Research On Event Detection And Analysis And Applications In Social Networks

With the rapid development of the Internet,the way people obtain and diffuse in-formation has undergone revolutionary changes.In particular,the emergence of online social networks(ie,Social Network Service,SNS)has changed the traditional way of information diffusion,and social networks have gradually become the mainstream in-formation publishing and dissemination platform.An event is an abstract concept that has a topic,a time dimension,and a set of related entities(such as location,people,organization,and so on).This paper carries out data mining and method research on information diffusion on social networks.Through the research on the occurrence,out-break and development of events on social networks,we can understand the mode and characteristics of information dissemination,and then provide effective support for prac-tical applications such as computing advertising and data journalism.The specific work and innovations are as follows.First,this paper proposes a method for detecting,tracking and evolving trending events on social platforms.Social events generally involve information such as time,place,and subject,which may be related to current affairs politics,popular events,etc.,and may also be bad information such as online rumors and false advertisements.In turn,people can understand what happened through the content of heated discussions on social networks.Trending event detection and tracking is also an important part of the recommendation system.It is necessary to detect hot topics and emergencies in the data on the social platform,and hope to discover potential hot news in a timely manner when the trending has not completely erupted,combined with social communication data In the development process of the event,the latest progress of the event is tracked in time,and a key time series of event development is finally formed.How to auto-matically discover these social events and their evolutionary relationships to help users filter and organize information is an urgent task to be solved.Although there have been many related works on the evolution analysis of news events and document topics before,social media brings new challenges.Social media is usually short text,with high dimen-sionality,sparseness,and large amount of data.Due to the number of words,slang or abbreviations are generally used to refer to things.A tweet usually only describes one event,and it is not possible to mine the event co-occurrence relationship from itself.At the same time,social media is interactive,allowing users to forward and comment,which provides the possibility to analyze the evolution of social events from the user's perspective.We use the event sequence detection and tracking methods to experiment with a large amount of data on Twitter.The experiments show that the method proposed in this article can detect and track events on social platforms very well.Then,based on event detection,tracking,and evolutionary analysis methods,this paper proposes a method for representing and predicting event sequences on social net-works.The current events are related to the past,and the development of the sequence of events has an inherent pattern.Understanding these internal patterns can help re-searchers better predict the type and timing of events that will occur next.In the liter-ature,researchers mainly use two types of methods to model event sequences,namely feature-based methods and generative approaches.Feature-based methods extract mul-tiple types of features,and then train a regression model or classification model to make predictions.However,the performance of feature-based methods depends on the qual-ity of feature extraction.The generation method usually assumes that the evolution of the event follows a random point process(such as a Poisson process or a more compli-cated point process).However,the true distribution of the event sequence is often that a specific point process cannot be completely characterized,and the performance of the model depends on the design of the random point process.In order to solve the prob-lem that both methods have deficiencies,we propose a new deep probability generation model for time series.The model performs a low-dimensional representation of the se-quence of events detected on social networks,and combines random point processes and Variational Autoencoders to make better use of hidden information and obtain the distribution of the next event arrival time and type.Experiments on real data sets prove the effectiveness of the proposed model.Finally,we propose a method for predicting CTR based on social analysis.With the rapid development of e-commerce,online advertising on social media has exploded in recent years,and the annual online advertising business in the United States has reached hundreds of billions of dollars.Due to the development of the real-time bidding model,in the calculation of advertising,the primary task is to set an appropriate price for each advertisement,so as to maximize the revenue of advertisers.Among the charging meth-ods for advertising,per-click charging,that is,the mode in which an advertiser pays a platform for advertising based on the number of user clicks,is the most popular advertis-ing transaction mode.Therefore,the accurate estimation of the probability of each time the user clicks on the advertisement is directly related to the revenue of the advertising platform.Advertising transaction data has the following characteristics.First,the ad-vertising transaction data is very sparse.Second,the real advertising data usually has a huge amount of data.In view of the above characteristics of advertising transaction data,this paper proposes a factorization machine model for sparse data.In the model,using the Laplace distribution to model the parameters can produce fewer non-zero elements,and can highlight the relevant features and feature pairs.Moreover,this paper designs a distributed implementation of the sparse factorization machine.Finally,experiments on two real-world datasets show the effectiveness of this method.