Long-and Short-term Interest Mining And Recommendation Algorithms For Microblog Users

Microblog is one of the most popular social media in the world.User interests are rich and varied.In order to better recommend relevant content to Microblog users,user interest mining and analysis are required.Currently,Neural Topic Model(NTM)and Neural Network have become the hotspots for research on long and short-term interest mining and recommendation respectively.In order to further improve the effect of user’s long-and short-term interest mining and recommendation,this thesis conducts an in-depth study from two aspects,namely,optimised neural topic model and long and short-term memory network(LSTM),which are mainly as follows.(1)Lack of effective training for complex unsupervised microblog data.The Chinese word boundaries are not clear and the semantics are not easily understood.The model is unable to extract valid features.A Prompt-optimized Double-Tower microblog user interest mining Contextualized Topic Model(PDTCTM)is proposed.A feature extraction model based on a double-tower structure is used to obtain word embeddings.The double-tower structure is a dynamic fusion of the pre-training model SBERT and the Chinese-Ro BERTa.Keywords are introduced into the pre-trained models SBERT and Chinese-Ro BERTa as Prompt information.The optimized double-tower model performs feature vectorization of words and phrases.Sampling and generating vocabulary based on topics using the CTM topic model.Topic modelling is completed.(2)To address the problems of dynamic time intervals between behavioral sequences,semantic irregularities and intertwined long-and short-term interests of users,a Time-Aware recommendation algorithm for user Long-and Short-term interest features Separation(TALSS)is proposed.Long-term interest features are captured using user personalized time-aggregated interval perception and temporal location multi-headed attention.Time-LSTM with dynamic time interval perception and latent intent attention is used to capture short-term interest features.In order to capture user interests on two time scales separately and independently,a separate acquisition method for long-and short-term interest features is proposed.To improve the accuracy of user interest feature capture,long and short-term interest features are adaptively fused through an attention mechanism.The experimental results show that the PDTCTM model improves the PMI and cos α metrics by an average of 4.37% and 9.77% respectively over the Prod LDA,CTM and zeroshot TM models on the real Microblog dataset.The TALSS algorithm predicted accuracy metrics AUC and GAUC on the Microblog dataset,improving by1.06% and 10.24% respectively over Sli-Rec and TLSAN.The ablation experiments also further demonstrate the need for the TALSS algorithm to choose dual long-and short-term interest features,the Time-LSTM model for short-term interest feature acquisition and the adaptive fusion of long-and short-term.