You are viewing a plain text version of this content. The canonical link for it is here.

Posted to user@predictionio.apache.org by Daniel Gabrieli <dg...@salesforce.com> on 2017/06/21 21:27:36 UTC

UR Model Metrics Introspection

Hi,

Question about the Universal Recommendation Engine (UR).

The challenge I am experiencing right now is that all my recommendations
come back with a score of 0 (and are the same recommendations for all
users).

After training the engine, is there a way to get the metrics about the
trained model?  For example, which  (if any) indicators have a meaningful
impact on the recommendation?

I would like to tune the parameters maxCorrelatorsPerItem and minLLR but I
am not sure how to figure out how the default parameters differ from the
trained models results.  For example, what is the LLR for the indicators in
the trained model currently?  How many correlators are currently considered
significant?


Thanks!

Re: UR Model Metrics Introspection

Posted by Daniel Gabrieli <dg...@salesforce.com>.

Thanks Pat.  The error was on my side.  All good now.  Cheers!

On Wed, Jun 21, 2017 at 3:08 PM Pat Ferrel <pa...@occamsmachete.com> wrote:

> If all recs come back score 0 then they are from the most popular items
> only, not from the Collaborative Filtering algorithm.
>
> If you truly mean “all” then you have a problem in config or nearly 0
> data. Don’t mess with tuning until you solve this.
>
>
> On Jun 21, 2017, at 2:27 PM, Daniel Gabrieli <dg...@salesforce.com>
> wrote:
>
> Hi,
>
> Question about the Universal Recommendation Engine (UR).
>
> The challenge I am experiencing right now is that all my recommendations
> come back with a score of 0 (and are the same recommendations for all
> users).
>
> After training the engine, is there a way to get the metrics about the
> trained model?  For example, which  (if any) indicators have a meaningful
> impact on the recommendation?
>
> I would like to tune the parameters maxCorrelatorsPerItem and minLLR but I
> am not sure how to figure out how the default parameters differ from the
> trained models results.  For example, what is the LLR for the indicators in
> the trained model currently?  How many correlators are currently considered
> significant?
>
>
> Thanks!
>
>
>
>

Re: UR Model Metrics Introspection

Posted by Pat Ferrel <pa...@occamsmachete.com>.

If all recs come back score 0 then they are from the most popular items only, not from the Collaborative Filtering algorithm.

If you truly mean “all” then you have a problem in config or nearly 0 data. Don’t mess with tuning until you solve this.


On Jun 21, 2017, at 2:27 PM, Daniel Gabrieli <dg...@salesforce.com> wrote:

Hi,

Question about the Universal Recommendation Engine (UR).  

The challenge I am experiencing right now is that all my recommendations come back with a score of 0 (and are the same recommendations for all users).

After training the engine, is there a way to get the metrics about the trained model?  For example, which  (if any) indicators have a meaningful impact on the recommendation?

I would like to tune the parameters maxCorrelatorsPerItem and minLLR but I am not sure how to figure out how the default parameters differ from the trained models results.  For example, what is the LLR for the indicators in the trained model currently?  How many correlators are currently considered significant?


Thanks!