predictionio-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Pat Ferrel <>
Subject Re: New items in UR
Date Mon, 27 Feb 2017 16:17:18 GMT
We have an in-house Template to implement “Behavioral Search” with the algo inside the
UR called Correlated Cross-Occurrence. It allows you to augment your Search index, which usually
contains mainly content from items. These augmentation fields have information about what
item other people searched for or bought or whatever events you track. In the Search you then
use the search the user typed in as well as other events they have in their history. 

When you make the search it find items with the search terms, as well as items the user is
likely to convert on AND items that other people have searched and converted but with terms
may match one the other people used. The later bit means that if there is a term people use
that may not be in the content but is common, it will be put in the index to augment it. This
could be slang, a common misspelling, or nicckname for the item. 

We call this “Behavioral Search” because the augmentation data comes from either other
people’s behavior or the individual’s behavior. In the blog post below, we make a slight
distinction between Personalized Search and Augmented Search because one uses user-specific
data in the query and the other just uses other people’s data and so is not “personalized”.
In practice it’s almost always better to use both.

TLDR; Behavioral Search will return items based on their content alone, if no behavioral match
is found. it will boost items that the user is likely to convert on. Boost here means that
it will return items with the search terms but favor items with the right behavioral data
when available. <>

On Feb 27, 2017, at 3:48 AM, Masha Zaharchenko <> wrote:

Hi, everyone!
I want to use UR to get  scores for items in search results(to range them). But it`s possible
that an item hasn`t got any interactions yet(hasn`t been viewed or purchased, etc.).
So I have the following question:
Will an item with no events be included in recommendations based only on its properties or
will it be ignored?

View raw message