predictionio-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Pat Ferrel <...@occamsmachete.com>
Subject Re: Data UR
Date Tue, 16 May 2017 17:42:29 GMT
Answers below:


On May 16, 2017, at 10:19 AM, Dennis Honders <dennishonders@gmail.com> wrote:

Hi,

1. 
I already used similar product template for experimenting. 
https://predictionio.incubator.apache.org/templates/similarproduct/quickstart/ <https://predictionio.incubator.apache.org/templates/similarproduct/quickstart/>

For UR, are the data queries for the eventserver about the same, but can take more properties?
In my case three events. Set users, set items and set buys. 

The UR only needs the buys and determines users and items from the buys, you’d do better
is you have other events like product detail views, or category of item bought, etc.

2. 
I have coordinates for the users. Is this supported as property?

Yes to location but lat/lon is problematic. Some area location like postal code or something
like country+province+city works much better. These need to be able to contain more than one
person so lat/lon is theoretically not applicable since it is too fine grained.

Note: in my case I like to make predictions by user id and by an array of item ids which is
supported, also for products that are never bought for cold start. I have item properties
like category id, manufacturer id, label and price range. 

All are supported but I’ll warn that you should test these results, mixing user-id and item-sets
has no theoretical basis for working and without correct boosting of one over the other may
interfere and create less good results. Also item-sets can work to produce either "similar
items" or “complimentary items” as in things you might want in the same shopping cart.
These require different model building.

How are you generating the array of items? what is your goal for this? If you want items similar
to the one being viewed—on the current page for instance, use an item-based query, it will
return similar items to the one viewed and can mix with user-based items.

In general everything you mention is supported but my gut feel is that it may be overly complicated
so I’d advise A/B testing with a stripped down simple query against this query to see if
it really does produce better conversions. Let you data be your guide—intuition must be
tested. Adding rules is often needed and is supported but may also reduce conversion lift
in unexpected ways.

Thanks in advance


Mime
View raw message