The Research On Representation Of Network Information Diffusion's Predictability Based On Information Entropy

When choosing a suitable model,how to compare the prediction models naturally becomes an urgent problem to be solved.At present,researchers simply put each diffusion prediction model on their respective diffusion datasets to conduct experiments and compare their prediction accuracy.Due to the lack of a unified representation of the prediction tasks,the comparison result is just a single case which lacking universality.Therefore,this article explores this issue,and proposes to use an indicator to represent the predictability of the diffusion datasets.From the perspective of predictability,the different diffusion datasets are uniformly represented.The prediction results on can refer to the predictability of the data set,making the results comparable.This article is divided into five sections.The first section is the introduction.This part clarifies the reason for this topic,and sorts out the research status.First of all,this part makes a general summary and classification of information diffusion prediction models on social networks,and then puts forward the current research need due to the lack of unified standard of prediction models comparison and analyzes the shortcomings of the current model comparison method.The second part introduces the concept of information entropy firstly.In information theory,information entropy is used to represent the uncertainty of random events.Therefore,this paper proposes an index for representing the predictability of the network diffusion datasets,average entropy of cascades' order(ACE for short),and defines the calculation formula of ACE based on the formula of information entropy.ACE measures the amount of information about the order of the users contained in the diffusion cascade sets.In the third part,we first introduced representation learning and the advantages of network information diffusion prediction using it,then introduced two classic diffusion prediction models based on network representation learning.Based on these two models,this paper considered the influence among different diffusion cascades in the latent spaces.By embedding the diffusion cascades from different sources into different latent spaces,an improved prediction model is proposed.Finally,the calculation method of model prediction precision is introduced.In order to verify whether ACE proposed in this paper can be used to represent predictability,this paper uses the prediction models' precision as a bridge,and explores whether ACE is related to predictability by studying the correlation between it and precision.The fourth part firstly proves that there is a stable correlation between ACE and the precision of the model through experiments.It is concluded that ACE can be used to express predictability.Based on this,this paper proposes an application of ACE which is to compare prediction models by ACE.That is,to generate a set of diffusion cascade sets with different ACE values,then compare these datasets' precision separately.This new comparison method not only has a relatively small workload,but also enables the comparison results to be more representative due to the uniformity of the different datasets in the predictability dimension.At the same time,the comparison method can also investigate the trend of prediction models' precision when the difficulty of the prediction task changes.The fifth part is the conclusion.Firstly,the three main works of the paper are summarized and the future research work is put forward.The paper proposes that the quantitative relationship between predictability and information entropy can be further studied.And the influence of network structure on predictability is also a research direction worthy of attention.