key: cord-0836520-11f7pij3 authors: Cao, Ren-Meng; Liu, Xiao Fan; Xu, Xiao-Ke title: Why cannot long-term cascade be predicted? Exploring temporal dynamics in information diffusion processes date: 2021-09-01 journal: Royal Society open science DOI: 10.1098/rsos.202245 sha: aa55092c2866ac43df9604d1fc6335fdbf42f7ee doc_id: 836520 cord_uid: 11f7pij3 Predicting information cascade plays a crucial role in various applications such as advertising campaigns, emergency management and infodemic controlling. However, predicting the scale of an information cascade in the long-term could be difficult. In this study, we take Weibo, a Twitter-like online social platform, as an example, exhaustively extract predictive features from the data, and use a conventional machine learning algorithm to predict the information cascade scales. Specifically, we compare the predictive power (and the loss of it) of different categories of features in short-term and long-term prediction tasks. Among the features that describe the user following network, retweeting network, tweet content and early diffusion dynamics, we find that early diffusion dynamics are the most predictive ones in short-term prediction tasks but lose most of their predictive power in long-term tasks. In-depth analyses reveal two possible causes of such failure: the bursty nature of information diffusion and feature temporal drift over time. Our findings further enhance the comprehension of the information diffusion process and may assist in the control of such a process. The study of information diffusion in online social networks has practical values in various domains, such as advertisement 'retweeting'. A user's original tweets and retweets will appear on its followers' timelines. The dataset contains three parts: original tweets: the tweet identity (ID), posting user's ID, release time and content for 30 010 original tweets. These tweets were posted between 1 January 2015 and 20 May 2016; retweets: the original tweet's all retweets, including retweeter ID and retweeting time. There are 15 975 387 retweets from 7 445 377 users; and following relationship: the lists of users that each of the posting users and retweeters followed, including 7 977 942 users and 700 434 403 following relations. In this section, we adopt a best-practice method to predict the information diffusion scale. First, we extract features of the users' following and retweeting network, tweet content and early popularity dynamics. Then, we use a robust ML model (i.e. XGBoost) to forecast the cascade sizes of the original tweets based on these features. Finally, the prediction performance and loss of it for each group of features are analysed and compared. We have developed an extensive list of features, including both previously reported and newly coined ones, for ML model training (table 1) . These features can be grouped into four categories: the following network structure, retweeting network structure, early-stage popularity dynamics and tweet content. On online social media platforms, people can follow the ones that they are interested in. Based on the following relationships among users, a network G F = (U F , E F ) can be constructed, where each user u ∈ U F is represented as a node and each directed edge e(u s , u t ) ∈ E F from u s to u t is established when u t follows u s . This network not only provides pathways for information diffusion but also affects the size of the information cascade. For example, prior studies [20, 35] have shown that users with a large number of followers could lead to large retweet cascades. Five network structural features are constructed to measure the user influence in the following network: out_degree_F, in_degree_F, all_degree_F, bi_degree_F and pagerank_F, where out_degree_F and in_degree_F were both studied in previous works [15, 20, 28, 30] . Retweeting is a key mechanism of information diffusion and a specific manifestation of information popularity [36] . Maleewong [37] found that the retweets from active users (who have stable and frequent participation in posting and retweeting) could receive a large number of retweets. For each original tweet, a retweeting network G R = (U R , E R ) based on the retweet data in the first 60 min is constructed. In G R , each user u ∈ U R is represented as a node and each directed edge e(u s , u t ) ∈ E R from u s to u t is established when u t retweets one of the original tweets or retweets from u s . Five network structural features are constructed to measure the user influence in the retweeting network: out_degree_R, in_degree_R, all_degree_R, bi_degree_R and pagerank_R, in which pagerank_R was studied in previous works [13, 30] . The early-stage retweet time series contains rich temporal information. Szabo & Huberman [16] found a strong linear correlation between the log-transformed popularity at early and later periods in Digg and Youtube. Hu et al. [14] quantified the temporal features from three dimensions: the overall level of the popularity, increasing or decreasing trends, and cyclicity. where peak represents the largest net increase of the retweet time series in the observation period; stability: this index measures the stability of the