Michael Ekstrand
Michael Ekstrand
When the interaction entities have count attributes, we should support those counts in our popularity statistics.
We should create a general-purpose Lucene-based CBF recommender. Needed: - recommender logic, possibly built on item-item CF - a configurable model builder that knows how to take various item attributes...
It would be useful to allow, in at least certain circumstances, nullable dependencies on components whose defaults have unsatisfiable dependencies to be null rather than result in a resolution failure....
We would like a traditional SVD recommender in `lenskit-svd`. This can be built on the `BiasedMFModel`, using the SVD solver from `commons-math3` (and a `BiasModel` to provide the biases to...
It'd be nifty for the evaluator to be able to update a Slack channel with the status of an in-progress experiment. The Slack webhook integration is perfect for this. This...
We can implement sequential RMSE by preparing data sets for the train-test evaluator. For day _d_, the training data will consist of all ratings prior to _d_, and the test...
We need to add timestamp limiting to the new DAO infrastructure (in the static file DAO provider), possibly along with support for packed data.
We need to be able to have live updates for FunkSVD again, in a well-debugged and documented fashion. Either gradient descent, like we have been doing, or using `commons-math3`'s least-squares...
The CLI would benefit from a `run-script` command that loads a recommender, and then runs a Groovy script on it. This would be great for custom data export, etc.
Proposal: factor out a `lenskit-util` module that `lenskit-core` and others will depend on. The goal is to make `lenskit-core` primarily concerned with two things: - Configuration and lifecycle support -...