TrendLearner: Early prediction of popularity trends of user generated content

Abstract

Predicting the popularity of user generated content (UGC) is a valuable task to content providers, advertisers, as well as social media researchers. However, it is also a challenging task due to the plethora of factors that affect content popularity in social systems. Here, we focus on the problem of predicting the popularity trend of a piece of UGC (object) as early as possible. Unlike previous work, we explicitly address the inherent tradeoff between prediction accuracy and remaining interest in the object after prediction, since, to be useful, accurate predictions should be made before interest has exhausted. Given the heterogeneity in popularity dynamics across objects, this tradeoff has to be solved on a per-object basis, making the prediction task harder. We tackle this problem with a novel two-step learning approach in which we: (1) extract popularity trends from previously uploaded objects, and then (2) predict trends for newly uploaded content. Our results for YouTube datasets show that our classification effectiveness, captured by F1 scores, is 38% better than the baseline approaches. Moreover, we achieve these results with up to 68% of the views still remaining for 50% or 21% of the videos, depending on the dataset.

Early Prediction of Scholar Popularity

“…After applying the β CV heuristic in our data, we identified 5 different clusters, as further discussed in Section 5. Regarding the parameters of Algorithm 1, namely vectors θ and γ, we adopt the same parametrization approach as in [12]. Notably, we apply an One-Vs-All classification (OVA) algorithm [4] for all classes separately.…”

“…Using the selected value for γ [i] , the minimum confidence θ [i] is the average probability computed for all scholars in class K i . Regarding parameter γ max , we set its value equal to the total number of points in the popularity time series (i.e., 20), as done in [12]. Table 2 shows the best parameter values obtained following the aforementioned procedure for each of the 5 identified classes.…”

“…First they classify UGC objects (e.g., videos) based on their popularity trends and then predict the popularity of that object in the future. Unlike those studies, our method considers the tradeoff between remaining citations after prediction and prediction effectiveness, adapting a model called TrendLearner, previously proposed to the context of UGC popularity trend predictions [12], to the particular context of scholar popularity trends. Like TrendLearner, our approach determines the duration of the period during which each scholar should be monitored before prediction on a per scholar basis, while other studies such as [2,6,29,26,19] considered fixed monitoring periods for all objects/scholars.…”

“…However there is no simple manner to find the optimal q. Thus, as in [16,12], we search for the optimal value of q considering all integers in the range (−n, n), where n is the length of the input vectors p x and p y . Note that K-SC requires all time series have the same size n. Thus, we represent each scholar by a vector with n elements where each element represents the scholar popularity (i.e., number of citations) in one year.…”

“…We tackle the problem of predicting popularity trends of individual scholars by applying a novel two-step learning approach called TrendLearner which was originally proposed for user generated content (UGC) on the Web [12]. In this approach the popularity trends of UGC are predicted based on a tradeoff between prediction effectiveness and remaining interest in the content after prediction.…”