Skip to main content
dstillery dstillery
dstillery

9.8.16

Read the full article on forbes.com 

How is machine learning impacting digital advertising? originally appeared on Quora: the knowledge sharing network where compelling questions are answered by people with unique insights.

Answer by Claudia Perlich, Chief Scientist Dstillery , Adjunct Professor at NYU, on Quora:

Machine learning is virtually everywhere behind the digital advertising scenes. There have been a number of ADKDD workshops in conjunction with KDD, one of the oldest machine learning conference. The advertising environment is (for the better or worse) very data rich, and the rise of programmatic buying and selling of ads in real time has provided ample opportunity for machine learning to play a key role. For a one hour overview, please see my recent talk at the Institute of Advanced Study here.

For the most part, machine learning is being applied to a number of different components of advertising:

  • Measurement and attribution (market mix models, causal modeling from observational data, propensity matching, etc.)
  • Cross device association (predicting the probability that two devices belong to the same person based on usage patterns, IP overlap, etc.)
  • Intent prediction (what is the probability that a consumer is going to buy that new car in the next month or so) on an individual level.
  • Response prediction on an ad impression level (probability of a click or a video completion).
  • Fraud detection (how can I tell a bot from a real person, spoofed URLs from real ones, click fraud from true interest, etc.)
  • Audience insights (looking at a model that is good at predicting intent; can I extract some behavioral patterns for instance to inform creative design).

 

Here are a few more controversial topics on the role and effect of machine learning in advertising. Work by Professor Sweeney suggests that machine learning algorithms can potentially reflect racial discrimination because they will reflect existing racial biases of our society.

Some of my recent work points at the tendency of predictive models to ‘go where the signal is’, and in the business of weak metrics that can be problematic. Consider for instance predicting clicks on an ad. It turns out that it is much easier to predict that you are very likely to click on an ad on the flashlight app. Not because you are interested, but because you are fumbling in the dark. While being very good at optimizing towards a high click rate, the model is only going to reach people who are almost surely not interested in the product.