Source-linked AI summary
Predicting the popularity of online content
Gabor Szabo, Bernardo A. Huberman
TL;DR
The paper addresses the difficulty of predicting how much attention individual online submissions will receive over time. It models early view and vote accrual on Youtube and Digg, finding that Digg popularity reaches about 10% accuracy within two hours while Youtube requires about 10 days, with errors larger for evergreen content.
Problem
The paper asks whether the future attention devoted to individual online content can be predicted from early access measurements.
Method
The authors model popularity as views on Youtube and diggs on Digg, using log-scale correlations and three prediction models to forecast later popularity from early measurements.
Results
About 10% prediction accuracy is reached within two hours for Digg and 10 days for Youtube; Digg predictions converge faster, and the CS model outperforms two alternatives on both portals.
Takeaways & Limitations
Prediction is more accurate when attention decays quickly, whereas evergreen content produces larger errors; Digg stories change little after two days.
Takeaways & Limitations
The study does not explore how prediction differs across portal sections or whether Digg popularity can be forecast when votes come from a small known group of users.
Abstract
from arXiv · showhide
We present a method for accurately predicting the long time popularity of online content from early measurements of user access. Using two content sharing portals, Youtube and Digg, we show that by modeling the accrual of views and votes on content offered by these services we can predict the long-term dynamics of individual submissions from initial data. In the case of Digg, measuring access to given stories during the first two hours allows us to forecast their popularity 30 days ahead with remarkable accuracy, while downloads of Youtube videos need to be followed for 10 days to attain the same performance. The differing time scales of the predictions are shown to be due to differences in how content is consumed on the two portals: Digg stories quickly become outdated, while Youtube videos are still found long after they are initially submitted to the portal. We show that predictions are more accurate for submissions for which attention decays quickly, whereas predictions for evergreen content will be prone to larger errors.
1. INTRODUCTION
Online services have made content production and distribution widespread, but attention is concentrated unevenly and remains difficult to predict for individual submissions over time.
- Web 2.0 services enable content producers to reach very large audiences through sharing, blogging, social bookmarking, and collaborative platforms.
- Most content receives limited attention, while a small fraction attracts the bulk of views and downloads.
- Although aggregate attention distributions can be predicted, the amount devoted over time to individual items has been difficult to forecast.
- The study examines attention dynamics on Digg, where users vote on linked submissions, and Youtube, where users upload and interact with videos.
2. SOURCES OF DATA
The study measures popularity as views on Youtube and votes on Digg, using portal-specific access data and submission samples to model subsequent content popularity.
- Popularity is defined as video views on Youtube and votes, or diggs, collected by Digg stories.
- 7,146 Youtube videos were sampled from the recently uploaded listing to obtain an unbiased set with complete lifetime view histories.
- Youtube view counts were collected daily beginning April 21, 2008, with linear interpolation used to estimate counts between measurements.
- Digg submissions begin in an upcoming section, where users vote and sufficiently fast-growing stories can be promoted to the front page.
- Digg data comprised about 60 million diggs from 850 thousand users across approximately 2.7 million submissions.
- Among 1,321,903 Digg submissions during the collection period, 94,005, or 7.1%, were promoted to the front page.
3. DAILY CYCLES
Digg activity varies substantially by time of day and week, so the study adjusts submission age using activity-based digg time to make early popularity comparisons more reliable.
- Digging, story submission, and promotion rates show daily and weekly cycles, with weekdays about 50% more active than weekends.
- Stories submitted during inactive periods initially receive fewer diggs, which can make equally interesting stories appear less promising.
- Digg time measures submission age by the number of diggs cast on promoted stories rather than by wall-clock seconds.
- Because 72% of diggs go to promoted stories, one digg hour is defined as the time required for 5,478 new diggs to arrive there.
4. PREDICTIONS
The paper models early-to-late popularity using logarithmic correlations and evaluates prediction schemes for Digg stories and Youtube videos. Correlations emerge earlier for Digg, while prediction errors decrease faster for Digg than Youtube and depend on the chosen error measure.
- Correlations between early and later times: 11% of Digg stories receive most of their eventual diggs within their first hour, forming a slower-growing cluster distinct from the majority.The two clusters become indistinguishable after approximately the seventh digg hour, so subsequent calculations use the upper cluster.
- Correlations between early and later times: 0.98 is the Pearson correlation coefficient for Digg after 5 digg hours, rising to 0.993 after 12 digg hours; Youtube reaches 0.92 after about 5 days.For Digg, measuring age in digg hours produces stronger correlations than measuring age in absolute time through approximately the twelfth hour.
- Modeling popularity: Log-transformed popularity exhibits strong early-to-late correlations, motivating a linear prediction model with additive noise on the logarithmic scale.The model relates popularity at two times through a submission-independent linear relationship plus a zero-mean noise term.
- Model validation: Residuals approximately match normal distributions, although they are slightly right-skewed because highly popular content is overrepresented.Youtube residuals are also bounded from below because view counts can only increase.
- Prediction performance: QRE becomes negligible for Digg after 12 hours, whereas Youtube predictions converge more slowly to actual popularity.The CS model performs best for relative squared error, while the LN model is preferable for absolute squared error.
5. SATURATION OF THE POPULARITY
Popularity predictions converge faster for Digg than Youtube because Digg stories saturate quickly, whereas Youtube videos continue accumulating views and fluctuate near their reference values.
- Digg predictions converge faster than Youtube predictions to their respective reference values.
- Digg popularity grows quickly but saturates, while Youtube videos keep receiving views throughout the observed period.Digg stories saturate in about one day; Youtube videos continue gaining views throughout data collection, with an expected near-linear trend.
- Youtube’s relative squared error does not decrease monotonically until near the reference time, reflecting fluctuations even when popularity is nearly final.
- Different discovery mechanisms explain the saturation gap: Digg stories become obsolete quickly, while Youtube videos are found through search after leaving the recently added section.Digg content often concerns breaking news or fleeting fads; Youtube videos can remain discoverable through keyword search or related videos.
- Digg articles reach 10% accuracy within about 2 hours, whereas Youtube videos require about 10 days.Digg popularities do not change considerably after 2 days, while Youtube videos remain discoverable through search and related-video links.
6. CONCLUSIONS AND RELATED WORK
The paper develops and tests early-access models for predicting individual content popularity, compares them on Youtube and Digg, and discusses applications, assumptions, and unresolved scope questions.
- The paper presents a method and experimental verification for predicting content popularity from measurements taken soon after submission.
- A strong linear correlation links logarithmically transformed early and later popularities, enabling three prediction models for Youtube and Digg.
- Relative error measures have lower dispersion than absolute squared errors, so the paper recommends relative measures when estimating prediction error.
- Predicted popularity can support advertising-revenue estimation and content ranking, although ranking comparisons may be sensitive to prediction uncertainty.
- The predictions use observable popularity time series and do not model content semantics or why some submissions become more popular.
- The study does not examine how prediction varies across portal categories or whether Digg popularity can be forecast from a small number of users’ voting histories.
- Related work has characterized online video and media properties, streaming workloads, and Digg participation, ratings, activity, and moderation.