[Paper Review] Improving the Effectiveness of Content Popularity Prediction Methods using Time Series Trends

Flávio Jr. B. Figueiredo, Marcos André Gonçalves+1 more|arXiv (Cornell University)|Aug 29, 2014

Technology and Data Analysis13 references3 citations

TL;DR

This paper proposes a two-stage model that improves web content popularity prediction by identifying common time series trends in early engagement data and using linear regression to predict 48-hour user engagement (visits, Facebook likes, Twitter mentions). By incorporating distances to learned popularity trends, the method achieves 15–27% higher accuracy than state-of-the-art baselines, with the Mixed-Trend K-Means variant proving most effective and scalable.

ABSTRACT

We here present a simple and effective model to predict the popularity of web content. Our solution, which is the winner of two of the three tasks of the ECML/PKDD 2014 Predictive Analytics Challenge, aims at predicting user engagement metrics, such as number of visits and social network engagement, that a web page will achieve 48 hours after its upload, using only information available in the first hour after upload. Our model is based on two steps. We first use time series clustering techniques to extract common temporal trends of content popularity. Next, we use linear regression models, exploiting as predictors both content features (e.g., numbers of visits and mentions on online social networks) and metrics that capture the distance between the popularity time series to the trends extracted in the first step. We discuss why this model is effective and show its gains over state of the art alternatives.

Motivation & Objective

• To improve early prediction of web content popularity using only first-hour data.
• To address the challenge of predicting user engagement metrics (visits, likes, mentions) 48 hours after upload.
• To enhance prediction accuracy by modeling temporal popularity trends rather than relying solely on static features.
• To evaluate the effectiveness of trend-based modeling in comparison to existing baseline methods.
• To demonstrate that incorporating trend similarity improves robustness and performance in popularity forecasting.

Proposed method

• Time series clustering (K-Means or KSC) is applied to extract common popularity trends from early engagement data.
• Each web page's popularity curve is compared to the identified trends, and distance metrics are computed as predictive features.
• Linear regression models are trained using both content features (e.g., initial visits, mentions) and trend-distance features.
• The Mixed-Trend model combines content features with trend similarity, enabling specialized prediction for different popularity dynamics.
• Model training uses generalized cross-validation (GCV) for hyperparameter tuning, with evaluation on a held-out test set.
• Regularization (Ridge, Lasso) is tested but found not to improve performance over ordinary least squares regression.

Experimental results

Research questions

• RQ1 Can modeling temporal popularity trends improve early prediction of web content engagement?
• RQ2 How does incorporating trend similarity enhance prediction accuracy compared to baseline models?
• RQ3 Which clustering method (K-Means vs. KSC) provides better trade-off between performance and scalability for trend extraction?
• RQ4 Does the proposed model generalize well without overfitting on unseen data?
• RQ5 To what extent can early engagement patterns predict long-term popularity?

Key findings

• The Mixed-Trend K-Means model achieved 15–27% improvement in prediction accuracy over state-of-the-art baselines across all three response variables (visits, Facebook likes, Twitter mentions).
• The Mixed-Trend model outperformed all baselines, particularly for predicting number of visits, with a 27% improvement over the SH model.
• The model showed minimal overfitting, as test set RMSE on the challenge server was comparable or better than GCV-based estimates.
• K-Means and KSC produced nearly identical performance, but K-Means was preferred due to its superior scalability.
• Regularized regression (Ridge, Lasso) did not improve results over ordinary least squares, indicating no need for penalty in this dataset.
• The model's AIC and BIC values were consistently better than baselines, supporting its generalization and robustness.