Research On Video Popularity Prediction Based On Knowledge Graph And LSTM Neural Network

In recent years,research on predicting the popularity of online content,especially videos,has attracted widespread attention.Successful popularity prediction can benefit many practical applications,such as recommendation systems,edge caching,optimizing advertising strategies,and balancing network throughput.Although some progress has been made in the field of online video popularity prediction,it still faces the following challenges:1)large fluctuations in popularity,which are easily affected by various external factors,making it difficult to predict and capture popularity trends;2)diverse,sparse,and noisy characteristics of online video features make the prediction task complex and unstable;3)Some data possess time-sensitive characteristics,necessitating an exploration of how to integrate temporal and content features in modeling.Given the numerous issues associated with predicting video popularity,this paper conducts research in three areas: capturing temporal features,extracting cross-features from metadata,and mining associated features from knowledge graphs.Furthermore,it considers how to effectively combine these three aspects.The main contributions and innovations of this paper are as follows:(1)To address the difficulty in capturing the dynamic changes in video popularity trends,this paper proposes a Multi-Feature Multi-Channel Attention-based LSTM(MFC-ALSTM)This method captures long-term dependencies in sequences effectively by introducing three gate controllers.Additionally,the inclusion of multi-feature and multi-channel allows the LSTM to process various temporal features and automatically learn the contribution of each channel to the prediction results.Meanwhile,by incorporating temporal attention,the MFCALSTM can adaptively allocate different weights to each time step.The integration of feature attention and channel attention enables the MFC-ALSTM to learn the underlying patterns of data from multiple dimensions.Experimental results show that the proposed MFC-ALSTM model achieves a prediction accuracy of 61.6% on the Bilibili movie dataset,which represents a 10.59% improvement compared to the LSTM model.(2)Considering that static features of movies play a vital role in short-term popularity prediction,this paper proposes an Adaptive Temporal Knowledge Graph Network(ATKN)video popularity prediction model.This model optimizes the previously mentioned MFCALSTM by preserving its ability to capture movie time-series features while introducing static movie features to enhance prediction accuracy.Firstly,an Attention-based Factorization Machine(AFM)is introduced to handle movie static metadata features.This model can not only handle high-dimensional sparse data but also process various types of features,assigning different weights according to the input feature differences.Secondly,knowledge graphs are introduced,and the Relational Graph Convolutional Neural Network(RGCN)model is utilized to learn relationship information within the graph,learning node representations by propagating and aggregating different types of edges in the graph.This helps capture the dependencies and interaction relationships between videos.Finally,the feature fusion module is redesigned,and a dynamic feature fusion method is proposed.This method dynamically integrates time-series features,metadata features,and knowledge graph association features by introducing gated mechanisms and attention mechanisms.Experimental results show that the proposed ATKN model in this chapter achieves a prediction accuracy of 68.2% on the Bilibili movie dataset,which represents a 10.71%improvement compared to the MFC-ALSTM model presented in the previous chapter.