popularity time series in the observation period, which is defined as the ratio of the standard deviation δ of the time series cascade to the mean u, that is, [15, 20, 28, 30] number of people the posting user follows in_degree_F [15, 20, 28, 30] number of people who follow or are followed by the posting user whether the original tweet has URLs URL [15, 20] whether the original tweet has hashtags hashag [15, 20] whether the original tweet has photos photo [15] whether the original tweet contains '@user name' terms mention [15, 20] topics category of the original tweet contentgory [13] text length of the original tweet wordlength [15, 28, 30] royalsocietypublishing.org/journal/rsos R. Soc. Open Sci. 8: 202245 We use the term frequency-inverse document frequency algorithm [38] to extract the feature vectors of content text and used the latent Dirichlet allocation (LDA) topic model [39] to obtain the topic distributions of each original tweet. Apart from text, multimedia and hyperlinks can be used to convey information, which increases the diversity of information expression and affects the popularity of information. For example, Suh et al. [20] found that the numbers of hyperlinks and hashtags have a strong correlation with retweet likelihood by using principal component analysis (PCA). Zhao et al. [40] found that tweets with multimedia information have a greater number of retweets and a longer life span than plain text messages. Besides, the topics, length and release time of original tweets also affect retweet probability [13, 28, 41] . Considering the influence of contents of original tweets on the retweet likelihood, six features are extracted from the tweet content, where the feature of contentgory was mentioned in [13] . The rest were mentioned in [15, 20] . The detailed descriptions of the features are as follows: contentgory: the topics category of an original tweet extracted by the LDA topic model; wordlength: the text length of the original tweet; -URL: whether the original tweet has URLs; hashtag: whether the original tweet has hashtags; photo: whether the original tweet has photos; and mention: whether the original tweet contains @user name terms. The prediction problem is defined as follows: given the observation of retweeting dynamics of an original tweet m after being posted for a period t, predict the cascade size of the original tweet at time t + Δt. The cascade size at a certain time can be viewed as a continuous value; thus, we treat the prediction task as the classical regression problem. The XGBoost model [42] is used as the training algorithm. We use the root mean square error (RMSE) [43] to evaluate the predictive performance of our model. The RMSE measures the deviation between real and predicted values. The smaller the value of RMSE, the better predictive performance we achieve. Let N m (t) be the real cascade size of the original tweet m at time t andN m ðtÞ be the predicted cascade size of m at time t. For a dataset with M original tweets, the prediction error in RMSE at time t is defined as We consider two prediction scenarios: short-term and long-term prediction. In the short-term prediction, we take the first 60 min of retweeting dynamics as training samples to predict the information cascade size at the 75th min. In the long-term prediction, we take the same training samples to predict the information cascade at the 4380th min. For each target, we retrain our prediction model; 26 955 Weibo original tweets posted in 2015 are randomly divided into two groups: 80% for training and 20% for testing. Figure 1 shows the prediction performance and the loss of the predictive power of different feature groups. In the short-term prediction, features of early popularity dynamics outperform other groups of features. In the long-term prediction, the RMSEs of all feature groups substantially increase, and the most significant growth is observed in early popularity dynamics, despite that it is still the most informative group of features. In summary, features of early popularity dynamics not only are the key predictors of cascade size but also lose their predictive power most severely from short-term to longterm prediction. The results from the last section showed two apparent trends. First, the accuracy of the prediction tasks drops sharply when the prediction target becomes remote. Second, the most informative features for short-term prediction (i.e. the early popularity dynamics) lose their predictive power most significantly in the long-term prediction tasks. Considering that the structural properties of the royalsocietypublishing.org/journal/rsos R. Soc. Open Sci. 8: 202245 following and retweeting networks and the tweet content features have very little predictive power, we conjecture that the difficulty of predicting long-term information cascade lies in the decay of the autocorrelation between the early and future popularity. The possible factors that break such a link may come from many different aspects. Several previous works showed that the bursty nature of online content popularity evolution and the dynamic change of user online activity can break the correlation between early and future popularity [11, 12, 19] . On