Research On User Interactive Behavior Modeling And Recommending Application On Social Media

In recent years,social media has been rapidly developed due to its unique advantage of content generation method and information diffusion speed.More and more people are beginning to use social media as their daily tool for information sharing and online communication.Different from the traditional instant communication applications,the online interactions between social media users are based on messages and leads to the enrichment of the content types and forms of user interactive behavior data.The analysis and modeling of such high-quality data provide opportunities for revealing users' preferences and behavior patterns behind their online behaviors,and further filling in the supporting data gap for personalized recommending systems.Existing studies on user interactive behavior have mainly focused on the information propagation aspect for specific user behaviors,which severely limits their deployment feasibility.Moreover,the influencing factors on users' interaction selection haven't been fully investigated.It's hard for existing efforts to achieve anticipant prediction and recommendation performances as they ignored the multi-modal content information like text and images,the context information like time and Geo-location,and user personal features like sentiment orientation.In this work,the author takes the two most representative and most common user interactive behavior —“mention” and “retweet” —as the main research objects.From the perspective of common users' online interactions,this work explores user behavior under different scenarios by integrating multi-type data sources and further making efficient user and content recommendations.Specifically,this work performs a systematic study on social media user interactive behavior from three specific aspects.The researching aims of this work are to explore the connection between users' implicit personalities and explicit behaviors,and to promote the in-depth study of user interaction behavior in academic communities,as well as to provide new data models and algorithms for industrial applications such as information monitoring,business intelligence,and social marketing.More specifically,the main work and contributions are as follows:(1)Multi-modal Content-Based User Mention Behavior Modeling and RecommendationAlong with the enrichment of information acquisition methods and uploading channels,the trend of user-generated data in social media lies on the increasing amount of multi-modality information,especially on the visual resources.After analyzing several social media services,this work finds that the visual resources can also provide valuable information to reveal users' mentioning tendencies.Hence,this work proposes a novel generative model,named Multi-modal Mention Topic Model(MMTM),to modeling users' online mention behaviors.By modeling the textual and visual contents synthetically,MMTM is capable of learning users' multi-type semantic patterns,the correlations between contents in different modalities,and their joint influence on users' mentioning tendencies in a unified way.For a multi-model mentioning message,this work retrieves the top-k most likely mentionees to construct mentionee recommendation.The result of experiment constructed on large real-world dataset shows that,the proposed method performed significantly better than existing methods in all metrics.(2)Spatial Context-Aware Fast Mentionee RecommendationTriggered by the popularity of localizable devices,especially smart phones,the diffusion of the location dimension in social media provides us valuable opportunity to understand the physical activity-derived user online behaviors.In this case,we can capture users' behavioral patterns and personal preferences more accurately by incorporating the spatial factors in user behavior models.After analyzing two real-world social media datasets,the author observed two geographical phenomenons embedded in users' mentioning activities which reveals the geographical relationships between mentioners and mentionees and further motivate the need for spatial context-aware user mention behavior modeling and mentionee recommendation.In this work,a joint latent-class probabilistic model,named Spatial COntext aware Mention behavior Model(SCOMM),is proposed to generate users' location-tagged mentioning activities by considering both semantic and spatial context factors synthetically.SCOMM is able to learn and model the semantic patterns of mentioners,the geographical clustering areas of mentionees,and their joint effects on mentioners' movement patterns in a unified way.Moreover,an efficient hyper pruning algorithm is designed to accelerate the query answering speed and to improve the recommendation efficiency by pruning the searching space in both spatial and semantic dimensions simultaneously.Besides,this work designs a location inferring method based on the spatial distribution variation of words and user social graph to retrieve users' fine-grain home locations,and further alleviate the decline of modeling accuracy cased by the sparsity of location data.Extensive experiments are conducted on large real-world datasets.The results demonstrate the superiority of the proposed approach by making more effective and efficient recommendations.Moreover,the proposed user location inference method finds the living cities of more than 80% of users accurately,which proves the feasibility and effectiveness of location inference method based on the geographic attributes of text vocabulary.(3)Sentiment-Enhanced Dynamic User Retweet Behavior Modeling and RecommendationThe retweet mechanism in social media plays a critical role in improving message quality,enhancing user communication experience and tracking network information flow.Existing research have mainly focused on retweeting effects like message dissemination and information diffusion,few attentions have been paid on the problem of how to find the most likely message for a common user to retweet,i.e.,the personal retweeting recommendation problem.Although some recent studies proposed to improve the effectiveness of retweeting prediction/recommendation by exploring features with regard to user,content and network,there lacks a comprehensive analysis of the joint influence on users' retweeting selections of all these factors,especially for the retweeting recommendation problem.Moreover,the impact of dynamic public interest and user sentiment bias has been seldom investigated yet.To deal with that,this work proposes a joint probabilistic generation model SDRM(Sentiment-enhanced Dynamic user Retweet behavior Model)to simulate the decision-making process of user retweet behavior.SDRM learns the user's intrinsic interest and time-sensitive public interestrelated topics.Based on that,SDRM models users' sentiment tendencies,social graphes and message popularity synchronously.The results of experiment constructed on realworld datasets show that,the proposed method performed significantly better than existing methods in all metrics.Moreover,the experimental results demonstrate that the time-sensitive public interest and user's sentimental bias are vital factors affecting the retweetability.