PaperlayerView exact revision on arXiv

Source-linked AI summary

Predicting popularity of online videos using Support Vector Regression

Tomasz Trzcinski, Przemyslaw Rokita

arXiv:1510.06223v4cs.SIcs.CV

TL;DR

The paper addresses the challenge of predicting online video popularity despite its dependence on complex external, social, and content factors. It proposes Support Vector Regression with a Gaussian Radial Basis Function, combining temporal and visual cues. Across YouTube and Facebook datasets, the method outperforms the state of the art, and combined features reach a Spearman rank correlation of up to 0.94 six hours after Facebook publication.

  • Problem

    Online video popularity is difficult to predict because external context, publisher networks, viewer relevance, and real-world events influence distribution patterns.

  • Method

    The paper uses Support Vector Regression with a Gaussian Radial Basis Function, combining visual cues with early popularity and social-interaction patterns.

  • Results

    The method outperforms the state of the art across evaluated datasets, with combined visual, social, and early-view features reaching a Spearman correlation rank of up to 0.94 six hours after Facebook publication.

  • Takeaways & Limitations

    Early distribution patterns can be improved for online video popularity prediction by adding visual and social metadata.

Abstract

from arXiv · show

In this work, we propose a regression method to predict the popularity of an online video based on temporal and visual cues. Our method uses Support Vector Regression with Gaussian Radial Basis Functions. We show that modelling popularity patterns with this approach provides higher and more stable prediction results, mainly thanks to the non-linearity character of the proposed method as well as its resistance against overfitting. We compare our method with the state of the art on datasets containing over 14,000 videos from YouTube and Facebook. Furthermore, we show that results obtained relying only on the early distribution patterns, can be improved by adding social and visual metadata.

I. INTRODUCTION

Online video popularity is difficult to predict because it depends on external context, publisher networks, viewer relevance, and real-world events. The paper proposes Popularity-SVR and combines early popularity patterns with visual and social cues, while releasing a dataset of over 1,800 videos.

  • Motivation: Online video popularity depends on external context, publisher-network structure, viewer relevance, and relationships between content and real-world events.These factors make future popularity difficult to capture from a single information source.
  • Research gap: Combining visual, textual, social, and early-distribution cues received limited attention for online video popularity prediction.Prior methods typically relied on intrinsic content, social-network, or early-distribution features separately.
  • Approach: Popularity-SVR uses Support Vector Regression with a Gaussian Radial Basis Function kernel to predict online video popularity.The method incorporates visual cues available before publication and early popularity patterns after release.
  • Contributions: Adding visual and social features to early distribution patterns improves popularity prediction results.Examples include the number of faces shown throughout a video and the number of recorded comments.
  • Contributions: Over 1,800 online videos from the largest social network were collected and released publicly with corresponding temporal and visual features.The dataset supports evaluation and reuse of the proposed prediction approach.

II. RELATED WORK

Prior work predicted online-content popularity from access patterns, temporal evolution, social context, or intrinsic content cues. This paper combines visual features with early evolution data for longer social-media videos, addressing a comparatively underexplored setting.

  • Early popularity prediction: Earlier popularity research examined user web-access and video-distribution patterns, partly to understand caching benefits for Internet traffic.These studies established access-pattern analysis as an early direction in popularity prediction.
  • Video popularity prediction: Most video-popularity methods used early evolution patterns or social context, without exploiting additional visual cues to improve prediction accuracy.This contrasts with the paper’s multimodal feature strategy.
  • Video popularity prediction: Szabo and Huberman reported a Pearson correlation exceeding 0.9 between log-transformed YouTube views at seven and thirty days.Their observation motivated using early views as indicators of later popularity.
  • Video popularity prediction: The Multivariate Linear model extended early-view prediction with multiple inputs, while an RBF-based Ridge regressor reduced prediction error by 20% on average relative to the referenced method.The passage describes these as extensions of the earlier log-linear approach.
  • Social context: Social-Forecast added publisher-network information and optimized a trade-off between prediction accuracy and forecast timing.Its approach used viewer-sharing proportions and follower counts.
  • Paper positioning: The paper combines computer-vision features with early evolution data for longer videos, whereas cited multimodal work focused on micro-videos lasting only a few seconds.The authors describe this as one of the first attempts in this context.

