Predict Short-Video Engagement with Skewed Data

# Predict Short-Video Engagement with Skewed Data You need to predict a future engagement quantity for a short video. Explain what data you would request, how you would formulate the modeling problem, and how you would handle skewed data. The exact engagement event and forecast horizon are not supplied: identify the definitions you need and make any working assumptions explicit. Describe both model development and evaluation. Distinguish a skewed target distribution from unrepresentative sampling or an imbalanced binary label; these problems need different remedies. ### What a Strong Answer Covers - An explicit target event, prediction time, horizon, and distinction between total engagement and engagement conditional on exposure. - Features available at prediction time and protection against later video-performance information leaking into training. - A baseline and modeling choices appropriate to a count, rate, or other agreed target. - A diagnosis of the type of skew, appropriate loss or transformation, and evaluation of both typical and high-engagement videos. - Validation that reflects future use and separates content quality from the platform's exposure policy. ```hint Separate exposure from response A video can receive more engagements because it is shown more often, because viewers respond more often, or both. ``` ### Follow-up Questions - What changes if predictions are made before a video has received any impressions? - Why might a log-transformed target improve typical-video error but underestimate total engagement? - How would you test whether good offline results are driven by the historical recommendation policy?