User Follow Prediction of Microblog Based on the Activeness and Interest Similarity

[a] School of Management and Economics， University of Electronic Science and Technology of China， 610054， Chengdu， China； Business School， Sichuan Agricultural University， 611830， Chengdu， China.

[b] School of Management and Economics， University of Electronic Science and Technology of China， 610054， Chengdu， China.

[c] School of Management and Economics， University of Electronic Science and Technology of China， 610054， Chengdu， China.

[d] College of Economics and Management， Sichuan Agricultural University， 611130， Chengdu， China.

* Corresponding author.

Received 13 May 2013； accepted 16 July 2013.

Abstract

Many microblog fans have formed the basis of microblog information dissemination and diffusion， so accurate prediction and attracting more potential customers to follow microblogger becomes very necessary. By employing interpersonal relationship network of microblog fans， this study integrates microblogger popularity and user activeness into interest similarity in order to explain user follow predictor and propose user follow prediction models. Support vector machine （SVM） is used to train this model. Open data from Tencent microblog in KDD Cup 2012 prove that the proposed prediction model has higher prediction accuracy and stability.

Key words： Microblog； User follow； Microblogger popularity； User activeness； Interest similarity； Prediction model

INTRODUCTION

Microblog， as an important kind of online social networking platform has become an important means of communication， exchange and sharing knowledge in virtual environment （McFedries， 2007； Ebner， 2008）. With the rapid development of microblog services， enterprise has begun to develop marketing activities using microblog platform. On the one hand， the enterprise attracts users’follow， comment and retweet， and converts some of its followers to consumers via Microblog post information about products or services. On the other hand， convenient and effective collection of the users’ comments help to learn the users’ feedback in time and to grasp their interest and potential demand （Hsu， 2010）. Information diffusion in microblog platform not only depends on such factors as the microblogger popularity， the quality and timing of released information， but also relies on the participation in comments and retweets. Therefore， by predicting what users will focus on microblogger， attracting the potential microblog fans to follow the microblogger becomes very important. By employing interpersonal relationship network of microblog fans， this study integrates microblogger popularity and user activeness into Interest similarity to propose user follow prediction model. Support vector machine （SVM） is used to train the proposed model.

The paper is structured as follows. The second part of the paper reviews related work. The user follow predictor and prediction model are proposed in the third part. The fourth part testifies the feasibility of the predictor and accuracy of the prediction model by using actual data. Eventually we summarize the conclusion and present the future work.

1. RELATED WORKS

With the widely application of the microblog， more and more researchers have been engaged in microblog related research. There are three aspects related to this research： microblog adoption， continuance use and influence of microblog. In adoption research， Guardia et al. （2011） established the microblog adoption model （μBTAM） by integrating with such factors as subjective norms， image， perceived usefulness， perceived usefulness， and behaviour intention. It confirmed that the TAM model can be used for modelling microblog adoption. Gunther et al. （2009） proposed the theoretical models of the enterprise staff adoption of the internal microblog by expanding the UTAUT model via conducting focus group interview and introducing such factors as the information privacy concern， prestige， ratio of signal-to-noise. Schoendienst et al. （2011） found that performance expectations and privacy concerns have significant positive effect on behaviour intention by empirical research.

Microblog continuance use model was proposed by employing information systems continuance model and habit in order to explore antecedents of participation in interaction with microbloggers （Barnes B？hringer， 2011）. They found that satisfaction was affected by perceived usefulness and expectation confirmation. Zhao explored the factors affecting users’ satisfaction and continuance intention of microblogging. This study proposed that user satisfaction toward microblog service was positively affected by perceived interactivity， ease of use， telepresence， enjoyment and intimacy. It confirmed that microblog continuance use was significantly affected by such factors as perceived usefulness， satisfaction and habits （Zhao Lu， 2010）.

Along with the development of the recommendation system， recommending microbloggers to users and recommending active users to microbloggers is also experiencing good development. On the one hand， in order to expand information dissemination， microbloggers need to attract more users’ follow and keep interactions to improve its activity. Such studies focus on recommending users to specific microbloggers. On the other hand， in the vast information， recommending interesting microbloggers and useful information to users is also an important research direction. The above two types of research are of great significance both in theory and application. Calculation of user model and the semantic similarity of Microblog referring users to implement， Mei implemented recommendation by similarity between user model and semantic microblog （Mei et al.， 2012）. In order to identify microblog users providing useful information， Armentano proposed an algorithm， based on a user PageRank， to implement user prediction （Armentano et al.， 2012）.

In addition， there are some researchers studied this problem in social networks from the aspects of demographic characteristics， keywords， tags， and the interaction between the users （Zhao， 2012； Chen et al.， 2012； Chen Tang， 2012）. In such studies multi-layer factor decomposition model is used for user modeling and forecasting potential fans， but this method used excessive properties to lead to more computational complexity.

