1
$\begingroup$

Edit: increased generality.

I have an ad placement optimization problem and I am brainstorming to determine which ML techniques are well suited to it.

Basically, I have some objective that involves optimizing ad placements for a given customer based on historical data on impressions, interactions, revenue and viewership history. There are also certain constraints on the ad placements due to market, content, and repetition etc...

This history goes back over two years and covers a very large customer base. What I would like to do is use machine learning to optimize ad slot placement that maximizes impressions based on the aforementioned inputs.

This sort of problem can be solved using linear programming methods, but I could also see RNNs as a viable option. Any insights here would be greatly appreciated.

$\endgroup$
5
  • 1
    $\begingroup$ Are you sure you have a sequential environment here - does user state evolve meaningfully due to the actions taken in your data (e.g. are you tracking how many times an individual user has seen or clicked through an ad, and have a strict sequence of user behaviour for each individual)? $\endgroup$ May 2 '19 at 6:42
  • $\begingroup$ The state here is the historical viewership which will change over time but not due to the action taken. So perhaps it's not a sequential environment as you say. $\endgroup$
    – Sledge
    May 2 '19 at 11:47
  • 1
    $\begingroup$ Yes, sounds more like a contextual bandit problem, as there is not a strong sequence. Bandit problem optimisation is often about minimising "regret" during continuous learning, and is commonly used for web ad placements. It would seem well suited to address your concern in the penultimate paragraph. $\endgroup$ May 2 '19 at 12:17
  • 1
    $\begingroup$ There's still a question here about how to best use the historical data to pre-train your bot, but perhaps you should edit the question so that it doesn't presume a Reinforcement Learning solution - as you can pre-train RL from history, but it would not be a good match to your problem if someone answered that question for you $\endgroup$ May 2 '19 at 12:18
  • $\begingroup$ I rephrased the question to be less opinionated about methods. $\endgroup$
    – Sledge
    May 2 '19 at 12:52

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.