A Comparison of Retweet Prediction Approaches: The Superiority of Random Forest Learning Method

Abstract

We consider the following retweet prediction task: given a tweet, predict whether it will be retweeted. In Keywords: retweet prediction, machine learning algorithms, performanceCopyright © 2016 Universitas Ahmad Dahlan. All rights reserved. IntroductionSocial media like Twitter has provided a platform for spreading information among users [1,2]. In this work we focus on the retweet prediction problem. Given a tweet, we would like to predict whether it will be retweeted. Applications of this task are, for example, to help decision makers propagate their issues and facilitate companies to promote their products.A wide range of learning methods and features have been proposed for retweet prediction; see, e.g., [3][4][5][6][7]. In addition, different modeling setups have been proposed; e.g., Petrovic, et al.,[3] propose a time-sensitive model that builds separate models depending on the tweet's creation time and show that it substantially improves performance over the passiveaggressive learning algorithm. This large variety of features, learning methods, and ways of modeling the task calls for a systematic comparison on a single dataset. Therefore, we propose a systematic comparison of the selected learning methods and features, both within and without time-sensitive framework, all on a single dataset. To the authors' knowledge this is the first attempt to collect best performing features and contrast linear and non-linear learning methods.We consider the following research questions: (i) Which of the proposed learning methods is the most effective for the retweet prediction task? (ii) Which of the proposed features are most discriminative features for the learning methods considered? And (iii) To which degree does time-sensitive modeling help improve the performance of learning methods on our dataset?Prior work has documented several techniques to solve the retweet prediction problem. Naveed, et al., [5] state that the problem of finding "interestingness" on Twitter is the same as predicting whether a tweet will be retweeted. They employ logistic regression to do the prediction and find that content features such as initial negative sentiments make a tweet more likely to be retweeted.

Assessing the reTweet proneness of tweets: predictive models for retweeting

“…The experiments have been compared by using the AUC (area under the precisionrecall curve) demonstrating the dependency of the model with respect to the user feature (e.g., followers counts), hashtag used popularity, user network features. Bunyamin and Tunys [9], have provided a comparison of the performance for different learning methods and features, in terms of retweet prediction accuracy and feature importance, to understand what kind of tweets would be retweeted, by using as response variable a dummy variable representing the two states of being retweeted or not retweeted. They have found that Random Forests method archives the best performance.…”

“…A part of the identified metrics has been also used in [55], where a simple descriptive and Principal Component Analysis have been provided without deriving a predictive model. In the paper of Bunyamin and Tunys [9], a comparative analysis of several methods has been proposed without considering all metrics we identified, and without addressing the prediction of the degree of retweeting.…”

User Connectivity and Event Popularity Based Re-Tweet Prediction in Social Networks

“…The propagation of information in micro blogs has brought phenomenal enhancement and has accelerated inter-personal communication. Re-tweet mechanism gives an approach to enable social clients to hold the most recent news and help undertakings to do advertising on social-media stage [2]. Along these lines, it is of extraordinary practical importance to examine and investigate the re-tweet practices for improving the data spread and user involvement in micro blogs.…”

Prediksi Penyebaran Informasi di Twitter dengan Metode Pembelajaran Mesin dengan Fitur Linimasa

“…Xu et al [7] melakukan prediksi dalam perspektif twitter individu dan mencoba melakukan prediksi dengan metode decision tree, regresi logistik dan support vector machine (SVM) [7]. Metode pembelajaran random forest telah diujikan dalam penelitian sebelumnya dan menunjukan hasil yang lebih baik dibanding metode yang sebelumnya diuji [8][9] [10].…”

“…Bunyamin [8] dalam deteksi popularitas menemukan metode pelatihan yang memiliki performa terbaik adalah random forest dengan memanfaatkan fitur pengguna dan fitur konten. Dalam pencarian feature importance digunakan metode mean decrease importance (MDI) pada random forest dan recursive feature elimination (RFE) pada pelatihan regresi logistik, lalu nilai kepentingannya diurutkan.…”