Whether users accept and follow specific microbloggers is associated with the popularity and the user’s activeness of microbloggers， and also related with interest similarity between them. Based on the above research conclusion， user follow prediction model is proposed to identify potential followers， and then motivate them to become followers by active concerns.

2. RELATED DEFINITION AND USER FOLLOW PREDICTION MODEL

2.3.1 Microblogger Popularity and User Activeness

In topic propagation model， blogger popularity and user activeness are the influence factors of propagation process（Zhao et al.， 2009）. In the microblog network， the main property index to depict the Item include the number of fans， the number of “AT” （@）， the total number of retweets and total number of comments， etc. The more numbers or times of the previous factors， the more popular is the microblogger.

At the same time， the number of microblog post， the number of “AT”， the total number of retweets and the total number of comments are also the important indicators of user activeness. User think that participation in the microblogging can meet the purpose of information communication and exchange， and the more active of its retweets and comments， the higher is the activeness. Robert pointed out that the interaction is an important factor to keep users to use microblog （Robert Pongsajapan， 2009）. Meanwhile， the number of “AT”reflects relationship between microblog users and describes the interaction between them. Popularity and activeness constitute the necessary premise of the users’following of the Item. Once the Item with high popularity is recommended to high active user， the latter tend to accept and follow the former. Therefore， this paper will take popularity and activeness as a predictor.

2.3.2 Interest Similarity

In the era of micro propagation， quality of microblog content is associated with the user’s following. The more humorous and interesting the content is， the more satisfactory and attractive is the content to the users. Therefore， users’ following of the item is not only associated with popularity and activeness， but also associated with the similarity of interest.

In recent years， previous research has shown that the rate of microbloggers’ being followed and retweeted is based on the microblog content and the basic characteristics of microblog content. From microblog content （posts and comments） the users’ interest can be excavated. In the context of this study， interest similarity is the important factor of the microbloggers’ being followed. Therefore， interest similarity can be used as prediction index of following.

3. THE EMPIRICAL RESEARCH AND ANALYSIS

3.1 Data description

To test the feasibility of characteristics of prediction of microblog user follow， this paper uses public datasets from KDD Cup 2012 provided by the tencent Microblog Track 1. The datasets contain seven text file complex number， containing information of items， users and their mutual relationship.In addition， the training set file contains over 70 million records of 31 days. Each record is composed by ‘USER’， ‘ITEM’， and ‘RESULT’. The meaning of ‘USER’ and ‘ITEM’ are stated earlier， the value of ‘RESULT’ equal 1，means following ，-1， means rejection. Test datasets， containing 19 million records， are used to verify prediction accuracy. Predictors selected in this paper involves five files， the file name and data description are listed in Table 2.

3.3 The Empirical Results and Analysis

3.3.1 Support Vector Machine （SVM）

Standard support vector machine （SVM）， based on structural risk minimization principle ，which is put forward by Vapnik in the 1990s （Cortes Vapnik， 1995）， is a statistical learning method. SVM not only considers the training sample of empirical risk， but also takes into account the generalization ability of the algorithm， so as to effectively improve the accuracy of the algorithm with other machine learning method based on empirical risk minimization principle. SVM， by maximizing the interval of two support planes， realizes the structural risk minimization principle. It transforms a classification or regression problem into a quadratic convex programming problem， so as to make the classifier obtain the global optimal. In the case of nonlinear separation， the nonlinear mapping can be used to map samples of low-dimensional input space to high dimension linear feature space. Therefore， the linear algorithm can be employed to analyse the nonlinear characteristics of the sample in high dimensional feature space. In microblog network， items are recommended to users， these users can choose whether to follow the items or not. Therefore， user follow predict can be transformed into a classification issue， and because the SVM has good classification ability， the SVM is used to complete the prediction of user follow.

3.3.2 Prediction Results and Analysis

In order to test validity of the prediction feature of user follow， this study need to constitute test dataset， and take the way of sampling produce training and test dataset. To avoid different value among attributes influence the prediction results， it needs to normalize such properties as interest similarity， popularity and activeness

The findings shown from experimental results in Table 4：

（1） Except experiment 8， the experiment accuracy of the two types of sample is similar. Every experiment achieves the approximate prediction accuracy. However， experiment 8 gets a low prediction precision in the first sample， while at the same time， the prediction is high for the second sample. But the total accuracy of other experiments are still very close. Through the analysis， the causes of this deviation may be that the sample set contains more isolated points. It can be improved by using fuzzy SVM in the future.

（2） All kinds of precision standard deviation of the sample and the total sample pay little， showing that the multiple sampling has little affect to prediction accuracy， and the stability of this prediction method is verified.

