lenskit / lenskit Goto Github PK
View Code? Open in Web Editor NEWLensKit recommender toolkit.
Home Page: http://lenskit.org
License: Other
LensKit recommender toolkit.
Home Page: http://lenskit.org
License: Other
(originally reported in Trac by @elehack on 2011-04-05 22:36:41)
We would like to make the sparse vector framework copy-on-write.
The following value sharing strategy will be safe:
MutableSparseVector
from another vector always creates a fresh, unshared copy.ImmutableSparseVector
from another vector copies the value array and counter and increments the atomic counter.ImmutableSparseVector
decrements the counter.MutableSparseVector
that is shared (counter > 1) results in copying the value array, decrementing the counter, and creating a new counter initialized to 1.We can also allow sharing between mutable sparse vectors when one is created from another, as they will not be allowed to cross thread boundaries.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:53.246648+00:00, last updated: 2013-02-01T22:35:40.926617+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:34:46)
LensKit needs support for handling data updates in a consistent and well-defined fashion. This is necessary in particular for continuous update evaluation a la #13.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:38:44.755007+00:00, last updated: 2013-02-01T22:35:44.661125+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:24:57)
We need a design document explaining how LensKit's architecture works and why.
This document has been pretty much written, we just need to review it and make sure that it's up to date.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:38:32.658057+00:00, last updated: 2013-02-01T22:35:45.688634+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:38:13)
Consider supporting multiple evaluations being run in parallel (multiple algorithms or multiple folds).
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:38:50.162828+00:00, last updated: 2013-02-01T22:35:44.539723+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:26:52)
We want to be able to serialize and deserialize recommenders (or their models) so that a recommender can be built and then used by a different process.
It would also be interesting to explore alternative models (e.g. memory-mapped similarity matrices).
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:38:36.048971+00:00, last updated: 2013-02-01T22:35:45.578125+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:18:33)
A lot of APIs are not well-documented yet. We need thorough API documentation.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:37:02.246349+00:00, last updated: 2013-05-31T22:23:14.291139+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:23:29)
The cross-fold evaluator should support time-based instead of random splitting (try to predict the last third).
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:38:31.110415+00:00, last updated: 2013-02-01T22:35:45.815275+00:00
Placeholder for deleted ticket.
Note: This issue has been automatically migrated from Bitbucket
Created by Anonymous on 2013-04-01T00:00:00.0+00:00
(originally reported in Trac by @elehack on 2011-04-04 17:36:24)
It would be useful to have a JDBC-based implementation of RatingDataAccessObject
. It should be configurable w.r.t. what tables & fields to use.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:42.122763+00:00, last updated: 2013-02-01T22:35:41.541898+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:44:10)
Separate the algorithms into lenskit-knn
and lenskit-svd
projects that depend on lenskit-core
so that algorithms can be loaded independently, and to stand as examples of how to package new algorithms.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:12.914258+00:00, last updated: 2013-02-01T22:35:43.896767+00:00
(originally reported in Trac by @elehack on 2011-03-24 17:56:17)
Right now, the only normalization we support in item-item and user-user CF is subtracting baselines. It would be useful to abstract this into a strategy so we can do things like z-score normalization.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:24.439719+00:00, last updated: 2013-02-01T22:35:43.098551+00:00
Placeholder for deleted ticket.
Note: This issue has been automatically migrated from Bitbucket
Created by Anonymous on 2013-04-01T00:00:00.0+00:00
(originally reported in Trac by @elehack on 2011-03-24 17:47:42)
Right now our rating implementations fail to implement .equals()
and .hashCode()
.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:23.746940+00:00, last updated: 2013-02-01T22:35:43.213464+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:32:26)
Build necessary infrastructure to use LensKit recommenders with RecLab.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:38:44.128606+00:00, last updated: 2013-02-01T22:35:44.771635+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:19:28)
So far, we do not have any support for basket-based recommendation. The APIs are there, but no data sources or recommenders actually do it.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:38:24.655224+00:00, last updated: 2013-02-01T22:35:46.144753+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:20:49)
We should support Karypis's conditional probability asymmetric similarity function.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:38:27.483717+00:00, last updated: 2013-07-19T21:51:59.761117+00:00
(originally reported in Trac by @elehack on 2011-03-25 16:04:08)
In addition to our fractional holdout scheme for splitting query and probe sets for user-based crossfold evaluation, it would be useful to support leave-''N''-out and retain-''N'' protocols.
All that's required to finish this is support for retain-N protocols.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:29.120420+00:00, last updated: 2013-03-05T18:29:43.403522+00:00
(originally reported in Trac by @elehack on 2011-04-05 14:44:19)
In order to support basket-based recommenders, we need the ability to present unary purchase or click data in the data source. In order to support this, we will introducing the concept of Events which are stored in the DAO.
