Understanding Power Law Distributions in Content Ecosystems: Data Science Insights and Applications

The session, presented by Dr. Yu Yang from Tencent, introduces the role of data science in content ecosystems, outlining three main questions: what the data looks like, what can be done with it, and how the analysis is performed.

1. What does the data look like? The talk starts with a review of the normal distribution, illustrating its symmetric bell shape and typical examples such as exam scores and newborn weights. It then contrasts this with the heavy‑tailed Power Law, showing examples like city population sizes, earthquake depths, and startup valuations, and explaining why people perceive Power Law more strongly in daily life.

2. Why is Power Law so common? An abstract mechanism is described: human‑related networks tend to produce Power Law patterns because of shared preferences. The talk outlines three theoretical mechanisms—proportional random growth, transformations of Power Laws, and matching theory (e.g., CEO salary distribution)—supported by equations presented as images.

3. Content ecosystem data: the Power Law Real data from three products (short video, two information‑flow apps) shows that daily user consumption time and content flow both follow a heavy‑tailed distribution, where roughly 10% of content generates 90% of traffic.

What can we do? The presenter discusses thick‑tail distributions, their properties (Pareto principle, conditions for finite mean/variance), and risks such as biased A/B tests and unexpected 10‑sigma events. Applications include web search (Zipf’s law) and the survival of niche bookstores under the 80/20 rule.

How we do it – a content middle‑platform enables full‑link optimization. Examples include time‑series forecasting for article volume, integer programming to select high‑value creators, and probabilistic attribution for publishing anomalies.

Graph‑theoretic methods are used to model creator expertise: tags become nodes, co‑occurrence creates edges weighted by odds ratios, and community detection reveals dominant topics, improving creator‑to‑business matching from 10% to 50% accuracy.

Settlement‑strategy impact is analyzed with Difference‑in‑Differences (DID). Findings show that increasing settlement rates boosts video publishing by 9% for new creators, while price cuts reduce retention of newly active creators by 13%, guiding policy adjustments.

Q&A The audience asks how to handle unknown distributions (use histograms, MLE, KS tests) and how to avoid unfair weighting of head samples in Power Law models; the answers emphasize careful analysis and balanced strategy design.