Using this method to establish follow predictor， the prediction precision of the 0.5541 is achieved by using support vector machine （SVM）. The prediction accuracy obtained by proposed model is slightly better than the previous literatures（Zhao， 2012； Chen et al.， 2012； Chen Tang， 2012）， and the selection of indicators and the prediction method is relatively simple and easy to implement.

CONCLUSION AND FUTURE WORK The increase of active followers to microbloggers help the dissemination and diffusion of microblog information. For microbloggers， by active communication with active users， converting potential active users to their followers to increase popularity and influence is very impotant. User follow predictor and prediction model are proposed in this paper on the basis of research of information dissemination. By using data from KDD Cup 2012 Track 1， and the support vector machine （SVM） for empirical research， This paper achieves a better prediction precision， and comfirms that the proposed indicators such as popularity， activeness and interest similarity are appropriate for predicting user follow.

There are still some shortages in this study. for example， other features such as demographic characteristics， category and tag attributes， etc.， may be used for user modeling. In addition， some isolated data points are found in experimental samples， and these isolated points will impact prediction precision of SVM. Therefore， the choice of predictor and classification method still needs to be studied in further research.

REFERENCES

Armentano， M. G.， Godoy， D.， Amandi， A. （2012）. TopologyBased Recommendation of Users in Micro-Blogging Communities. Journal of Computer Science and Technology， 27（3）， 624-634.

Barnes， S. J.， B？hringer， M. （2011）. Modeling Use Continuance Behavior in Microblogging Services： The Case of Twitter. Journal of Computer Information Systems， 51（4）， 1-10.

Chen， T. Q.， Tang， L. P. （2012）. Combining Factorization Model and Additive Forest for Collaborative Followee Recommendation. http：//www.kddcup2012.org/workshop.

Chen， Y. W.， Liu， Z. T.， Ji， D. Q.（2012）. Context-aware ensemble of multifaceted factorization models for recommendation Prediction in Social Networks. http：// www.kddcup2012.org/workshop.

Cortes， C.， Vapnik， V. N. （1995）. Support-Vector Networks. Machine Learning， 20（3）， 273-297.

Ebner， M.， Schiefner， M. （2008）. Microblogging - more than Fun. In Proceeding of the IADIS Mobile Learning Conference. Algarve， Portugal， 155-159.

Guardia， F. R.， Liebana， F. J.， Fiestas， M. M. （2011）. Motivational Factors that Influence the Acceptance of Microblogging Social Networks： The μBTAM Model. Proc 3rd Int Conf on Education and New Learning Technologies， Barcelona， Spain： IATED， 3196-3206.

Günther， O.， Krasnova， H.， Riehle， D.， Sch？ndienst， V. （2009）. Modeling Microblogging Adoption in the Enterprise（p.544）. Proc 15th Americas Conf on Information Systems， San Francisco， California.

Hsu， C. L.， Liu， C. C.， Lee， Y. D. （2010）. Effect of Commitment and Trust Towards Micro-Blogs on Consumer Behavioral Intention： A Relationship Marketing Perspective. International Journal of Electronic Business Management， 8（4）， 292-303.

McFedries， P. （2007）. Technically Speaking： All a-Twitter. IEEE Spectrum， 44（10）， 84-84.

Mei， K. B.， Zhang， B. F.， Zheng， J. X.， Zhang， L. X.， Wang， M. （2012）. Method of Recommend Microblogging Based on User Model （pp.1056-1060）. IEEE 12th International Conference on Computer and Information Technology.

Robert， A.， Pongsajapan， B. A. （2009）. Liminal Entities： Identity Governance and Organizations on Twitter. Washington： Georgetown University.

Sandes， E.， Li， W. G.， Mello， A. （2012）. Logical Model of Relationship for Online Social Networks and Performance Optimizing of Queries （726-736）. Proceedings of the 13th International Conference on Web Information System Engineering Championship on T1： Scalability.

Schoendienst， V.， Krasnova， H.， Günther， O. （2011）. MicroBlogging Adopting in the Enterprise： An Empirical Analysis（pp.931-940）. Proc 10th Int Conf on Wirtschaftsinformatik， Zurich， witzerland.

Zhao， L.， Lu， Y. （2010）. Perceived Interactivity： Exploring Factors Affecting Micro-Blogging Service Satisfaction and Continuance Intention. Proc PACIS 2010 on Web 2.0， Social Computing and Data Mining， 940-949.

Zhao， L.， Yuan R. X.， Guan， X. H.， Jia， Q. S. （2009）. Bursty Propagation Model for Incidental Events in Blog Networks. Journal of Software， 20（5）， 1384-1392.

Zhao， X. （2012）. Scorecard with latent Factor Models for User Follow Prediction Problem. http：//www.kddcup2012. org /workshop.