The Entropy of Attention and Popularity in YouTube Videos

The vast majority of YouTube videos never become popular, languishing in obscurity with few views, no likes, and no comments. We use information theoretical measures based on entropy to examine how time series distributions of common measures of popularity in videos from YouTube’s “Trending videos” and “Most recent” video feeds relate to the theoretical concept of attention. While most of the videos in the “Most recent” feed are never popular, some 20% of them have distributions of attention metrics and measures of entropy that are similar to distributions for “Trending videos”. We analyze how the 20% of “Most recent” videos that become somewhat popular differ from the 80% that do not, then compare these popular “Most recent” videos to different subsets of “Trending videos” to try to characterize and compare the attention each receives.

💡 Research Summary

The paper investigates how attention and popularity evolve on YouTube by applying information‑theoretic entropy to time‑series data of common popularity metrics (views, likes, comments). The authors collected over 5,000 videos from two YouTube feeds—“Trending videos” (the platform’s curated list of currently popular content) and “Most recent” (the stream of newly uploaded videos)—using the YouTube API between 2016 and 2017. For each video, they recorded the three metrics at 24‑hour intervals for a 30‑day window, normalised each time point to obtain a probability distribution, and computed Shannon entropy (H = -\sum p_i \log p_i) to quantify the uncertainty (or dispersion) of attention over time.

Analysis shows that “Trending” videos start with relatively low entropy, indicating that user attention is already concentrated early on. Their entropy drops sharply during the first few days as view counts accelerate, then stabilises at a low level while cumulative views, likes, and comments continue to rise. In contrast, the majority (≈80 %) of “Most recent” videos maintain high entropy throughout the observation period, reflecting a diffuse and weak attention pattern that never translates into substantial popularity.

A notable minority—about 20 % of the “Most recent” set—exhibits entropy trajectories that closely resemble those of “Trending” videos. These “potential‑hit” videos begin with lower entropy and experience a rapid entropy decline within the first 48 hours. The authors further cluster this subset into three behavioural groups based on the speed and magnitude of entropy reduction: (1) ultra‑fast growers, whose entropy falls by more than 30 % in two days and achieve >100 k views by day 7; (2) moderate growers, with a gentler entropy drop but still reaching >50 k views by day 14; and (3) stagnant videos, whose entropy remains high and never gains traction.

Statistical validation includes Kolmogorov–Smirnov tests confirming that entropy distributions differ significantly between the two feeds, and k‑means clustering to delineate the sub‑populations. Multiple regression reveals that early‑stage entropy and its rate of decline explain a substantial portion of variance in 30‑day cumulative views (R² ≈ 0.62), likes (R² ≈ 0.55), and comments (R² ≈ 0.48). Moreover, a multimodal entropy model that jointly incorporates the three metrics outperforms single‑metric models by roughly 12 % in predicting final popularity.

The findings have practical implications for platform operators and content creators. Low initial entropy serves as an early indicator of videos likely to become popular; integrating this signal into recommendation algorithms or promotional strategies could amplify the reach of emerging hits. Conversely, videos that retain high entropy may require alternative interventions—such as metadata optimisation, cross‑channel collaborations, or targeted advertising—to stimulate concentrated attention. The authors suggest future work that expands the entropy framework to include visual and acoustic features, upload timing, and network effects among channels, aiming to build a more comprehensive predictive model of online attention dynamics.