The upshot of this is that we no longer need a hierarchy of DAO objects - we can just have DAO which provides users, items, and history queries. This also makes all data sources available uniformly - so more exotic recommenders can operate on combinations of purchases and ratings, for example. Subclasses can, of course, introduce additional information such as item metadata.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:47.002449+00:00, last updated: 2013-02-01T22:35:41.165673+00:00
(originally reported in Trac by @elehack on 2011-03-22 20:46:30)
Baseline predictors currently exist almost as their own little entities. We need to be able to evaluate them like any other predictor.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:20.855760+00:00, last updated: 2013-02-01T22:35:43.441992+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:40:55)
Implement the gradient descent fitted baseline with regularization described by Koren in Factorization meets the neighborhood. For this ticket, we aren't implementing the whole SVD++ or Asymmetric-SVD mode; we're just learning the regularized baseline.
This will be LeastSquaresPredictor
in the org.grouplens.lenskit.baseline
package in the lenskit-core
project.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:10.665857+00:00, last updated: 2013-02-05T21:16:27.708650+00:00
(originally reported in Trac by @elehack on 2011-03-22 16:57:00)
Implement Koren's jointly derived neighborhood interpolation weights.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:19.477803+00:00, last updated: 2013-02-01T22:35:43.554457+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:29:58)
Some people still have copies of the NetFlix data set; it would be nice to be able to process it with LensKit.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:38:43.073434+00:00, last updated: 2013-02-01T22:35:45.117116+00:00
(originally reported in Trac by @elehack on 2011-03-28 20:38:18)
FunkSVD has a marked performance decrease when the MeanDamping
parameter is enabled and set to 25 with the item-user mean predictor. This needs to be debugged. I suspect the problem is in ItemUserMeanPredictor
.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:37.180451+00:00, last updated: 2013-02-01T22:35:41.997519+00:00
(originally reported in Trac by @elehack on 2011-03-29 17:44:59)
Having SimilarityMatrixBuildStrategy
family of classes as well as the SimilarityMatrixBuilder
family is confusing. To fix this, we will rename SimilarityMatrixBuildStrategy
to ItemItemModelBuildStrategy
. Along with renaming ItemItemRecommenderBuilder
to ItemItemModelBuilder
, this will make for a much clearer source tree layout.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:38.191254+00:00, last updated: 2013-02-01T22:35:41.887240+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:31:36)
It could be interesting to allow Mahout recommenders to be used in LensKit and vice-versa. It at least has interesting evaluation and comparison potential.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:38:43.760915+00:00, last updated: 2013-02-01T22:35:44.887908+00:00
(originally reported in Trac by @elehack on 2011-03-25 16:04:50)
Support aggregating statistics like MAE and RMSE per-user rather than just overall.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:30.218995+00:00, last updated: 2013-02-01T22:35:42.743459+00:00
(originally reported in Trac by @elehack on 2011-03-24 14:03:48)
Running the evaluator is somewhat complex, particularly with the splitting out of algorithms into separate programs. We need a set of scripts to make this easy. I am not sure what scripts - whether they are in Ant, Maven, or shell scripts - but we need something.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:22.305557+00:00, last updated: 2013-02-01T22:35:43.336753+00:00
(originally reported in Trac by @mludwig on 2011-03-25 21:23:11)
The purpose of this is to enable the ability to include or exclude a user's rated items from the rating set. It is not reasonable to make someone pass in a candidate set consisting of the compliment of the user's rated items.
Instead we need to provide both an include and exclude set of items. Items in the exclude set will never be returned, even if they are in the include set. Items not in the include set will not be returned. If the include set is empty (default), it is the universe of items. The default exclude set is the user's ratings, although the empty set/null is still considered to be empty (to say "don't exclude anything").
This creates a little bit of an inconsistency between include/exclude default behavior, but we can document this or improve it.
