Video Ad Engagement Prediction: 3 Million Labeled Impressions Dataset

Description

This dataset consists of 3 million labelled advertising auction lines, aimed at fostering advancements in Machine Learning, particularly in user engagement prediction with video ads.

This dataset is a product of extensive work by Cyrille Dubarry and was initially used for a Machine Learning class competition at École Polytechnique.

Objective

This dataset is designed to facilitate the prediction of the duration for which a user will engage with a video advertisement. Each entry in the dataset, marked by a unique AuctionID, represents an individual ad impression and includes a variety of contextual information about the user, publisher, and advertiser.

Content and Features

• auction_id: unique id for identifying each line
• timestamp: the timestamp (in seconds) of the ad impression
• creative_duration: the total duration of the video that has been played
• campaign_id: the advertising campaign id
• advertiser_id: the advertiser id
• placement_id: the id of a zone in the web page where the video was played
• placement_language: the language of this zone
• website_id: the corresponding website id
• referer_deep_three: the URL of the page where the video was played, truncated at its 3rd level
• ua_country: the country of the user who watched the video
• ua_os: the user's Operating System
• ua_browser: the user's internet browser
• ua_browser_version: the user's browser version
• ua_device: the user's device
• user_average_seconds_played: the average duration the user watched video ads in the past. It can be null if the user never watched any ad.
• seconds_played: the observed time the video has been watched. This is the quantity we are trying to predict.