UR v0.6.0m RC1

[Pat Ferrel]

If anyone want to try out RC1 of the UR with pio 0.11.0 get the `develop` branch of the UR here: https://github.com/actionml/universal-recommender

pull the repo then `git checkout develop` before you build it with pio

I defaults to Spark 1.6, ES 1.7 and Scala 2.10, a UR v0.6.1 will have new build.sbt to support other configs. It does not support ES 5 but there is a PR that is nearly complete so expect that to be incorporated soon.

[Bolmo Joosten]

Thanks for sharing!

V0.6.0 training failed on one of our mock data sets. (Mahout: Incompatible operand geometry). In my case it fails if a user id has purchase event, but no view event.

Input data:

Purchase (userid, itemid

A, 1

B, 3

C, 3

View:

A, 1

B, 2

C, 2

If I remove the last event from view, training fails.

Bolmo

[David Litt]

I receive this error message as well.

[Pat Ferrel]

I’ll check that, thanks.

--
You received this message because you are subscribed to the Google Groups "actionml-user" group.
To unsubscribe from this group and stop receiving emails from it, send an email to actionml-use...@googlegroups.com.
To post to this group, send email to action...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/actionml-user/CANO2qVz3_1hK8a%3DTq%3DO_s0n8Yk%3DudgkZ%3D9cb9zfOfosieosh6A%40mail.gmail.com.
For more options, visit https://groups.google.com/d/optout.

[Pat Ferrel]

Indeed a bug. This is a blocker (no work around) in some conditions so tracking it down now.

To view this discussion on the web visit https://groups.google.com/d/msgid/actionml-user/7C6552B6-C553-4BF0-98B6-B35498DC43D3%40occamsmachete.com.

[Pat Ferrel]

This should be fixed in RC2, now in the UR develop branch

Thanks again Bolmo!

To view this discussion on the web visit https://groups.google.com/d/msgid/actionml-user/048C43ED-5CF9-47C2-9A0A-6FA8D31E33E6%40occamsmachete.com.

[David Litt]

Fixes my issue. Thank you!

[Pat Ferrel]

Good to hear, Thanks David. Any other testing is appreciated.

To view this discussion on the web visit https://groups.google.com/d/msgid/actionml-user/28a9e8ee-4f64-46e3-a0b1-8f68b3719fe4%40googlegroups.com.

[Pat Ferrel]

This will be released next week unless anyone finds an issue. 0.6.0 should be backward compatible with pervious data and run with PIO 0.10.0 and 0.11.0. 

Please have a look if you need any of it’s new features, which are: 

v0.6.0 https://github.com/actionml/universal-recommender/blob/develop/README.md

This is a major upgrade release with several new features. Backward compatibility with 0.5.0 is maintained. Note: We no longer have a default engine.json file so you will need to copy engine.json.template to engine.json and edit it to fit your data. See the Universal Recommender Congiuration docs.

• Performance: Nearly a 40% speedup for most model calculation, and a new tuning parameter that can yield further speed improvements by filtering out unused or less useful data from model building. See minEventsPerUser in the UR configuration docs.

• Complimentary Purchase aka Item-set Recommendations: "Shopping-cart" type recommendations. Can be used for wishlists, favorites, watchlists, any list based recommendations. Used with list or user data.

• Exclusion Rules: now we have business rules for inclusion, exclusion, and boosts based on item properties.

• PredictionIO 0.11.0: Full compatibility, but no support for Elasticsearch 5, an option with PIO-0.11.0.

• New Advanced Tuning: Allows several new per indicator / event type tuning parameters for tuning model quality in a more targeted way.

• Norms Support: For large dense datasets norms are now the default for model indexing and queries. This should result in slight precision gains, so better results.

• GPU Support: via Mahout 0.13.0 the core math of the UR now supports the use of GPUs for acceleration.

• Timeout Protection: Queries for users with very large histories could cause a timeout. We now correctly limit the amount of user history that is used as per documentation, which will all but eliminate timeouts.

• Bug Fixes: The use of blackListEvents as defined in engine.json was not working for an empty list, which should and now does disable any blacklisting except explicit item blacklists contained in the query.

To view this discussion on the web visit https://groups.google.com/d/msgid/actionml-user/658C6149-6476-4139-B1C1-B8737BC491F4%40occamsmachete.com.