Research And Implementation Of Social Media Information Cascade Prediction Method Based On Graph Neural Network

With the development of Internet tools becoming more and more mature,mankind has entered the era of cognitive surplus,and the emergence of increasingly perfect online social media and shared communities with assistance mechanism has rapidly promoted people to contribute their cognitive surplus in creativity and knowledge.Nowadays,social media has become an important medium for information sharing and interaction,and information dissemination has become a research hotspot.This paper focuses on the prediction of information cascade,which has important practical significance and research value.In recent years,with the development of graph deep learning technology,the existing information cascade prediction methods begin to use graph neural network to learn social network structure and information diffusion structure to obtain efficient user representation.Although the above methods have achieved some success,the existing cascade scale prediction methods based on graph neural network do not fully consider the role of text subject semantics,and in the micro prediction task of information cascade,the existing methods fail to make full use of the relationship between users’ social relations and forwarding behavior and users’ content preferences,which play an important role in inferring users’ activation probability.In view of these deficiencies,the main research work of this paper is as follows:For the macro prediction task of information cascade,this paper proposes a prediction model based on graph neural network and text topic semantics.Aiming at the characteristics of user generated content with short text and sparse semantics,this paper uses short text topic model and cyclic neural network to extract the topic semantic features of text content in social media.For the global social relationship between users,this paper uses graph neural network to model,and redesigns the neighborhood aggregation strategy to make it more suitable for cascade prediction tasks.The gating mechanism is used to simulate the activation probability of users under the influence of social network,topic embedding and self activation.Finally,the sum pool is used to output the cascade scale prediction results.Experiments on public data sets show the importance of semantic features,and mining information content with high coherence semantics is more likely to trigger users’ forwarding behavior.For the micro prediction task of information cascade,this paper proposes a prediction model based on dynamic heterogeneous graph convolution network and content preference.By constructing the forwarding relationship between users as a dynamic graph sequence and integrating it with the global social relationship graph into a dynamic heterogeneous graph,the heterogeneous graph convolution network is used to encode the user feature representation,and then the temporal attention mechanism is used to fuse the user’s historical representation.In order to model the non serialized information diffusion pattern,the method uses the multi head self attention mechanism to encode the cascaded sequence context.At the same time,the model calculates the user’s dynamic content preference representation,and finally infers the participation probability of potential users by the combination of sequence context coding and content preference representation.The experimental results show that the way of modeling the forwarding relationship and social relationship as a heterogeneous graph makes natural and efficient use of the user’s forwarding behavior and social structure information,and integrates the user’s content preference,which effectively improves the prediction accuracy.