III. FEATURES

The paper extracts visual features from raw video using computer-vision algorithms, including video characteristics, color, faces, text, scene dynamics, clutter, rigidity, thumbnails, and deep representations. These features provide cues for predicting popularity before publication.

  • Feature extraction: Visual features are extracted from raw video with computer-vision algorithms and are designed to be available before online publication.The feature set includes both conventional descriptors and deep visual representations.
  • Video characteristics: Video characteristics encode length, frame count, resolution, and frame dimensions.These descriptors capture basic properties of the video signal.
  • Color: Color features summarize dominant frame colors using a histogram over ten Hue-Saturation-Value color classes.The representation also records the dominant video color.
  • Face: Face features average detected face counts, face-containing frames, and face-region size across video frames.Faces are detected with a cascade-classifier-based detector.
  • Text: Text features measure the proportion of frames containing imprinted text and the average text-region size.Edge detection, morphological filters, and Tesseract-OCR validate text-region detection.
  • Scene dynamics: Scene-dynamics features describe shot boundaries, shot counts, average shot length, cut types, and shot histograms.Shot boundaries are obtained with the Edge Change Ratio algorithm.
  • Deep features: Deep features average ResNet-152’s 1000-dimensional probability outputs over thumbnails representing the video’s scenes.The resulting representation supplies additional visual cues for prediction.

B. Temporal features

After publication, the paper collects time-varying popularity information from views and social interactions. These temporal evolution patterns serve as prediction features.

  • Temporal features: Temporal features are information that changes over time, including aggregated views and received likes.They are recorded after the video is published online.
  • Views: Views are an aggregated count of how many times a video has been watched, increasing with time.This feature directly tracks accumulated viewing activity.
  • Social features: Social features aggregate likes, shares, and comments, whose evolution patterns provide cues for popularity prediction.The paper uses these patterns alongside early view information.

IV. METHODS

The paper casts online-video popularity prediction as regression: estimate views at a later target time from features available during an earlier reference period. It reviews univariate and multivariate linear approaches before introducing the proposed nonlinear method.

  • Popularity prediction estimates a video's views at target time tt from features available during the first tr days after publication.
  • Univariate Linear (UL) Regression: UL regression uses the logarithm of early cumulative views and a learned weight to predict later popularity.The predicted count is expressed through a natural exponential function.

Multivariate Linear (ML) Regression

ML regression represents early popularity evolution with multiple view increments, while MRBF adds nonlinear similarity features based on representative training videos. Ridge regression is used because the added features increase overfitting risk.

  • Multivariate Linear (ML) Regression: ML regression forms a feature vector from view increments sampled between publication time t0 and reference time tr.
  • Multivariate Radial Basis Function (MRBF) Regression: MRBF combines ML regression with RBF features that measure distances between videos and selected representative training videos.
  • Multivariate Radial Basis Function (MRBF) Regression: The Gaussian RBF uses a σ parameter and a video set C selected during cross-validation.
  • Multivariate Radial Basis Function (MRBF) Regression: Ridge regression replaces ordinary least squares because the additional RBF features increase the risk of overfitting.

B. Popularity-SVR

Popularity-SVR predicts future video popularity with Support Vector Regression and a Gaussian RBF kernel. The kernel maps feature vectors into a nonlinear space, using support vectors and coefficients to model evolution patterns without explicitly assuming a linear relation.

  • Popularity-SVR predicts future video popularity using Support Vector Regression with a Gaussian RBF kernel.
  • The Gaussian RBF maps feature vectors into a nonlinear space where popularity-evolution relations are easier to capture.
  • The model uses a feature vector available at reference time tr, support vectors, coefficients, and an intercept.
  • Unless otherwise stated, the features are log-transformed view deltas observed before the reference time.
  • Grid search selected C = 10 and σ = 0.005 for the reported experiments.

