Decided on a recommendation system.

[dscolby]

Need to figure out which learning paradigm, package(s), and inputs we need.

[dscolby]

There are basically two approaches that I think would work here.

1. Slate recommendations from multi-armed contextual bandits

• Overview: Use features about a user to select a set or slate of events to recommend, observe a reward, and use the reward to update the parameters of our model.
• Pros: Pretty simple to understand and not much computation per update.
• Cons: has to constantly be running as users rate events and requires more features, including user location information.

3. Matrix factorization

• Overview: Basically a variant of the Netflix Challenge where we would impute missing entries in a user by event matrix where each entry is a user rating for an event. Using this we would learn a model to predict user interaction with events.
• Pros: Pretty simple to implement, only requires user ratings, and does not require constant data queries to update parameters.
• Cons: Doesn't use any context and can be computationally expensive if the matrix is very large.

Right now I'm leaning towards the matrix factorization approach but am open to other ideas and questions.

In either of these approaches, though, we will probably need to group events together. For example, if we were recommending movies, we would have users watching a movie many times> But in our case, an event only happens once, so using it as a column/arm to recommend when it will not happen again would be a waste of time and resources. My guess is that we would want to use some kind of clustering/embedding/PCA to group similar events together.

[dscolby]

We will go with the matrix factorization approach using neural collaborative filtering but also adding other features besides interactions to the user and event vectors. At inference time, we will just take the top k predictions and make sure they have not already occurred.

[dscolby]

Update, we are actually going to use a Deep Factorization Machine. It is similar to the NCF but more expressive and easier to make recommendations because it doesn't need a user or item ID.