Research On Microblog User Portrait And Recommendation Algorithm Under Biterm Topic Model

Social media platforms,represented by microblogs,are developing rapidly and gradually have a huge impact on people’s daily life.Using this information to carve a portrait of online social media users not only provides users with accurate and personalized services,but also facilitates online opinion monitoring.Currently,word-totopic model(Biterm Topic Model,BTM)and deep learning have become the research hotspots for user interest mining and sequential recommendation algorithms.To further improve the effect of microblog user portrait construction and sequential recommendation algorithm,this thesis conducts an in-depth study in two aspects,namely,optimizing word-to-topic model and time-aware attention network,as follows.(1)For the problem of user blog post timeliness and missing frequency words in short textbook modeling,a Forgetting Curve Time Function and BTM Word Frequency Double-weighted Microblog User Portrait(TW-BTM)is proposed.First,the temporal weights of microblogging terms are calculated by fitting a temporal function using forgetting curves.Then,the recomputed word frequency features are used as random values for Gibbs sampling to propose an improved word frequency weighted BTM topic model to increase the word frequency weights of medium frequency words.Finally,we use the microblogging user behavior influence calculation method to construct user portraits under hot topics.(2)Aiming at the problem that existing models fail to fully consider the latent intent in user interaction sequences and the over fitting of models,a Stochastic Shared Embeddings and Latent Intent Aware Self-Attention for Sequential Recommendation(SSELISR)is proposed.The temporal convolution network is used to deeply convolve the user interaction sequence to obtain the expression of the user’s potential intention to the project.The absolute position and time interval of the project are modeled by using the time interval perception self attention layer.The output of temporal convolution layer and time aware self attention layer is used as intention time aware attention layer to predict the query,key and value of the next item,and find the correlation of items with potential intention.The random sharing embedding technology is used to reduce the over fitting of the model caused by excessive parameterization and improve the accuracy of model recommendation through the random transformation between embeddings.The experimental results show that compared with BTM,SL-LDA and LDA methods,TW-BTM has the best performance under different time slice PMI-score indexes,can accurately mine the subject words of each topic in different time slices,construct a word cloud of subject words of user interests under hot topics,and accurately display the user interests under hot topics.The experimental results show that the SSELISR model has a minimum improvement of 1.43% and 3.99% over the GRU4Rec+,Caser,Ma Rank,and Ti SASRec baseline models,respectively,in both predictive evaluation measures NDCG@10 and Hit@10 on the microblogging dataset.The ablation experiments also validated the interpretability and effectiveness of the random shared embedding and latent intent modules.