timonvs / beam-recommender Goto Github PK

Random forest search can be used to find the optimal hyperparameters (optimal amount of factors, epochs and regularisation).

Implement a way to evaluate the recommender system (precision@k, recall@k, etc.)

Implement a dashboard that can be used to train and evaluate the recommender system.

We need to rethink the implementation of the Beam Discover user interface. This was made clear by insights that were gathered over the last few months and the Beam Discover meeting that was held last week (meeting notes).

Current design:

Related links:

• Future of Beam Discover
• Beam Discover meeting notes (14-01-2017)

Requirements

Requirements that are unclear or not confirmed are marked with (?)

• Discover in search overview
  • Or in notifications (?)
• Convey when recommendations were generated
  • Recommendations are not real-time, they will be generated in a certain interval (e.g. once a day or once a week).
• Browse categories/topics
• Show categories/topics as explanation for recommendations (?)

The verbose mode is useful for exploration, evaluation and debugging.

• Print which subreddits the user "liked" recently
• Print the current epoch while training/evaluating
• ...

Recently interacted subreddits should get a higher weight.

• Is using just Beam user data enough?
  • How many users do we need?
• How can we correctly combine Beam user data with Reddit user data (just comments)?

Should we normalise/threshold implicit ratings? It may or may not give a better result, especially when combining it with the Reddit comments data.
If we threshold the ratings we should also use weighting (subtract average amount of interactions per user per subreddit).