predictionio-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Pat Ferrel <...@occamsmachete.com>
Subject Re: Universal Recommender date boosts
Date Wed, 19 Oct 2016 15:10:12 GMT
Create categorical date ranges and attach them to items. They will be "0-10 days”, "11-20
days” etc. Something that encodes how far back from today the data is. This will, of course
need to be updated periodically with a property change event.

Then in your engine.json or query boost "0-10 days” with a bias of something above 1 like
10, and “11-20 days” by something less like 5.

If you limit the number of user history events you will get a recency effect also. This is
discussed in actionml.com/ur-config <http://actionml.com/ur-config>, look for maxQueryEvents.
This will allow you to look for the most current user intent, it will not explicitly boost
new things. IMHO It is likely to have a better effect on your KPIs than an explicit boost
but you’d have to A/B test to be sure.


On Oct 18, 2016, at 7:39 AM, Tom McCaul <tom@tomandco.co.uk> wrote:

Hi,

I'm setting up UR to return an ordered list of items for a category by popularity but I wanted
to boost new in items.

So far, I'm using an attribute to create a boost for those new products but this isn't the
most ideal set up.  I'd prefer for the boost to be weighted over time, so on day 1, the product
receives the biggest boost and by day 28 (just before it's no longer considered new) it's
at its lowest.

Does anyone have any suggestions how I could best approach this?

Thanks,
Tom
-- 
Tom&Co. Ltd.
26 Chelsea Wharf
15 Lots Road
London
SW10 0QJ

Office +44 207 351 3521
Mobile +44 774 010 4045
tom@tomandco.co.uk <mailto:tom@tomandco.co.uk>


Mime
View raw message