the basis of prior works, we further explore the unpredictable events in information diffusion processes and summarize several common scenarios in which the actual tweet popularity largely diverges from the predicted scales. Burst is a common phenomenon that exists in many real systems. It can be characterized by short periods of intense activity followed by long times of no or reduced activity [44, 45] . The appearance of bursts can break the correlation between early-stage popularity and long-term popularity in information diffusion. For example, two original tweets with similar early-stage popularity can either experience a burst in the future propagation or not ( figure 2a) . Burst can be characterized by peaks in the diffusion curves [46, 47] . Here, we design a peak detection algorithm. Given an equally spaced time series that describes the net growths of cascade size of an original tweet C(t) = {c(t i ), c(t i+Δx ), … , c(t i+nΔx )}, a peak is defined as the local maximum of the timeseries segment within a fixed window size w, namely peak = max{c(t j−w ), c(t j−w+1 ), … , c(t j+w )}. Besides, the height of a peak is at least k times larger than the mean time series C ¼ ð1=nÞ P n l¼0 cðt iþlDt Þ. The difference in height between any two adjacent peaks should be greater than m. In practice, we set k = 4, w = 2 and m = 10 and perform a peak detection for all original tweets in the testing set. The original tweets with one peak in their diffusion curves account for 33.4%, those with more than one peak 52.6%, and no peaks 14.0%. The result demonstrates that most original tweets in the testing set exhibit one or more bursts in the information diffusion process. To test the impact of burstiness to the cascade size predictability, we compared the prediction errors between original tweets that experienced strong, weak and no burst in the cascade. An original tweet is considered to have experienced a strong burst if its largest peak height is higher than the average height of all peaks of all the original tweets' cascades. Other original tweets with at least one burst are royalsocietypublishing.org/journal/rsos R. Soc. Open Sci. 8: 202245 considered to have experienced weak bursts. In the end, 38% of the original tweets are tagged with strong bursts, 48% with weak bursts and 14% with no burst. The prediction performance, i.e. RMSE, for original tweets that experienced strong, weak and no burst at different times in their cascades are shown in figure 2b. One model is trained and tested for each group of samples, by randomly selecting 80% of data for training and the remaining 20% for testing. Evidently, the original tweets that experienced strong bursts have the lowest prediction performance, while the less severe or fewer bursts an original tweet experiences, the easier it is to predict its cascade size. Therefore, the strength of the bursts indeed leads to the (un)predictability of information cascades. The number of peaks in the information cascades may also exert influences on the prediction performance. Particularly, multiple bursts in the information diffusion increase the uncertainty of cascade propagation, making long-term popularity prediction more difficult. Figure 3a shows the real and the predicted diffusion curves of a typical original tweet with multiple peaks in the diffusion curve. The real cascade size increases to about 150 at the end of the first hour and then experiences two large and rapid increases, with the cascade size rising from less than 200 to about 1100. However, the predicted cascade size grows in a relatively flat trend, from less than 200 to about 700 steadily, resulting in a large error between the empirical value and predictions. To further explain the cause of the large prediction error of the cascade size of the original tweet in the above experiment, the cascade propagation tree of the original tweet and its net growth curve are drawn in figure 3b,c. In the retweeting network, the red, yellow and blue nodes represent the users who retweet the post, corresponding to the three peaks in the cascade size curve. In the first peak, node A contributes to the major number of retweets. The second peak is caused by two influential users (i.e. nodes B and C). The two influential users spread the information cascades to their followers respectively, forming two new clusters. The third peak occurs because the two influential users (i.e. nodes B and C) promote the spread of information cascades once again. Also, two new influential users (i.e. nodes D and E) appear and further improve the diffusion of the information cascades. royalsocietypublishing.org/journal/rsos R. Soc. Open Sci. 8: 202245 The unexpected retweet by a key opinion leader in the online social network may cause unforeseeable new cascades of the original tweets. These random events happen at a relatively low probability, and therefore diverge from the most probable scenarios modelled by the ML algorithms and cause large prediction errors. Similar to the previous section, we stratified the original tweets that experienced at least one burst into those which experienced only one burst and those which experienced multiple bursts. One model is trained and