledge map,path extraction is carried out.For each user project pair,multiple different path information is extracted,and the multiple paths are input into the bidirectional recurrent neural network LSTM one by one.Secondly,for each node hidden vector trained by LSTM,a layer of attention mechanism is added to avoid the information loss of each node.Finally,for each path,the recurrent neural network and attention mechanism jointly predict its eigenvectors,and input the eigenvectors of each path into the average pooling layer.The pooling layer is used to distinguish the importance of each path,so as to make the final prediction result more accurate.（2）Aiming at the cold start problem of recommendation algorithm,a multi-attribute graph convolution neural network recommendation algorithm based on knowledge graph（KG-SimGCN）is proposed.Firstly,users,items and attribute nodes are precoded to generate embedded vectors.Secondly,the attribute nodes and connections of the items to be predicted are extracted by using the constructed knowledge map of the items,which are input into GCN for training.For the project node,the similarity between the project node and its neighbor node is obtained by similarity calculation,and the attribute information of the neighbor node is fused according to the similarity value.Thus,attribute feature and similar node feature are combined with item feature vector to ensure the integrity of item feature.Finally,a dropout layer is added to discard the node features randomly to ensure the robustness of the model.The product operation is used to combine the user’s prediction vector with the item’s prediction vector.Finally,the user’s score of an item is obtained.