Another option is to provide an include set and an exclusion predicate, but this provides the same capabilities as the system above.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:32.023931+00:00, last updated: 2013-02-01T22:35:42.483440+00:00
(originally reported in Trac by @elehack on 2011-03-25 22:31:40)
We would like to upgrade to Maven 3 to take advantage of the new reporting configuration.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:33.505606+00:00, last updated: 2013-02-01T22:35:42.373687+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:46:59)
For [[milestone:1.0]], we need to have some good documentation. This tracker ticket records specific documentation tasks for 1.0.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:15.789898+00:00, last updated: 2013-02-05T21:47:54.992869+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:42:56)
The =Parameter= annotation should get a type parameter to specify the type expected by the parameter. This would enable parameter type-checking in an annotation processor and aid with documentation.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:12.171813+00:00, last updated: 2013-02-01T22:35:44.007590+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:22:19)
Item-item similarity computation can be parallelized by writing a new similarity matrix build strategy. We should do this.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:38:28.772959+00:00, last updated: 2013-02-01T22:35:45.926735+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:17:45)
The Getting Started guide is good, but we need a more extensive user guide to cover setting up LensKit, writing recommenders, etc.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:35:44.177522+00:00, last updated: 2013-02-01T22:35:46.337593+00:00
(originally reported in Trac by @elehack on 2011-04-04 15:54:56)
The CachingNeighborhoodFinder
class needs to use the rating update listener capabilities to update itself.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:41.424757+00:00, last updated: 2013-02-01T22:35:41.655756+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:30:36)
The evaluator needs to support Burke's TemporalMAE and ProfileMAE metrics.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:38:43.416145+00:00, last updated: 2013-02-01T22:35:45.014660+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:40:13)
The mean damping parameter needs to be reviewed for sane naming. It might be somewhat confusing.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:38:52.597584+00:00, last updated: 2013-02-01T22:35:44.326551+00:00
(originally reported in Trac by @elehack on 2011-04-05 14:26:36)
The current prediction/recommendation interface for rating-based systems takes the user as a long ID and a SparseVector
of their ratings. This may couple the interface too closely to LensKit, making bridges (#14, #15) to other systems more difficult to write. Further, it does not allow the prediction or recommendation method to use the timestamps of the user's ratings, or to deal with duplicate ratings.
Therefore, we should consider changing the interfaces to use Collection<Rating>
as the rating input.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:43.807165+00:00, last updated: 2013-02-01T22:35:41.274590+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:39:23)
We want to be able to easily generate multiple variants of an algorithm in the evaluator. In particular, we want to be able to vary a parameter and record that parameter's particular values in the output so we can plot e.g. RMSE vs. neighborhood size.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:38:51.539327+00:00, last updated: 2013-02-01T22:35:44.434178+00:00
(originally reported in Trac by @elehack on 2011-03-28 19:13:28)
LensKit will have a much cleaner API if we drop the RecommenderService
class and allow recommenders and predictors to be directly injected.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:35.733505+00:00, last updated: 2013-02-01T22:35:42.255617+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:28:52)
If the data set has multiple ratings for the same user-item pair, something intelligent should happen. Most of the current code ''should'' be OK with this, but we need to audit the code for good behavior and design APIs as appropriate to allow multiple ratings to be dealt with sanely.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:38:38.699012+00:00, last updated: 2013-02-01T22:35:45.369663+00:00
(originally reported in Trac by @elehack on 2011-03-25 16:05:43)
Create statistic aggregator classes to abstract mean and RMS accumulation.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:30.951936+00:00, last updated: 2013-02-01T22:35:42.623591+00:00
(originally reported in Trac by @elehack on 2011-03-28 20:13:43)
We need a RatingRecommender
implementation for the SVD recommenders.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:36.441028+00:00, last updated: 2013-02-01T22:35:42.113345+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:42:10)
It would be possible to make =RecommenderModuleComponent= automatically bind class parameters. We should think about doing this.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:11.325949+00:00, last updated: 2013-02-01T22:35:44.120179+00:00
(originally reported in Trac by @elehack on 2011-04-05 22:35:08)
We want a separate ImmutableSparseVector
class that extends SparseVector
and represents immutable - as opposed to read-only - sparse vectors. ImmutableSparseVector
s are also safe to share across thread boundaries. Eventually, then, we can introduce value sharing to avoid excess copying of vectors.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:51.605893+00:00, last updated: 2013-02-01T22:35:41.050916+00:00
(originally reported in Trac by @elehack on 2011-03-16 22:27:19)
We would like to have a PLSI recommender in LensKit. References:
The PLSI recommender should be in a new Maven project lenskit-prob
.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:38:37.682862+00:00, last updated: 2013-02-01T22:35:45.472453+00:00
(originally reported in Trac by @elehack on 2011-04-04 15:11:59)
Currently, user-user CF is entirely unnormalized. This needs to be fixed.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:39.657595+00:00, last updated: 2013-02-01T22:35:41.765732+00:00
(originally reported in Trac by @elehack on 2011-04-04 18:13:28)
The algorithm graphing mojo needs to be finished to support multiple algorithm files.
Note: This issue has been automatically migrated from Bitbucket
Created by grouplens on 2013-02-01T21:55:43.115305+00:00, last updated: 2013-02-01T22:35:41.402987+00:00
A declarative, efficient, and flexible JavaScript library for building user interfaces.
๐ Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. ๐๐๐
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google โค๏ธ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.