tested for each group of samples, by randomly selecting 80% of data for training and the remaining 20% for testing. Based on the previous result of peak detection, the original tweets with one peak in their cascade account for 52% while the tweets with multiple peaks occupy 48%. Figure 3d shows that it is easier to predict the cascade size for original tweets that experienced only one peak in their cascades. This finding confirms that the complex burstiness in information cascades hinders the predictability of the cascade process. The prediction performance of XGBoost models for original tweets that experienced multiple peaks or a single peak in their cascades, respectively. Higher RMSE reflects lower prediction precision. Prediction performance for original tweets with multiple peaks in their cascades is lower than those with a single peak. royalsocietypublishing.org/journal/rsos R. Soc. Open Sci. 8: 202245 original tweets and retweet in the daytime rather than at midnight. In this study, we consider 0.00 to 8.00. as the low activity period and 8.00 to 23.00 as the high activity period. Figure 4b shows the real and predicted diffusion curves of an original tweet, which is posted around 5.00. The original tweet receives a small number of retweets after being posted for 1 h. With the passage of time, the tweet information cascades transits gradually to the periods of high user activity. From the 75th to 435th min, the real cascade size of the original tweet increases at a relatively steady rate, and the errors between the real and the diffusion curve are very small. From the 435th to 675th min, user activity reaches a relatively high level, while the predicted cascade size still increases at a relatively slow rate, resulting in a significant prediction error. These results show that the dynamic change of user activity leads to a significant change in cascade growth trends of early-stage and long-term popularity dynamics, which significantly reduces the prediction ability of the features built on the early popularity dynamics. To further test the difference in predictability for cascades that started in high and low activity periods, respectively, we stratified the original tweets by their posting time into two groups. One model is trained and tested for each group of samples, by randomly selecting 80% of data for training and the remaining 20% for testing. Figure 4c shows that it is easier to predict the cascade size for original tweets posted during a high activity period than those posted during a low activity period. This finding confirms that original tweets that are posted during a low activity period may experience unexpected cascading behaviour and diverge from prediction severely. Feature drift occurs when the validity of features in the data stream changes over time, requiring ML models to be retrained to ignore irrelevant features and consider the newly relevant ones [48] . royalsocietypublishing.org/journal/rsos R. Soc. Open Sci. 8: 202245 Barddal et al. [49] analysed the impact of feature drift on some ML algorithms and found that the existing ML algorithms did not solve the problem of feature drift. Therefore, when performing an ML task on the data with the feature drift phenomenon, the influence of temporal factors on the prediction results should be considered. To confirm the existence of feature drift in cascade propagation prediction, a predictive model is constructed and trained based on 10 000 original tweets published in 2015, and 3000 original tweets published in 2015 and 2016 are randomly selected for testing, respectively. There is no overlap between the training data and the testing data. Figure 5 shows the curve of RMSE, the prediction accuracy index, with the growth of spreading time when the prediction is made based on the testing sets of different years. As can be seen from the figure, the accuracy of the prediction result based on the testing data of 2015 is better than that based on the testing data of 2016, and this experimental result proves that the phenomenon of feature drift does exist in the prediction of cascade propagation. Therefore, in the cascaded propagation prediction using ML methods, it is necessary to construct the features with a strong time correlation with the prediction target variables; otherwise, the prediction accuracy will be greatly affected. This paper attempts to explore the possible causes of difficulties in predicting information diffusion in the long run. First, we use a robust algorithm (i.e. XGBoost) to build ML models using various features from the early-stage dissemination of original tweets in a popular online social network. Then, we compare the predictive power of different groups of features in the short-and long-term cascade sizes and found that the prediction error increases sharply with the prediction gap. The most predictive features (i.e. early popularity dynamics) in the short-term prediction tasks lose most of their predictive power in the long-term tasks. In the light that the early-stage information propagation dynamic features lose their predictive power in the long-term prediction tasks, we conjecture that the difficulty of predicting long-term popularity lies in the