1. August 2001 Inventory Management in Internet Advertising Tom Shields HTPS – October 15, 2001

2. Inventory Management Web sites and ad serving systems have collected data on users for years With the internet advertising downturn, advertisers are demanding more targeted ads Ad sales people need to know availability of impressions meeting arbitrary target criteria Ad servers need to be able to fulfill promises made by salespeople

3. Requirements Query availability across arbitrary criteria Response times in 5-10 sec range Maximize successful delivery to flight goal Handle large data sets –Large systems serve over 1 billion ads per day –200,000 ad flights, up to 10,000 simultaneously –Many dimensions, some of high cardinality, e.g. Zip Common dimensions –CZAG (Country, Zip, Age, Gender) –Marital status, household income, education level –Reported interests –Site, section, or page –Keyword or keyword pairs

4. Example Forecast impressions: –From Men ages 18-25: 20,000 –From Men in California: 15,000 Sell 15,000 impressions to Men 18-25 in P&G Query: How many avails for Men in California? Forecast overlap: Men, 18-25, in CA: 6,000 Calculate impact: ¾ of overlap already sold Query answer: 9,000 non-overlap + 1,500 in overlap If we sell 7,500 of Men in CA, only 1/5 can be 18-25, even though proportionately 2/5 should be

5. The Problems How to represent the impression data Standard forecasting issues: –Time of day, day of week, seasonal variations –Discontinuous growth due to “site of the day” Algorithm for calculating avails for any set of criteria Algorithm for discounting existing flights from availability Algorithm for delivery that maximizes probability that all goals will be met

6. Attempted Solutions Large multi-dimensional database –Couldn’t afford the hardware and software to scale it Sampling –Sample size needed to be large to handle high cardinality dimensions –Still requires table scans and lots of computation Simulation –Accurate, but very slow

7. Today Where are we now? –Calculate first level of overlap between flights –Add in 10% “fudge factor” to handle the rest –Make up flights necessary about 10% of the time What should we be doing?