V. RESULTS

The experiments compare UL, ML, MRBF, and Popularity-SVR on YouTube datasets using rank correlation at a 30-day target. The reported results emphasize higher performance and more stable accuracy for Popularity-SVR.

  • The evaluation compares UL, ML, MRBF, and Popularity-SVR on three datasets containing almost 24,000 videos.
  • At tr = 6 days and tt = 30 days, Table I reports Spearman rank correlation with 95% confidence intervals for the four methods.
  • Popularity-SVR outperforms the competitors while providing more stable prediction accuracy, indicated by a smaller confidence interval.
  • For both Random and Top YouTube datasets, Popularity-SVR outperforms state-of-the-art methods, with greater improvement for tr < 12.

A. Datasets

The study evaluates popularity prediction on large YouTube and Facebook video datasets using temporal, social, visual, and view-related information. Performance is assessed with 10-fold cross-validation and Spearman rank correlation, with Popularity-SVR showing higher and more stable results on YouTube.

  • YouTube datasets: 16,132 Random and 5,811 Top YouTube videos remained after removing incomplete statistics and videos with less than 30 days of data.The original datasets contained 24,484 randomly selected and 27,212 top-ranked videos.
  • Facebook dataset: 1,820 Facebook videos were collected from several publishers between August 1 and October 15, 2015.The collection used a crawler based on the Facebook Graph API to retrieve publicly available interaction information.
  • Evaluation: 10-fold cross-validation trains on nine folds and tests on the remaining fold, repeating this across all folds.The reported results are averages across the ten test sets with corresponding 95% confidence intervals.
  • Evaluation: Spearman rank correlation is used to evaluate prediction accuracy.This metric follows the evaluation approach of prior work.
  • Results: Popularity-SVR outperforms competing methods across YouTube datasets and reference times, with up to 40% smaller 95% confidence intervals.The advantage is especially significant for popular videos shortly after publication, while methods converge as the reference time approaches the target time.

D. Facebook dataset

On Facebook videos, visual, temporal, social, and view-based inputs are compared using Spearman correlation. Popularity-SVR performs best across configurations, with the strongest results from combining view evolution, social interactions, and visual features, although correlated social signals add only modest information beyond views.

  • Visual features: Deep visual features provide the highest visual-feature accuracy, while combining all visual features yields a Spearman correlation above 0.23.Clutter, scene dynamics, and thumbnail popularity rank are also identified as useful visual measures.
  • Feature combinations: The highest Facebook prediction accuracy uses view counts, social features, and visual features together.Among separate input groups, view counts perform best, followed closely by social features; visual features are available before publication.
  • Feature comparisons: Social signals are substantially stronger predictors of online-video popularity than visual signals in social media.The paper presents this comparison as empirical evidence from the Facebook results.
  • Feature comparisons: Views correlate with comments, shares, and likes at 0.86, 0.88, and 0.93, respectively.This multicollinearity explains why adding social features produces only a relatively small information gain beyond view evolution.
  • Statistical comparison: Popularity-SVR outperforms competitors across input configurations and reference times, with mean p-values below 0.02 against baseline outputs using view counts.The approximately 1% improvement over MRBF with all features motivated the additional significance analysis.

E. Runtime evaluation

The runtime evaluation compares Popularity-SVR with competing methods as dataset size increases. Popularity-SVR offers competitive or lower training time and faster prediction than MRBF, while prediction time remains stable.

  • Training time: Popularity-SVR has lower training time than the other methods except Univariate Linear Regression.The measurements use subsets of different sizes from the YouTube Random dataset and average results over 10 runs.
  • Prediction time: Popularity-SVR predicts faster than the competing MRBF method.The authors attribute this to a simplified prediction model using support vectors and the RBF kernel without MRBF’s additional ML regression term.
  • Scalability: Increasing training size increases training time for all evaluated methods, while prediction time remains fairly stable.This pattern supports the scalability of the evaluated methods.
  • Evaluation setup: The runtime comparison evaluates methods on YouTube Random dataset subsets of different sizes.Training time increases with subset size, whereas prediction time remains stable across sizes.
Loading 1510.06223v4…