z****d
发帖数: 14 | 1 Most consumer brands outsource their digital ad buying to some type of
intermediary (called a broker, media buyer, network, or trade desk). In one
common such arrangement, the brand agrees to pay the intermediary a specific
dollar amount per click and leaves it up to the intermediary to decide
where to buy and how much to pay per digital ad or impression. We call
advertising campaigns that buy using this logic "performance campaigns."
Brokers often carry out this arbitrage by buying impressions (digital ads)
one by one on real-time ad exchange. When a user loads a web page with an ad
slot in it, a call is made to the exchange, an auction is carried out in
real time (<40 ms), the highest bid wins, and the corresponding brand or
broker sends their ad of choice to load in the originating browser.
Suppose you are a broker and you have performance data on N different
performance campaigns across M different websites over a moving 4 week time
window. Specifically, for each campaign-site-ymdh (year, month, day, hour)
combination, you have a record of the number of impressions won, clicks
obtained, and total cost accrued by the campaign on that website in that
hour. You also know the value that the advertiser will pay for a click;
assume this value is constant across sites for any given campaign.
Summary of data available
Columns:
campaign_id
site_id (a unique identifier for each url and/or mobile app)
ymdh
impressions
clicks
total_cost
click_value
Further, for simplicity, suppose that the sites in this set of data
represent all the possible sites that your campaigns can buy and that this
is all the data that is available to you. You may assume that all auctions
are "second price."
How would you design an algorithm for determining how much you would bid on
impressions when a campaign, with a daily budget B, had yet to win any
impressions from a site? How would this value change as the campaign began
to win impressions from these new sites?
* Hint: Note that in order to generate impression and click data for a
campaign, site combination, the campaign needs to bid enough to win on that
site.
* Further hint: consider the consequences of over and under-estimating the
click rate.
Optional: discuss how your algorithm would scale if it had to be
implemented with N = 200,000, M = 100,000. |
|
