predictionio-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Pat Ferrel <...@occamsmachete.com>
Subject Re: UR: getting score for the given item
Date Wed, 01 Mar 2017 20:17:08 GMT
Yes, but in 0.5.1 (only waiting for PIO release) you can do that with one query:

    { "user": "u1",”item-set”:[“i1”, “i2”],"num”:2,"returnSelf":true}

This will do the score based ranking on the server in one query to the model. BTW you can
have as many things in the item-set as you want.

In general you can combine scores and re-rank because they are all measuring the same type
of thing in the same way, but beware of combining with any other types of scores that are
not calculated in the same. For instance ensemble recommenders, which combine results for
more than one recommender. More sophisticated combination schemes are required.


On Mar 1, 2017, at 12:06 PM, Masha Zaharchenko <mashastudy@gmail.com> wrote:

Yes, I want to use these scores to rank search results. And it`s unclear to me if I can compare
scores returned by different queries(especially queries which use item id to get recommendations
of similar items).

E.g. A user u1 had searched for keyword "cat" and two items were returned. Items ids are id1
and id2. 

We make two queries which look like this:
curl -H "Content-Type: application/json" -d '{ "user": "u1","item":"i1","num":1,"returnSelf":true}'
 http://localhost:8000/queries.json <http://localhost:8000/queries.json>
curl -H "Content-Type: application/json" -d '{ "user": "u1","item":"i2","num":1,"returnSelf":true}'
 http://localhost:8000/queries.json <http://localhost:8000/queries.json>

And we get two responses:
{"itemScores":[{"item":"i1","score":5}]}
{"itemScores":[{"item":"i2","score":4}]}

The question is: Can we use these scores(4 and 5) to rank the items?

On Wed, Mar 1, 2017 at 7:01 PM, Pat Ferrel <pat@occamsmachete.com <mailto:pat@occamsmachete.com>>
wrote:
The scores are the sum of dot product between the user’s history vectors and the item’s
indicator vectors, with norms and boosts applied for score calculation. Search this list for
more discussion. The math means the more indicators or boosts the higher the possible value.
These are exactly like the scores you get back from Searching with keywords. They have a precise
mathematical meaning but are only used to rank results. they are returned in a form sorted
by score.

Are you trying to use the scores?


On Feb 28, 2017, at 11:23 PM, Masha Zaharchenko <mashastudy@gmail.com <mailto:mashastudy@gmail.com>>
wrote:

Hi, everyone!

I`m currently trying to use the UR to solve the ranking problem by getting scores for the
each item on the list, but results are quite ambiguous and I`m not sure how to interpret them.
(I`m aware of the existence of Product Ranking Engine Template, but I`m curious if  I can
do it with the UR)

Suppose I send the following queries:
A) $ curl -H "Content-Type: application/json" -d '{ "user": "5881C656F1284616BA5D47C42F930497","item":"242234","num":1,"returnSelf":true}'
 http://localhost:8000/queries.json <http://localhost:8000/queries.json>
Response: {"itemScores":[{"item":"242234","score":50.42802810668945}]}

B)$ curl -H "Content-Type: application/json" -d '{ "user": "5881C656F1284616BA5D47C42F930497","item":"83364","num":1,"returnSelf":true}'
Response: {"itemScores":[{"item":"83364","score":51.85137176513672}]}

Are these scores comparable or they only have a meaning within the limits of the query(i.e.
recommending the given item as the most similar to itself)?

Thanks,
Maria




Mime
View raw message