Michael Ekstrand

There is a bug when isolating train-test sets. If a list of train-test sets has `isolate: true`, then it does not correctly have any test entities, at least when using...

We should implement a 'hit rate' top-N metric that measures whether the recommendation list contains any test items. When user test sets have size 1, this is equivalent to recall,...

We need to support recommendation based on frequency of occurance of an item in a user's neighborhood, such as in http://delab.csd.auth.gr/papers/ESWA08snpm.pdf.

Remove nDCG predict metric

2

Our use of nDCG as a prediction accuracy metric is non-standard and should probably be dropped. Equivalent behavior can be obtained by using a `recommend` task with unlimited size and...

I would like to redo LensKit's parallelism to use fork/join pervasively, with a single pool parallelizing evaluation tasks, model builds, and evaluations themselves. The fork/join pool would be injected into...

It would be useful - for Samantha integration, at least - to support general-purpose APIs for scoring arbitrary entities with respect to other arbitrary entities. This will also require some...

Add average per-user coverage metric

1

Add a per-user coverage metric (average %covered over users).

Update normalizer APIs

1

We need to reconsider and update the normalizer APIs. Not sure we want to use them in their current form.

It may be useful for the crossfolder to be able to emit query or runtime data (see #1001). This is half of #561.

When the interaction entities have count attributes, we should support those counts in our popularity statistics.