decay of correlation between early and future popularity. We design a peak detection algorithm to examine the unpredictable events in the information diffusion process and summarize two common scenarios in which the actual information cascades largely diverge from the predicted scales: (i) the burstiness of popularity; and (ii) the dynamic change of user activity. Also, the phenomenon of feature temporal drift reduces the validity of feature engineering we build, making long-term popularity more difficult to predict. Our findings have two-sided implications for the prediction and control of information diffusion in online social networks. On the one hand, we suggest that there is potentially an upper bound of accuracy when trying to predict future information cascade size using early-stage dynamics only. That is, early popularity does not mean long-term popularity and vice versa. The complex online environment may generate many unexpected factors that can severely affect the information propagation process and the diffusion scale. On the other hand, we suggest that the information diffusion process is generally royalsocietypublishing.org/journal/rsos R. Soc. Open Sci. 8: 202245 controllable. The controllability embodies in two ways: micromanagement during the diffusion process, e.g. the introduction, or the suppression, of a key opinion leader, which can boost an already started diffusion process or prevent undesired future diffusion; and changing the underlying online environment can also alter the information diffusion mechanism, hence information popularity in the new environment. Data accessibility. Data and relevant code for this research work are available at https://doi.org/10.5061/dryad.44j0zpccv [50] . The tube over time Attention prediction on social media brand pages From popularity prediction to ranking online news On popularity in the blogosphere Blocking information on COVID-19 can fuel the spread of misinformation Misinformation of COVID-19 on the internet: infodemiology study. JMIR Public Health Surveill. 6, e18444 A peek into the future Predicting tweet retweetability during hurricane disasters Proc. of the 21th ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining -KDD '15 Modeling and predicting popularity dynamics via reinforced Poisson processes Modeling and predicting retweeting dynamics on microblogging platforms Popularity prediction in microblogging network Predicting popular messages in Twitter Predicting the popularity of viral topics based on time series forecasting Predicting information diffusion on Twitter -analysis of predictive features Predicting the popularity of online content A novel time series approach for predicting the longterm popularity of online videos Using early view patterns to predict the popularity of youtube videos LARM: a lifetime aware regression model for predicting you tube video popularity Want to be retweeted? Large scale analytics on factors impacting retweet in Twitter network Can cascades be predicted? The untold story of the clones Predicting the popularity of web 2.0 items based on user comments Modeling and predicting popularity dynamics via reinforced Poisson processes SEISMIC: a self-exciting point process model for predicting tweet popularity Expecting to be HIP: Hawkes intensity processes for social media popularity Feature driven and point process approaches for popularity prediction Rt to win! predicting message propagation in Twitter Temporal prediction of retweet count Prediction of retweet cascade size over time Predicting the popularity of news articles Endogenous versus exogenous shocks in complex networks: an empirical test using book sale rankings Predicting the long-term citation impact of recent publications Page rank beyond the web Everyone's an influencer Who says what to whom on twitter An analysis of influential users for predicting the popularity of news tweets Document clustering: TF-IDF approach Latent dirichlet allocation Impact of multimedia in Sina Weibo: popularity and life span Predicting the lifespan and retweet times of tweets based on multiple feature analysis XGBoost Another look at measures of forecast accuracy Burstiness and memory in complex systems The origin of bursts and heavy tails in human dynamics The bursty dynamics of the Twitter information network Analysis of collective action propagation with multiple recurrences Heterogeneous ensemble for feature drifts in data streams Analyzing the impact of feature drifts in streaming learning Data from: Why cannot long-term cascade be predicted? Exploring temporal dynamics in information diffusion processes Authors' contributions. R.M.C., X.F.L. and X.K.X. designed the study. R.M.C. and X.K.X. designed the experiments. R.M.C.analysed the experimental data. X.F.L. and X.K.X. interpreted the results and wrote the manuscript. All authors gave final approval for publication.Competing interests. We declare we have no competing interests. Funding. This work was supported by the National Natural Science Foundation of China (61773091 and 61603073) , the LiaoNing Revitalization Talents Program (XLYC1807106), and the Major Project of the National Social Science Fund of China (grant/award no. 19ZDA324).